{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000001", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [232, 352, 243, 366]\n- T2: [246, 453, 277, 500]\n- T3: [398, 562, 403, 575]\n- T4: [178, 453, 197, 472]\n\nImage 2 (GROUND) labeled targets:\n- T1: [522, 275, 531, 287]\n- T2: [537, 250, 579, 294]\n- T3: [845, 242, 878, 302]\n- T4: [470, 278, 489, 292]\n\nAnswer options:\nA: T3 -> T2 -> T1 -> T4\nB: T3 -> T4 -> T2 -> T1\nC: T3 -> T2 -> T4 -> T1\nD: T2 -> T3 -> T4 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/65bc64984b2063a131f48fe526cf6a7b07da1d78c38097687558bd599f26a609.jpg", "reference_inputs/images/326e309493b74656ab397a5c7fe4a7d782f7c238e5d6a88d3395323169ed9146.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000002", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [232, 352, 243, 366]\n- T2: [398, 562, 403, 575]\n- T3: [246, 453, 277, 500]\n- T4: [178, 453, 197, 472]\n\nImage 2 (GROUND) labeled targets:\n- T1: [522, 275, 531, 287]\n- T2: [845, 242, 878, 302]\n- T3: [537, 250, 579, 294]\n- T4: [470, 278, 489, 292]\n\nAnswer options:\nA: T2 -> T3 -> T4 -> T1\nB: T2 -> T3 -> T1 -> T4\nC: T2 -> T4 -> T3 -> T1\nD: T3 -> T2 -> T4 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/65bc64984b2063a131f48fe526cf6a7b07da1d78c38097687558bd599f26a609.jpg", "reference_inputs/images/326e309493b74656ab397a5c7fe4a7d782f7c238e5d6a88d3395323169ed9146.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000003", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [486, 418, 498, 430]\n- T2: [482, 354, 493, 370]\n- T3: [439, 454, 460, 479]\n- T4: [650, 529, 663, 551]\n\nImage 2 (GROUND) labeled targets:\n- T1: [482, 281, 495, 290]\n- T2: [479, 274, 491, 288]\n- T3: [422, 275, 449, 293]\n- T4: [698, 265, 714, 294]\n\nAnswer options:\nA: T3 -> T4 -> T1 -> T2\nB: T4 -> T3 -> T2 -> T1\nC: T4 -> T1 -> T3 -> T2\nD: T4 -> T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2dc7235177eab99e00758787c60e4230c1c1263feedc79745fdcd2eae1d0ed46.jpg", "reference_inputs/images/9ad8a50cccc69d4d19587748ca75a29e617773e0bff87ecdaec8ddeb22d583f1.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000005", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [942, 43, 992, 66]\n- T2: [327, 553, 345, 575]\n- T3: [652, 54, 673, 81]\n\nImage 2 (GROUND) labeled targets:\n- T1: [262, 263, 276, 295]\n- T2: [700, 270, 706, 291]\n- T3: [482, 258, 493, 296]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: unable to determine\nC: T3 -> T1 -> T2\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b8792d16dea90b6ef0611eaee2ce16526d9ae336d106646657cea1b42a3c1d7a.jpg", "reference_inputs/images/82cfc3331a51c8535d78f6cbee980b1c6984fdb50a3c30fe64984dbe6581143a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000006", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [524, 422, 539, 434]\n- T2: [718, 553, 725, 575]\n- T3: [586, 496, 591, 517]\n\nImage 2 (GROUND) labeled targets:\n- T1: [819, 279, 886, 297]\n- T2: [104, 253, 118, 299]\n- T3: [272, 265, 291, 294]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T2 -> T1 -> T3\nC: T3 -> T2 -> T1\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e0381b03e29afa40e6d18828c0f36bc0d89bb5a74d329b1a9516c1a79e4b51a4.jpg", "reference_inputs/images/57e4694e7f958f730076926ae88c4407181e8cbbd683ad2fdc003377784f44e2.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000008", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [507, 418, 520, 430]\n- T2: [458, 422, 474, 433]\n- T3: [489, 385, 499, 396]\n\nImage 2 (GROUND) labeled targets:\n- T1: [472, 245, 594, 338]\n- T2: [0, 262, 219, 346]\n- T3: [437, 270, 473, 298]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T3 -> T1\nC: T1 -> T3 -> T2\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ce8efefe6296b719d57192eb8ddd3beade6a537c2ce5be71960d4a9c54866577.jpg", "reference_inputs/images/360617ebd1ddf8668083527a4aeebbdda1341224f852b2297b1aa4e0edd6dbf3.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000009", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [643, 550, 654, 566]\n- T2: [642, 482, 646, 501]\n- T3: [616, 494, 622, 515]\n\nImage 2 (GROUND) labeled targets:\n- T1: [797, 231, 848, 318]\n- T2: [264, 234, 283, 307]\n- T3: [334, 218, 361, 318]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T1 -> T3 -> T2\nC: T3 -> T1 -> T2\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bea921d528598e67ce2c2c61b17817b00fbeb16fe4424686f86ef1b39ff8552c.jpg", "reference_inputs/images/0434ee60b659113cb97e064670e8b02e9955eeff7f24ccedb8f970345927aa67.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000010", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [751, 522, 786, 548]\n- T2: [92, 330, 136, 348]\n- T3: [504, 51, 555, 166]\n\nImage 2 (GROUND) labeled targets:\n- T1: [829, 246, 873, 304]\n- T2: [158, 254, 173, 298]\n- T3: [507, 274, 538, 297]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T1 -> T2\nC: T2 -> T1 -> T3\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ed6a292d6cd804eddcda4961ead9b4a2b5e9bae9f81b95a5aabdcde5d500b5d6.jpg", "reference_inputs/images/95c96102e8a9126a8dc2362e391865f4cccb2be1df7cbe5c3bfa84cbdc0dab96.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000011", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [512, 0, 564, 93]\n- T2: [487, 318, 538, 429]\n- T3: [667, 402, 730, 514]\n\nImage 2 (GROUND) labeled targets:\n- T1: [374, 279, 399, 294]\n- T2: [318, 278, 354, 296]\n- T3: [399, 282, 434, 298]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T2 -> T3 -> T1\nC: T3 -> T2 -> T1\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f6a79034b298f8ff2f1b93aaa565a24635361b61b7454da8eb86027d450c4efa.jpg", "reference_inputs/images/136f851f4ce6fede76cc50fdc3fc6048d01d9a5ec12763732777c13fbb606be9.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000012", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [974, 564, 997, 575]\n- T2: [329, 0, 349, 41]\n- T3: [381, 422, 497, 488]\n\nImage 2 (GROUND) labeled targets:\n- T1: [1016, 270, 1023, 294]\n- T2: [555, 259, 574, 289]\n- T3: [914, 278, 994, 306]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T3 -> T1\nC: T3 -> T2 -> T1\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/90f04a6a2238a7fdcc466fd00f9d19ae60103ef575e4a7d3a7eff0223c6367b0.jpg", "reference_inputs/images/6e0075721026d6fd27359c56d0e9c4227a96e72271da50ba1dddeb99595ec730.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000013", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [407, 350, 418, 364]\n- T2: [358, 464, 377, 486]\n- T3: [422, 478, 458, 535]\n\nImage 2 (GROUND) labeled targets:\n- T1: [818, 285, 830, 296]\n- T2: [781, 290, 795, 299]\n- T3: [835, 273, 862, 300]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T2 -> T1 -> T3\nC: T3 -> T1 -> T2\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/61e5799ec4365ae66e47f25111b7621c86ca0ffc24894e6fee6505c7db389c9c.jpg", "reference_inputs/images/33479d35b44aa18cd3869f33debd528efca86b2a06e5055948e2dd6a95cfdcf7.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000014", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [407, 350, 418, 364]\n- T2: [422, 478, 458, 535]\n- T3: [358, 464, 377, 486]\n\nImage 2 (GROUND) labeled targets:\n- T1: [818, 285, 830, 296]\n- T2: [835, 273, 862, 300]\n- T3: [781, 290, 795, 299]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T2 -> T1 -> T3\nC: T3 -> T2 -> T1\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/61e5799ec4365ae66e47f25111b7621c86ca0ffc24894e6fee6505c7db389c9c.jpg", "reference_inputs/images/33479d35b44aa18cd3869f33debd528efca86b2a06e5055948e2dd6a95cfdcf7.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000015", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [475, 412, 490, 423]\n- T2: [632, 510, 637, 531]\n- T3: [679, 565, 686, 575]\n- T4: [534, 448, 550, 465]\n\nImage 2 (GROUND) labeled targets:\n- T1: [476, 282, 490, 290]\n- T2: [628, 272, 633, 291]\n- T3: [673, 268, 679, 293]\n- T4: [534, 280, 550, 290]\n\nAnswer options:\nA: T3 -> T2 -> T1 -> T4\nB: T3 -> T4 -> T2 -> T1\nC: T3 -> T2 -> T4 -> T1\nD: T2 -> T3 -> T4 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a4cf592a72e7c039e89f2ce68a4f098467de07ca020316c2c156372cb9a1cfc5.jpg", "reference_inputs/images/9762b261fdab1ae82f5704a59390f7a7a5f789c680e4cf37c3c11ba8d2d2fc0f.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000016", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [632, 510, 637, 531]\n- T2: [534, 448, 550, 465]\n- T3: [499, 388, 517, 412]\n- T4: [679, 565, 686, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [628, 272, 633, 291]\n- T2: [534, 280, 550, 290]\n- T3: [500, 270, 517, 289]\n- T4: [673, 268, 679, 293]\n\nAnswer options:\nA: T1 -> T4 -> T2 -> T3\nB: T4 -> T2 -> T1 -> T3\nC: T4 -> T1 -> T3 -> T2\nD: T4 -> T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a4cf592a72e7c039e89f2ce68a4f098467de07ca020316c2c156372cb9a1cfc5.jpg", "reference_inputs/images/9762b261fdab1ae82f5704a59390f7a7a5f789c680e4cf37c3c11ba8d2d2fc0f.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000017", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [480, 355, 492, 372]\n- T2: [466, 391, 490, 399]\n- T3: [394, 463, 398, 482]\n- T4: [446, 446, 463, 464]\n\nImage 2 (GROUND) labeled targets:\n- T1: [469, 271, 486, 289]\n- T2: [446, 282, 480, 290]\n- T3: [250, 257, 262, 297]\n- T4: [381, 274, 420, 298]\n\nAnswer options:\nA: T3 -> T4 -> T1 -> T2\nB: T3 -> T2 -> T4 -> T1\nC: T3 -> T4 -> T2 -> T1\nD: T4 -> T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/54a45b98db6cf9dcae184341b11c3ada9efb63b5c8c0222a21ea269ead7746f1.jpg", "reference_inputs/images/22e82f3f5da3ca8e0a41507973a1054fc81a0436bd364afb7118546fe3ec6147.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000018", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [394, 463, 398, 482]\n- T2: [466, 391, 490, 399]\n- T3: [446, 446, 463, 464]\n- T4: [510, 570, 532, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [250, 257, 262, 297]\n- T2: [446, 282, 480, 290]\n- T3: [381, 274, 420, 298]\n- T4: [485, 233, 753, 396]\n\nAnswer options:\nA: T1 -> T4 -> T3 -> T2\nB: unable to determine\nC: T4 -> T1 -> T3 -> T2\nD: T4 -> T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/54a45b98db6cf9dcae184341b11c3ada9efb63b5c8c0222a21ea269ead7746f1.jpg", "reference_inputs/images/22e82f3f5da3ca8e0a41507973a1054fc81a0436bd364afb7118546fe3ec6147.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000019", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [454, 428, 470, 447]\n- T2: [494, 430, 516, 466]\n- T3: [458, 468, 480, 490]\n\nImage 2 (GROUND) labeled targets:\n- T1: [390, 267, 424, 285]\n- T2: [469, 223, 529, 299]\n- T3: [343, 273, 424, 306]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T2 -> T1 -> T3\nC: T3 -> T2 -> T1\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0338e6668ccc2980e70ebd4a9e32d2deccdc84389a69907b91810ac1184821af.jpg", "reference_inputs/images/a4ff943138300e14df5dde6b5b4d48e1550533ac77ffe671ed28e02e96353024.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000020", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [454, 428, 470, 447]\n- T2: [458, 468, 480, 490]\n- T3: [494, 430, 516, 466]\n\nImage 2 (GROUND) labeled targets:\n- T1: [390, 267, 424, 285]\n- T2: [343, 273, 424, 306]\n- T3: [469, 223, 529, 299]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T2 -> T3 -> T1\nC: T1 -> T2 -> T3\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0338e6668ccc2980e70ebd4a9e32d2deccdc84389a69907b91810ac1184821af.jpg", "reference_inputs/images/a4ff943138300e14df5dde6b5b4d48e1550533ac77ffe671ed28e02e96353024.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000022", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [713, 515, 736, 542]\n- T2: [98, 82, 134, 106]\n- T3: [374, 63, 432, 177]\n\nImage 2 (GROUND) labeled targets:\n- T1: [715, 256, 730, 306]\n- T2: [268, 269, 275, 294]\n- T3: [414, 282, 452, 299]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T3 -> T2\nC: T1 -> T2 -> T3\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8771c85ad7e20bee39b6eb3be777a3478801bbd3fe47bf6dc3ea6bcf510f6924.jpg", "reference_inputs/images/5656546a0d9e30f3637c2bbbabe2b6fcd0e76f79f385253514533b4f71a004c0.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000023", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [374, 63, 432, 177]\n- T2: [713, 515, 736, 542]\n- T3: [98, 82, 134, 106]\n\nImage 2 (GROUND) labeled targets:\n- T1: [414, 282, 452, 299]\n- T2: [715, 256, 730, 306]\n- T3: [268, 269, 275, 294]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T2 -> T1 -> T3\nC: T1 -> T3 -> T2\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8771c85ad7e20bee39b6eb3be777a3478801bbd3fe47bf6dc3ea6bcf510f6924.jpg", "reference_inputs/images/5656546a0d9e30f3637c2bbbabe2b6fcd0e76f79f385253514533b4f71a004c0.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000024", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000024", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [430, 458, 464, 520]\n- T2: [510, 466, 526, 485]\n- T3: [468, 399, 480, 408]\n- T4: [641, 464, 644, 482]\n\nImage 2 (GROUND) labeled targets:\n- T1: [13, 152, 344, 330]\n- T2: [502, 270, 558, 309]\n- T3: [434, 282, 454, 292]\n- T4: [926, 244, 946, 302]\n\nAnswer options:\nA: T1 -> T2 -> T3 -> T4\nB: T2 -> T1 -> T4 -> T3\nC: T1 -> T4 -> T2 -> T3\nD: T1 -> T2 -> T4 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/09d6bbd65d5060eae93487145b4ecd97a46eec4da32909caff0ec56456d840de.jpg", "reference_inputs/images/b44ca937189d4b9529c968b0f3bea457bb9b34d3c8f1a0061bfd35bda0a95e8c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000027", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [162, 121, 205, 139]\n- T2: [10, 205, 58, 222]\n- T3: [322, 230, 372, 331]\n\nImage 2 (GROUND) labeled targets:\n- T1: [539, 235, 560, 311]\n- T2: [440, 251, 458, 304]\n- T3: [0, 250, 310, 358]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T3 -> T2 -> T1\nC: T3 -> T1 -> T2\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0270d8b538279fc1cc2d42738a8e13de18ac9681608ba8b8b64e1d8f6abef7bd.jpg", "reference_inputs/images/7bb72823dd1ff767bcfa0af18ffdf886134eb00fa75108eafdef971bd5ab925b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000028", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [322, 230, 372, 331]\n- T2: [162, 121, 205, 139]\n- T3: [10, 205, 58, 222]\n\nImage 2 (GROUND) labeled targets:\n- T1: [0, 250, 310, 358]\n- T2: [539, 235, 560, 311]\n- T3: [440, 251, 458, 304]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T3 -> T2\nC: T1 -> T2 -> T3\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0270d8b538279fc1cc2d42738a8e13de18ac9681608ba8b8b64e1d8f6abef7bd.jpg", "reference_inputs/images/7bb72823dd1ff767bcfa0af18ffdf886134eb00fa75108eafdef971bd5ab925b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000029", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [474, 361, 486, 379]\n- T2: [530, 448, 546, 465]\n- T3: [618, 463, 622, 480]\n\nImage 2 (GROUND) labeled targets:\n- T1: [190, 263, 225, 291]\n- T2: [342, 262, 424, 314]\n- T3: [767, 232, 791, 306]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T3 -> T1 -> T2\nC: T3 -> T2 -> T1\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c18ed58b26d234fb469fd862a56269881f2ef4151f5fe299c88cbab19df890eb.jpg", "reference_inputs/images/bf6e1f1858c88fb977f0dbc32936d9c13cdd676f51a1e07b64daf08ce71bb42b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000030", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [530, 448, 546, 465]\n- T2: [474, 361, 486, 379]\n- T3: [618, 463, 622, 480]\n\nImage 2 (GROUND) labeled targets:\n- T1: [342, 262, 424, 314]\n- T2: [190, 263, 225, 291]\n- T3: [767, 232, 791, 306]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T1 -> T2\nC: T3 -> T2 -> T1\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c18ed58b26d234fb469fd862a56269881f2ef4151f5fe299c88cbab19df890eb.jpg", "reference_inputs/images/bf6e1f1858c88fb977f0dbc32936d9c13cdd676f51a1e07b64daf08ce71bb42b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000034", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000034", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [469, 370, 484, 391]\n- T2: [534, 490, 555, 514]\n- T3: [503, 414, 532, 427]\n\nImage 2 (GROUND) labeled targets:\n- T1: [373, 262, 401, 291]\n- T2: [609, 218, 1023, 386]\n- T3: [427, 270, 497, 297]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T2 -> T3 -> T1\nC: T2 -> T1 -> T3\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6d76248e7087fb5fcf6decf03cdcb4e685fdbb05d11f13983b35973a1a64e154.jpg", "reference_inputs/images/9f8bd7afecd76acd880fa4f00ef12b38a4ea1e2d76f817d9fdbce48abab400e1.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000035", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000035", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [503, 414, 532, 427]\n- T2: [469, 370, 484, 391]\n- T3: [534, 490, 555, 514]\n\nImage 2 (GROUND) labeled targets:\n- T1: [427, 270, 497, 297]\n- T2: [373, 262, 401, 291]\n- T3: [609, 218, 1023, 386]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T1 -> T2\nC: T3 -> T2 -> T1\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6d76248e7087fb5fcf6decf03cdcb4e685fdbb05d11f13983b35973a1a64e154.jpg", "reference_inputs/images/9f8bd7afecd76acd880fa4f00ef12b38a4ea1e2d76f817d9fdbce48abab400e1.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000036", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000036", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [646, 557, 651, 575]\n- T2: [395, 494, 444, 575]\n- T3: [501, 376, 516, 396]\n- T4: [517, 459, 537, 474]\n\nImage 2 (GROUND) labeled targets:\n- T1: [862, 241, 884, 306]\n- T2: [210, 206, 374, 310]\n- T3: [498, 267, 518, 290]\n- T4: [520, 280, 554, 296]\n\nAnswer options:\nA: T2 -> T1 -> T4 -> T3\nB: T1 -> T2 -> T3 -> T4\nC: T1 -> T4 -> T2 -> T3\nD: T1 -> T2 -> T4 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d0c7a1e600af97e4214a01d83668ccb9038b75752431c9384f9b153978205fed.jpg", "reference_inputs/images/6c5bcc96e9ec88143a2af460a2c36c62a89566b5824bb951b130bc5915679c3b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000037", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000037", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [646, 557, 651, 575]\n- T2: [501, 376, 516, 396]\n- T3: [517, 459, 537, 474]\n- T4: [395, 494, 444, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [862, 241, 884, 306]\n- T2: [498, 267, 518, 290]\n- T3: [520, 280, 554, 296]\n- T4: [210, 206, 374, 310]\n\nAnswer options:\nA: T4 -> T1 -> T3 -> T2\nB: T1 -> T4 -> T3 -> T2\nC: T1 -> T3 -> T4 -> T2\nD: T1 -> T4 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d0c7a1e600af97e4214a01d83668ccb9038b75752431c9384f9b153978205fed.jpg", "reference_inputs/images/6c5bcc96e9ec88143a2af460a2c36c62a89566b5824bb951b130bc5915679c3b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000038", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000038", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [526, 378, 540, 398]\n- T2: [459, 407, 486, 418]\n- T3: [476, 363, 490, 382]\n\nImage 2 (GROUND) labeled targets:\n- T1: [546, 239, 606, 298]\n- T2: [264, 275, 405, 308]\n- T3: [424, 257, 463, 293]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T1 -> T2 -> T3\nC: T2 -> T3 -> T1\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ad7e06d19be14c7572094fcd2a8656811827fd7f8c582feb9f85922ac8d4a358.jpg", "reference_inputs/images/bb1234fc75db8b2004f6e5dad03553677922756abe73a2bc76ac3db2608e9339.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000039", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000039", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [476, 363, 490, 382]\n- T2: [526, 378, 540, 398]\n- T3: [459, 407, 486, 418]\n\nImage 2 (GROUND) labeled targets:\n- T1: [424, 257, 463, 293]\n- T2: [546, 239, 606, 298]\n- T3: [264, 275, 405, 308]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T1 -> T2\nC: T3 -> T2 -> T1\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ad7e06d19be14c7572094fcd2a8656811827fd7f8c582feb9f85922ac8d4a358.jpg", "reference_inputs/images/bb1234fc75db8b2004f6e5dad03553677922756abe73a2bc76ac3db2608e9339.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000040", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000040", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [748, 23, 778, 51]\n- T2: [822, 434, 857, 452]\n- T3: [753, 274, 786, 287]\n\nImage 2 (GROUND) labeled targets:\n- T1: [58, 216, 98, 316]\n- T2: [800, 238, 822, 310]\n- T3: [557, 232, 586, 318]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T3 -> T2 -> T1\nC: T1 -> T2 -> T3\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ab2188839eca588e14ececd553c05ab309b3db806f153f7d78edfdf9250a16d7.jpg", "reference_inputs/images/4ea6834c85dd9da4d0e393b1fbd27477d525a255f2c04252747fb6893e6cca2b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000041", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000041", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [410, 64, 463, 177]\n- T2: [782, 302, 810, 315]\n- T3: [318, 360, 376, 473]\n\nImage 2 (GROUND) labeled targets:\n- T1: [467, 275, 491, 293]\n- T2: [655, 268, 662, 292]\n- T3: [390, 280, 431, 297]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T3 -> T2 -> T1\nC: T2 -> T3 -> T1\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b355ce4a747caad0252734b0d1a30745dc9d99b13180da815406c6aa9dbfb60e.jpg", "reference_inputs/images/64aea7386a93b8d411c31c3c476f955f4bb9d442dea86752dfdbfd3a90eec182.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000042", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000042", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [452, 221, 537, 430]\n- T2: [54, 482, 110, 502]\n- T3: [151, 180, 193, 198]\n\nImage 2 (GROUND) labeled targets:\n- T1: [475, 234, 526, 299]\n- T2: [111, 257, 125, 298]\n- T3: [310, 265, 317, 293]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T1 -> T2 -> T3\nC: T2 -> T1 -> T3\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/de07fc500f433a6eb9993bf376f2d200ba9871b557f96ad81813d6e0f4570870.jpg", "reference_inputs/images/c75e576ecbd789e91ef2be95caf7d32345bf6782da6c9167586dc89ad790e35c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000043", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000043", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [312, 259, 365, 370]\n- T2: [162, 274, 183, 287]\n- T3: [100, 519, 138, 545]\n\nImage 2 (GROUND) labeled targets:\n- T1: [370, 234, 734, 354]\n- T2: [574, 241, 594, 268]\n- T3: [239, 248, 254, 308]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T1 -> T2 -> T3\nC: T3 -> T1 -> T2\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6af64a6f155a53fd7e9405fd0240f3d88f46f26297b3a634601d63726b1edce1.jpg", "reference_inputs/images/5c64e995683b1de21c283c671349dac82351d1b3a64375df3ccdb6bffaedb802.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000044", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000044", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [100, 519, 138, 545]\n- T2: [312, 259, 365, 370]\n- T3: [183, 486, 216, 513]\n\nImage 2 (GROUND) labeled targets:\n- T1: [239, 248, 254, 308]\n- T2: [370, 234, 734, 354]\n- T3: [218, 233, 246, 314]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T2 -> T1 -> T3\nC: T3 -> T2 -> T1\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6af64a6f155a53fd7e9405fd0240f3d88f46f26297b3a634601d63726b1edce1.jpg", "reference_inputs/images/5c64e995683b1de21c283c671349dac82351d1b3a64375df3ccdb6bffaedb802.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000045", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000045", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [89, 470, 94, 487]\n- T2: [170, 463, 176, 479]\n- T3: [52, 508, 118, 529]\n- T4: [385, 505, 432, 525]\n\nImage 2 (GROUND) labeled targets:\n- T1: [686, 279, 691, 301]\n- T2: [742, 280, 762, 304]\n- T3: [596, 287, 632, 303]\n- T4: [842, 279, 1023, 356]\n\nAnswer options:\nA: T4 -> T2 -> T3 -> T1\nB: T4 -> T3 -> T1 -> T2\nC: T4 -> T3 -> T2 -> T1\nD: T3 -> T4 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7585f41fe0b84cff321aa7f677c0cca23c67433ce4c183eb9baafe39b873f4d3.jpg", "reference_inputs/images/b829ddf4c587e79a67ffe16fe25873fe19448bfe353db00f3602571b6511c0a3.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000046", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000046", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [170, 463, 176, 479]\n- T2: [52, 508, 118, 529]\n- T3: [385, 505, 432, 525]\n- T4: [89, 470, 94, 487]\n\nImage 2 (GROUND) labeled targets:\n- T1: [742, 280, 762, 304]\n- T2: [596, 287, 632, 303]\n- T3: [842, 279, 1023, 356]\n- T4: [686, 279, 691, 301]\n\nAnswer options:\nA: T3 -> T2 -> T1 -> T4\nB: T3 -> T1 -> T2 -> T4\nC: T3 -> T2 -> T4 -> T1\nD: T2 -> T3 -> T1 -> T4\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7585f41fe0b84cff321aa7f677c0cca23c67433ce4c183eb9baafe39b873f4d3.jpg", "reference_inputs/images/b829ddf4c587e79a67ffe16fe25873fe19448bfe353db00f3602571b6511c0a3.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000047", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000047", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [162, 46, 215, 70]\n- T2: [300, 20, 359, 135]\n- T3: [121, 475, 156, 499]\n\nImage 2 (GROUND) labeled targets:\n- T1: [644, 242, 668, 281]\n- T2: [562, 260, 935, 350]\n- T3: [34, 250, 67, 310]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T1 -> T3\nC: T2 -> T3 -> T1\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6263899b1ea338624c0248869d9cfc22791beeb4099213fcee8cd99486bbd9a5.jpg", "reference_inputs/images/b7087cf8dfff20392c5cce16fb1cbc197f4b8a023fd6a1a14e127507670e1033.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000048", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000048", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [10, 205, 58, 222]\n- T2: [162, 121, 205, 139]\n- T3: [328, 180, 351, 194]\n\nImage 2 (GROUND) labeled targets:\n- T1: [294, 252, 312, 305]\n- T2: [330, 237, 350, 312]\n- T3: [1, 216, 46, 350]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T1 -> T3 -> T2\nC: T3 -> T2 -> T1\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0270d8b538279fc1cc2d42738a8e13de18ac9681608ba8b8b64e1d8f6abef7bd.jpg", "reference_inputs/images/3e7836c5acb2dcdcaa8ae2df4de0058e570a9f85222862e2d16221447f8c8257.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000049", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000049", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [10, 205, 58, 222]\n- T2: [328, 180, 351, 194]\n- T3: [162, 121, 205, 139]\n\nImage 2 (GROUND) labeled targets:\n- T1: [294, 252, 312, 305]\n- T2: [1, 216, 46, 350]\n- T3: [330, 237, 350, 312]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T2 -> T3 -> T1\nC: T3 -> T2 -> T1\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0270d8b538279fc1cc2d42738a8e13de18ac9681608ba8b8b64e1d8f6abef7bd.jpg", "reference_inputs/images/3e7836c5acb2dcdcaa8ae2df4de0058e570a9f85222862e2d16221447f8c8257.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000050", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000050", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [567, 108, 653, 309]\n- T2: [864, 513, 909, 542]\n- T3: [755, 113, 782, 133]\n\nImage 2 (GROUND) labeled targets:\n- T1: [545, 244, 590, 296]\n- T2: [834, 254, 846, 298]\n- T3: [643, 270, 648, 291]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T1 -> T3\nC: unable to determine\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5fcef0ed1e787fc5faa2ef566001887a3b83ebedf6e08c51a36e92a864bcc37c.jpg", "reference_inputs/images/f715261c31086ae963af6c95167737fcc81cedadfdf0f8460aaa5262d2a8969f.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000051", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000051", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [864, 513, 909, 542]\n- T2: [755, 113, 782, 133]\n- T3: [567, 108, 653, 309]\n\nImage 2 (GROUND) labeled targets:\n- T1: [834, 254, 846, 298]\n- T2: [643, 270, 648, 291]\n- T3: [545, 244, 590, 296]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: unable to determine\nC: T3 -> T2 -> T1\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5fcef0ed1e787fc5faa2ef566001887a3b83ebedf6e08c51a36e92a864bcc37c.jpg", "reference_inputs/images/f715261c31086ae963af6c95167737fcc81cedadfdf0f8460aaa5262d2a8969f.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000055", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000055", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [482, 354, 493, 370]\n- T2: [542, 463, 560, 480]\n- T3: [404, 543, 438, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [482, 275, 494, 287]\n- T2: [538, 281, 552, 290]\n- T3: [428, 282, 442, 293]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T1 -> T2\nC: T3 -> T2 -> T1\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2dc7235177eab99e00758787c60e4230c1c1263feedc79745fdcd2eae1d0ed46.jpg", "reference_inputs/images/4d1429f8a559ffa76ff6dd6181113d642a7d6ccac748309b027172a63950138b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000056", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000056", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [139, 361, 174, 382]\n- T2: [402, 31, 453, 146]\n- T3: [567, 284, 652, 500]\n\nImage 2 (GROUND) labeled targets:\n- T1: [2, 238, 29, 303]\n- T2: [424, 271, 466, 301]\n- T3: [573, 162, 749, 317]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T3 -> T1 -> T2\nC: T1 -> T2 -> T3\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0fc31c1a19b1bb2cdd002913a35bc79ef3f253358915b124d93a788326e3387b.jpg", "reference_inputs/images/e6f52202e903210d25243e3735afe40fa075f6ed21c211af9dc0feaa58a7cb34.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000058", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000058", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [506, 387, 515, 397]\n- T2: [460, 412, 474, 428]\n- T3: [467, 433, 481, 447]\n\nImage 2 (GROUND) labeled targets:\n- T1: [498, 274, 521, 293]\n- T2: [363, 266, 406, 299]\n- T3: [318, 270, 398, 310]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T2 -> T1 -> T3\nC: T3 -> T2 -> T1\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/760df4126276baeb1783940be483f8faf7c7cad7e4581beac622ddf0980c251b.jpg", "reference_inputs/images/b4d21a7e3f68d56255291f70c6384a51e8da3d9f3012b231848b23a57471c549.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000060", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000060", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [398, 562, 403, 575]\n- T2: [156, 423, 176, 436]\n- T3: [246, 453, 277, 500]\n- T4: [178, 453, 197, 472]\n\nImage 2 (GROUND) labeled targets:\n- T1: [358, 271, 369, 291]\n- T2: [176, 283, 192, 290]\n- T3: [240, 264, 274, 291]\n- T4: [194, 281, 210, 290]\n\nAnswer options:\nA: T3 -> T1 -> T4 -> T2\nB: T1 -> T3 -> T4 -> T2\nC: unable to determine\nD: T1 -> T4 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/65bc64984b2063a131f48fe526cf6a7b07da1d78c38097687558bd599f26a609.jpg", "reference_inputs/images/876792b574311b760babd853433624869dbf4494f45b87e39af7b6aa66402447.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000061", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000061", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [753, 274, 786, 287]\n- T2: [414, 325, 460, 426]\n- T3: [748, 23, 778, 51]\n- T4: [106, 422, 143, 453]\n\nImage 2 (GROUND) labeled targets:\n- T1: [675, 266, 698, 294]\n- T2: [434, 275, 472, 300]\n- T3: [635, 268, 641, 292]\n- T4: [178, 259, 190, 297]\n\nAnswer options:\nA: T4 -> T2 -> T1 -> T3\nB: T2 -> T4 -> T3 -> T1\nC: T2 -> T4 -> T1 -> T3\nD: T2 -> T1 -> T4 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ab2188839eca588e14ececd553c05ab309b3db806f153f7d78edfdf9250a16d7.jpg", "reference_inputs/images/72cd0a8adb1640bf26b9f9f7b03503233379e3d16e3754adda473802969f4c0f.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000062", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000062", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [414, 325, 460, 426]\n- T2: [753, 274, 786, 287]\n- T3: [106, 422, 143, 453]\n- T4: [748, 23, 778, 51]\n\nImage 2 (GROUND) labeled targets:\n- T1: [434, 275, 472, 300]\n- T2: [675, 266, 698, 294]\n- T3: [178, 259, 190, 297]\n- T4: [635, 268, 641, 292]\n\nAnswer options:\nA: T2 -> T1 -> T4 -> T3\nB: T1 -> T2 -> T3 -> T4\nC: T1 -> T2 -> T4 -> T3\nD: T1 -> T4 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ab2188839eca588e14ececd553c05ab309b3db806f153f7d78edfdf9250a16d7.jpg", "reference_inputs/images/72cd0a8adb1640bf26b9f9f7b03503233379e3d16e3754adda473802969f4c0f.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000063", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000063", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [480, 446, 495, 462]\n- T2: [417, 503, 445, 540]\n- T3: [478, 362, 490, 381]\n\nImage 2 (GROUND) labeled targets:\n- T1: [483, 282, 497, 290]\n- T2: [436, 276, 454, 292]\n- T3: [481, 274, 492, 287]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T1 -> T2 -> T3\nC: T2 -> T3 -> T1\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7abe8bf0e8316b5c50cf7a1af93b7def29fa22d68678887ccb5eace7bbf814b0.jpg", "reference_inputs/images/1ac8ae79665ede165e788f39ca756a98d62b27c8dc1a0d5c67b311d5992b3466.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000064", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000064", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [417, 503, 445, 540]\n- T2: [480, 446, 495, 462]\n- T3: [478, 362, 490, 381]\n\nImage 2 (GROUND) labeled targets:\n- T1: [436, 276, 454, 292]\n- T2: [483, 282, 497, 290]\n- T3: [481, 274, 492, 287]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2 -> T3\nC: T1 -> T3 -> T2\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7abe8bf0e8316b5c50cf7a1af93b7def29fa22d68678887ccb5eace7bbf814b0.jpg", "reference_inputs/images/1ac8ae79665ede165e788f39ca756a98d62b27c8dc1a0d5c67b311d5992b3466.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000065", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000065", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [120, 549, 157, 574]\n- T2: [478, 276, 563, 491]\n- T3: [304, 387, 358, 500]\n\nImage 2 (GROUND) labeled targets:\n- T1: [614, 272, 621, 296]\n- T2: [337, 241, 417, 302]\n- T3: [498, 278, 528, 300]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T2 -> T1 -> T3\nC: T3 -> T1 -> T2\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1dd5e7838e2c03695637a18be71970132c07aa954e052efb5afe736e1b8ede06.jpg", "reference_inputs/images/772ae829aab097e813edc986df6337aa53a4dfb46188b630cff56caf74228b46.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000066", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000066", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [120, 549, 157, 574]\n- T2: [304, 387, 358, 500]\n- T3: [478, 276, 563, 491]\n\nImage 2 (GROUND) labeled targets:\n- T1: [614, 272, 621, 296]\n- T2: [498, 278, 528, 300]\n- T3: [337, 241, 417, 302]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T3 -> T2 -> T1\nC: T1 -> T3 -> T2\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1dd5e7838e2c03695637a18be71970132c07aa954e052efb5afe736e1b8ede06.jpg", "reference_inputs/images/772ae829aab097e813edc986df6337aa53a4dfb46188b630cff56caf74228b46.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000067", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000067", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [469, 435, 486, 450]\n- T2: [540, 489, 560, 514]\n- T3: [497, 406, 516, 434]\n\nImage 2 (GROUND) labeled targets:\n- T1: [463, 282, 483, 291]\n- T2: [545, 277, 570, 294]\n- T3: [496, 265, 517, 290]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T1 -> T3\nC: T1 -> T3 -> T2\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/977acfa7137e557053b937cdac0f694f0400cda381f38b0fcb1b44864ca8defa.jpg", "reference_inputs/images/bcdc1e69c0a36fdd6f81b6dc138bac8948640e1f401b454a6715a6941bd5a4a5.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000068", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000068", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [710, 375, 745, 390]\n- T2: [746, 219, 769, 230]\n- T3: [526, 489, 582, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [898, 263, 913, 282]\n- T2: [892, 258, 901, 287]\n- T3: [1005, 274, 1023, 288]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T1 -> T2 -> T3\nC: T1 -> T3 -> T2\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1ee50cfdaac707213827426bf2a39ffa6b1fbf9a114cbcb4670402f1cb9ae1fc.jpg", "reference_inputs/images/deab17ec8340bbe5a1613810b85eed3465824705baa86b6815e514c648975542.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000069", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000069", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [52, 508, 118, 529]\n- T2: [712, 524, 773, 549]\n- T3: [550, 470, 556, 488]\n- T4: [385, 505, 432, 525]\n\nImage 2 (GROUND) labeled targets:\n- T1: [0, 274, 50, 314]\n- T2: [216, 280, 233, 290]\n- T3: [82, 271, 88, 290]\n- T4: [130, 279, 162, 295]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T4 -> T2 -> T3\nC: T1 -> T3 -> T4 -> T2\nD: T1 -> T4 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7585f41fe0b84cff321aa7f677c0cca23c67433ce4c183eb9baafe39b873f4d3.jpg", "reference_inputs/images/6b5ca664a0bec76fcb4e83634ea93c39a1cd48cb0e2e19cac5000ed55fa74d96.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000070", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000070", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [550, 470, 556, 488]\n- T2: [385, 505, 432, 525]\n- T3: [52, 508, 118, 529]\n\nImage 2 (GROUND) labeled targets:\n- T1: [82, 271, 88, 290]\n- T2: [130, 279, 162, 295]\n- T3: [0, 274, 50, 314]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1 -> T3\nC: T1 -> T2 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7585f41fe0b84cff321aa7f677c0cca23c67433ce4c183eb9baafe39b873f4d3.jpg", "reference_inputs/images/6b5ca664a0bec76fcb4e83634ea93c39a1cd48cb0e2e19cac5000ed55fa74d96.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000071", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000071", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [408, 545, 437, 575]\n- T2: [440, 463, 462, 480]\n- T3: [482, 426, 498, 445]\n\nImage 2 (GROUND) labeled targets:\n- T1: [437, 280, 456, 292]\n- T2: [456, 282, 471, 290]\n- T3: [488, 279, 499, 289]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T1 -> T3 -> T2\nC: T2 -> T3 -> T1\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cc654abf3625cebb4fb1ef067790e12e02268773b72d9deb070fed6671b3d91a.jpg", "reference_inputs/images/502fe2b7bf5d1fe5c44e96178c57cdc8f6cda7fad2cf013c6853ede35a442838.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000072", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000072", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [408, 545, 437, 575]\n- T2: [440, 463, 462, 480]\n- T3: [482, 426, 498, 445]\n\nImage 2 (GROUND) labeled targets:\n- T1: [437, 280, 456, 292]\n- T2: [456, 282, 471, 290]\n- T3: [488, 279, 499, 289]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T1 -> T2 -> T3\nC: T2 -> T3 -> T1\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cc654abf3625cebb4fb1ef067790e12e02268773b72d9deb070fed6671b3d91a.jpg", "reference_inputs/images/502fe2b7bf5d1fe5c44e96178c57cdc8f6cda7fad2cf013c6853ede35a442838.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000073", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000073", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [42, 446, 91, 467]\n- T2: [570, 0, 612, 83]\n- T3: [148, 172, 192, 189]\n\nImage 2 (GROUND) labeled targets:\n- T1: [36, 250, 58, 298]\n- T2: [542, 275, 570, 296]\n- T3: [278, 262, 285, 294]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T1 -> T2 -> T3\nC: T3 -> T1 -> T2\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/80411caf2c694c923da91acf9ce65757d6d0b3cc4e9f32a494dc213f2e0988f9.jpg", "reference_inputs/images/9ee52d4b590978577fe3a21b750b8093f46ed3b4efce315bf8098e1feb0e6325.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000074", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000074", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [146, 413, 180, 436]\n- T2: [42, 446, 91, 467]\n- T3: [570, 0, 612, 83]\n\nImage 2 (GROUND) labeled targets:\n- T1: [166, 256, 172, 299]\n- T2: [36, 250, 58, 298]\n- T3: [542, 275, 570, 296]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T2 -> T1\nC: T3 -> T1 -> T2\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/80411caf2c694c923da91acf9ce65757d6d0b3cc4e9f32a494dc213f2e0988f9.jpg", "reference_inputs/images/9ee52d4b590978577fe3a21b750b8093f46ed3b4efce315bf8098e1feb0e6325.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000075", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000075", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [189, 59, 231, 81]\n- T2: [320, 178, 378, 291]\n- T3: [145, 450, 179, 476]\n\nImage 2 (GROUND) labeled targets:\n- T1: [721, 237, 750, 312]\n- T2: [216, 259, 617, 360]\n- T3: [111, 242, 143, 310]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T2 -> T1\nC: T1 -> T2 -> T3\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d47c774f2e316eb76158b6303c0ca7e885e89db572b755d0cb89296d52521cfc.jpg", "reference_inputs/images/65afc4ef610accaede802be45ab069bc460e54708884cbaffae2455de23764a6.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000076", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000076", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [438, 530, 443, 558]\n- T2: [555, 530, 590, 575]\n- T3: [410, 517, 417, 542]\n\nImage 2 (GROUND) labeled targets:\n- T1: [56, 266, 65, 293]\n- T2: [163, 281, 230, 298]\n- T3: [38, 267, 47, 293]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T3 -> T2 -> T1\nC: T1 -> T2 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cfd3ae4956d0d912e42e8b069a600698afb87d5f906b5079bc3f973feb58a839.jpg", "reference_inputs/images/8f6e1d91122c91578b18dffedaa92d09784d3e28552882ecd25afe2c34326083.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000079", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000079", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [454, 428, 470, 447]\n- T2: [549, 555, 576, 575]\n- T3: [650, 502, 659, 524]\n- T4: [494, 430, 516, 466]\n\nImage 2 (GROUND) labeled targets:\n- T1: [443, 277, 462, 291]\n- T2: [563, 272, 606, 300]\n- T3: [690, 266, 703, 293]\n- T4: [491, 258, 518, 292]\n\nAnswer options:\nA: T2 -> T3 -> T4 -> T1\nB: T3 -> T2 -> T4 -> T1\nC: T2 -> T3 -> T1 -> T4\nD: T2 -> T4 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0338e6668ccc2980e70ebd4a9e32d2deccdc84389a69907b91810ac1184821af.jpg", "reference_inputs/images/34d027dbab11397f568f524e93a1358bb23b7502da274b0c04a1dba07548b837.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000080", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000080", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [650, 502, 659, 524]\n- T2: [454, 428, 470, 447]\n- T3: [458, 468, 480, 490]\n- T4: [549, 555, 576, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [690, 266, 703, 293]\n- T2: [443, 277, 462, 291]\n- T3: [446, 281, 474, 294]\n- T4: [563, 272, 606, 300]\n\nAnswer options:\nA: T4 -> T3 -> T1 -> T2\nB: T1 -> T4 -> T3 -> T2\nC: T4 -> T1 -> T3 -> T2\nD: T4 -> T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0338e6668ccc2980e70ebd4a9e32d2deccdc84389a69907b91810ac1184821af.jpg", "reference_inputs/images/34d027dbab11397f568f524e93a1358bb23b7502da274b0c04a1dba07548b837.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000082", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000082", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [530, 402, 543, 411]\n- T2: [534, 425, 549, 440]\n- T3: [462, 410, 477, 418]\n\nImage 2 (GROUND) labeled targets:\n- T1: [577, 277, 637, 302]\n- T2: [686, 202, 1000, 350]\n- T3: [274, 275, 367, 307]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T1 -> T3\nC: T2 -> T3 -> T1\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1c3d4109a38cb72dd2d349b03eb89d2366a1cfb46285ae129959aae657de9447.jpg", "reference_inputs/images/e8f86a144ff07fab1b6ed4595caeffbf95d4782f342c62d61c804499bb0065bf.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000083", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000083", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [530, 402, 543, 411]\n- T2: [534, 425, 549, 440]\n- T3: [462, 410, 477, 418]\n\nImage 2 (GROUND) labeled targets:\n- T1: [577, 277, 637, 302]\n- T2: [686, 202, 1000, 350]\n- T3: [274, 275, 367, 307]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T2 -> T3 -> T1\nC: T1 -> T2 -> T3\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1c3d4109a38cb72dd2d349b03eb89d2366a1cfb46285ae129959aae657de9447.jpg", "reference_inputs/images/e8f86a144ff07fab1b6ed4595caeffbf95d4782f342c62d61c804499bb0065bf.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000084", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000084", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [650, 519, 655, 542]\n- T2: [423, 481, 449, 512]\n- T3: [635, 562, 641, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [914, 262, 924, 294]\n- T2: [595, 274, 626, 294]\n- T3: [891, 258, 902, 297]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T1 -> T3 -> T2\nC: T2 -> T3 -> T1\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6d76248e7087fb5fcf6decf03cdcb4e685fdbb05d11f13983b35973a1a64e154.jpg", "reference_inputs/images/74f386330c7d9af7076f742a7d66db3df65f2fcc5027d9e6f0c2bb2d6c20ab41.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000085", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000085", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [635, 562, 641, 575]\n- T2: [534, 490, 555, 514]\n- T3: [650, 519, 655, 542]\n\nImage 2 (GROUND) labeled targets:\n- T1: [891, 258, 902, 297]\n- T2: [740, 276, 763, 295]\n- T3: [914, 262, 924, 294]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T2 -> T3\nC: T1 -> T3 -> T2\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6d76248e7087fb5fcf6decf03cdcb4e685fdbb05d11f13983b35973a1a64e154.jpg", "reference_inputs/images/74f386330c7d9af7076f742a7d66db3df65f2fcc5027d9e6f0c2bb2d6c20ab41.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000087", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000087", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [508, 518, 530, 546]\n- T2: [273, 557, 280, 575]\n- T3: [531, 416, 544, 428]\n\nImage 2 (GROUND) labeled targets:\n- T1: [507, 278, 535, 297]\n- T2: [182, 259, 198, 295]\n- T3: [535, 280, 549, 290]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T1 -> T2 -> T3\nC: T2 -> T1 -> T3\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1c4a984fa871acdb7a4fcedb40fb895e0a1f4c3d65635ac5eb19c51fb3d8efe8.jpg", "reference_inputs/images/dbdbafd1873f8374fc2da3b5dc8e39767781e722be8ed25fdcf2da41eac4a224.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000088", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000088", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [531, 416, 544, 428]\n- T2: [508, 518, 530, 546]\n- T3: [273, 557, 280, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [535, 280, 549, 290]\n- T2: [507, 278, 535, 297]\n- T3: [182, 259, 198, 295]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T3 -> T1 -> T2\nC: T3 -> T2 -> T1\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1c4a984fa871acdb7a4fcedb40fb895e0a1f4c3d65635ac5eb19c51fb3d8efe8.jpg", "reference_inputs/images/dbdbafd1873f8374fc2da3b5dc8e39767781e722be8ed25fdcf2da41eac4a224.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000089", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000089", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [505, 392, 521, 416]\n- T2: [476, 457, 494, 482]\n- T3: [510, 448, 525, 465]\n\nImage 2 (GROUND) labeled targets:\n- T1: [494, 239, 539, 298]\n- T2: [0, 171, 350, 426]\n- T3: [497, 247, 615, 334]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T2 -> T3 -> T1\nC: T2 -> T1 -> T3\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3e21d07491e09cd36408a4cbcdacf1e77d7462660023e5288805dc80014a0207.jpg", "reference_inputs/images/423edee79517840c130357a8a1a08ab8664be9cec297a266fc992caea0152d92.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000090", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000090", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [476, 457, 494, 482]\n- T2: [510, 448, 525, 465]\n- T3: [505, 392, 521, 416]\n\nImage 2 (GROUND) labeled targets:\n- T1: [0, 171, 350, 426]\n- T2: [497, 247, 615, 334]\n- T3: [494, 239, 539, 298]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T2 -> T3\nC: T1 -> T3 -> T2\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3e21d07491e09cd36408a4cbcdacf1e77d7462660023e5288805dc80014a0207.jpg", "reference_inputs/images/423edee79517840c130357a8a1a08ab8664be9cec297a266fc992caea0152d92.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000091", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000091", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [914, 390, 960, 406]\n- T2: [654, 340, 721, 456]\n- T3: [843, 219, 879, 233]\n\nImage 2 (GROUND) labeled targets:\n- T1: [350, 264, 370, 298]\n- T2: [67, 276, 162, 309]\n- T3: [248, 263, 258, 297]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T1 -> T3 -> T2\nC: T1 -> T2 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7d3719bc63e8a194e6b5bade4d3c3bfbb3b34170c480ec70ad99fbee5cfcb1c4.jpg", "reference_inputs/images/16652bd8a0749e2b8903fd22753cff94fa7d4fe455aeb70555cc2ffe81a55955.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000092", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000092", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [773, 402, 804, 424]\n- T2: [119, 438, 156, 455]\n- T3: [314, 238, 376, 350]\n\nImage 2 (GROUND) labeled targets:\n- T1: [737, 260, 747, 297]\n- T2: [174, 258, 188, 298]\n- T3: [365, 278, 415, 299]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T1 -> T2\nC: T3 -> T2 -> T1\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5efc52003cf953f94e3968827080733c9c7d3c87a5365c1c13b75ff2f9a21f03.jpg", "reference_inputs/images/1c86cc5951e53a469d0dfa1d49811abb2200a75b723703e0b03968751a5c81cc.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000093", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000093", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [441, 515, 464, 546]\n- T2: [434, 472, 457, 501]\n- T3: [503, 423, 517, 438]\n- T4: [686, 549, 691, 574]\n\nImage 2 (GROUND) labeled targets:\n- T1: [368, 274, 422, 303]\n- T2: [388, 269, 422, 289]\n- T3: [499, 275, 520, 292]\n- T4: [906, 243, 924, 301]\n\nAnswer options:\nA: T4 -> T1 -> T2 -> T3\nB: T4 -> T2 -> T1 -> T3\nC: T4 -> T1 -> T3 -> T2\nD: T1 -> T4 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/369d8c5a94fad57873a397e17382502d73f237070752a22f9e7d59767de6f1e1.jpg", "reference_inputs/images/52169b7a0b342f2d63699bcc716256cbbe071a240289254b98442a1483faace4.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000094", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000094", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [686, 549, 691, 574]\n- T2: [503, 423, 517, 438]\n- T3: [434, 472, 457, 501]\n- T4: [441, 515, 464, 546]\n\nImage 2 (GROUND) labeled targets:\n- T1: [906, 243, 924, 301]\n- T2: [499, 275, 520, 292]\n- T3: [388, 269, 422, 289]\n- T4: [368, 274, 422, 303]\n\nAnswer options:\nA: T1 -> T4 -> T2 -> T3\nB: T1 -> T3 -> T4 -> T2\nC: T4 -> T1 -> T3 -> T2\nD: T1 -> T4 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/369d8c5a94fad57873a397e17382502d73f237070752a22f9e7d59767de6f1e1.jpg", "reference_inputs/images/52169b7a0b342f2d63699bcc716256cbbe071a240289254b98442a1483faace4.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000095", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000095", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [500, 406, 551, 518]\n- T2: [52, 151, 97, 170]\n- T3: [325, 0, 378, 54]\n\nImage 2 (GROUND) labeled targets:\n- T1: [495, 247, 573, 315]\n- T2: [166, 258, 178, 290]\n- T3: [390, 279, 431, 297]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T1 -> T2 -> T3\nC: unable to determine\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/686afe55c2fa06299ea2faa81211ce9eaa6045b06db4b14b5dbd4117d79dc52f.jpg", "reference_inputs/images/176edff7c377ecfb0a85bbb5fe6f909debb69e63dd867a4d9757b6b09591aec1.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000096", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000096", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [500, 406, 551, 518]\n- T2: [325, 0, 378, 54]\n- T3: [52, 151, 97, 170]\n\nImage 2 (GROUND) labeled targets:\n- T1: [495, 247, 573, 315]\n- T2: [390, 279, 431, 297]\n- T3: [166, 258, 178, 290]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T1 -> T2 -> T3\nC: T2 -> T1 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/686afe55c2fa06299ea2faa81211ce9eaa6045b06db4b14b5dbd4117d79dc52f.jpg", "reference_inputs/images/176edff7c377ecfb0a85bbb5fe6f909debb69e63dd867a4d9757b6b09591aec1.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000097", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000097", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [467, 402, 480, 411]\n- T2: [470, 466, 490, 486]\n- T3: [508, 430, 523, 446]\n\nImage 2 (GROUND) labeled targets:\n- T1: [406, 280, 442, 295]\n- T2: [34, 258, 370, 360]\n- T3: [500, 258, 562, 306]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T3 -> T1\nC: T2 -> T1 -> T3\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e0c003ebe69411cc06a9beec7f3c8530af13d5a5add78963e2479095c488c6ba.jpg", "reference_inputs/images/4b219c2e1a5ddc8d0638c9babe4ad1646d1f1623be3f4678702dfca40e824bbc.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000098", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000098", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [446, 446, 465, 470]\n- T2: [466, 390, 478, 398]\n- T3: [515, 399, 528, 407]\n\nImage 2 (GROUND) labeled targets:\n- T1: [0, 230, 240, 339]\n- T2: [413, 281, 442, 294]\n- T3: [520, 280, 551, 295]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T3 -> T2\nC: T2 -> T1 -> T3\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/79c7088a1f08249b13cfb452a3aa80cf0acb81937c2f6e8724e40fe57ddd7298.jpg", "reference_inputs/images/140c37b0862350acdc31ccd5823d77a3406f34d820549a19c0ff03742502e98b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000099", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000099", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [446, 446, 465, 470]\n- T2: [466, 390, 478, 398]\n- T3: [515, 399, 528, 407]\n\nImage 2 (GROUND) labeled targets:\n- T1: [0, 230, 240, 339]\n- T2: [413, 281, 442, 294]\n- T3: [520, 280, 551, 295]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T3 -> T1 -> T2\nC: T1 -> T2 -> T3\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/79c7088a1f08249b13cfb452a3aa80cf0acb81937c2f6e8724e40fe57ddd7298.jpg", "reference_inputs/images/140c37b0862350acdc31ccd5823d77a3406f34d820549a19c0ff03742502e98b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000100", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000100", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [446, 446, 463, 464]\n- T2: [674, 554, 683, 575]\n- T3: [466, 391, 490, 399]\n\nImage 2 (GROUND) labeled targets:\n- T1: [435, 278, 455, 292]\n- T2: [729, 259, 740, 295]\n- T3: [462, 283, 482, 289]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T3 -> T1\nC: T2 -> T1 -> T3\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/54a45b98db6cf9dcae184341b11c3ada9efb63b5c8c0222a21ea269ead7746f1.jpg", "reference_inputs/images/389a42f43ad667172d6855c4182d3f7f74d9b73cea54b3b04b85012cce008507.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000101", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000101", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [422, 522, 429, 546]\n- T2: [512, 449, 529, 472]\n- T3: [501, 501, 522, 530]\n- T4: [554, 425, 567, 436]\n\nImage 2 (GROUND) labeled targets:\n- T1: [389, 264, 397, 294]\n- T2: [522, 275, 534, 293]\n- T3: [497, 274, 526, 297]\n- T4: [562, 281, 578, 291]\n\nAnswer options:\nA: T1 -> T2 -> T3 -> T4\nB: T3 -> T1 -> T2 -> T4\nC: T1 -> T3 -> T2 -> T4\nD: T1 -> T3 -> T4 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/29a6541a1c03ccfac6d073b89565f4d67321d2e2acdf67f5fe0a474b8c507135.jpg", "reference_inputs/images/314d01bbfb4d254b346e7f6bc4fb62da813832c57ce8b979e13a5d0aae67e2bd.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000102", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000102", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [422, 522, 429, 546]\n- T2: [501, 501, 522, 530]\n- T3: [554, 425, 567, 436]\n- T4: [512, 449, 529, 472]\n\nImage 2 (GROUND) labeled targets:\n- T1: [389, 264, 397, 294]\n- T2: [497, 274, 526, 297]\n- T3: [562, 281, 578, 291]\n- T4: [522, 275, 534, 293]\n\nAnswer options:\nA: T1 -> T4 -> T2 -> T3\nB: T1 -> T2 -> T3 -> T4\nC: T1 -> T2 -> T4 -> T3\nD: T2 -> T1 -> T4 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/29a6541a1c03ccfac6d073b89565f4d67321d2e2acdf67f5fe0a474b8c507135.jpg", "reference_inputs/images/314d01bbfb4d254b346e7f6bc4fb62da813832c57ce8b979e13a5d0aae67e2bd.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000103", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000103", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [494, 431, 518, 469]\n- T2: [345, 467, 356, 483]\n- T3: [326, 512, 335, 535]\n\nImage 2 (GROUND) labeled targets:\n- T1: [488, 254, 522, 294]\n- T2: [274, 269, 288, 292]\n- T3: [216, 258, 228, 295]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T2 -> T3 -> T1\nC: T1 -> T3 -> T2\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ee076bd854493700b9fab044dd65fda55b8ff5f246f0fb232ded9339d9b3fc6c.jpg", "reference_inputs/images/15efbf7d748243da6419b1350bea59113e625f891f627c70f979a4ee745a68a8.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000104", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000104", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [326, 512, 335, 535]\n- T2: [345, 467, 356, 483]\n- T3: [494, 431, 518, 469]\n\nImage 2 (GROUND) labeled targets:\n- T1: [216, 258, 228, 295]\n- T2: [274, 269, 288, 292]\n- T3: [488, 254, 522, 294]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2 -> T3\nC: T2 -> T1 -> T3\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ee076bd854493700b9fab044dd65fda55b8ff5f246f0fb232ded9339d9b3fc6c.jpg", "reference_inputs/images/15efbf7d748243da6419b1350bea59113e625f891f627c70f979a4ee745a68a8.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000106", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000106", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [422, 522, 429, 546]\n- T2: [509, 402, 520, 412]\n- T3: [554, 425, 567, 436]\n- T4: [512, 449, 529, 472]\n\nImage 2 (GROUND) labeled targets:\n- T1: [152, 270, 158, 292]\n- T2: [384, 284, 398, 291]\n- T3: [408, 284, 421, 293]\n- T4: [325, 278, 351, 294]\n\nAnswer options:\nA: T1 -> T3 -> T4 -> T2\nB: T1 -> T4 -> T3 -> T2\nC: T4 -> T1 -> T3 -> T2\nD: T1 -> T4 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/29a6541a1c03ccfac6d073b89565f4d67321d2e2acdf67f5fe0a474b8c507135.jpg", "reference_inputs/images/006e4b97e40fb9fb2f043974ca480bf6e4fbbcfac6f82f579c3ab98b45c6c1f4.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000107", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000107", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [147, 526, 180, 559]\n- T2: [479, 298, 532, 406]\n- T3: [818, 76, 858, 97]\n- T4: [164, 222, 209, 237]\n\nImage 2 (GROUND) labeled targets:\n- T1: [100, 250, 118, 302]\n- T2: [478, 265, 529, 301]\n- T3: [694, 266, 700, 292]\n- T4: [284, 266, 304, 294]\n\nAnswer options:\nA: T1 -> T2 -> T4 -> T3\nB: T1 -> T2 -> T3 -> T4\nC: unable to determine\nD: T2 -> T1 -> T4 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b6e69bb6b20e30c6251d467415a026ad170366486c78c77a0625195e38cab5b5.jpg", "reference_inputs/images/214450fcb9ec8f693510b00eb2718a24d336b00136b94d6785fc63ff6db98864.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000108", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000108", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [479, 298, 532, 406]\n- T2: [164, 222, 209, 237]\n- T3: [147, 526, 180, 559]\n\nImage 2 (GROUND) labeled targets:\n- T1: [478, 265, 529, 301]\n- T2: [284, 266, 304, 294]\n- T3: [100, 250, 118, 302]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T2 -> T3\nC: T1 -> T3 -> T2\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b6e69bb6b20e30c6251d467415a026ad170366486c78c77a0625195e38cab5b5.jpg", "reference_inputs/images/214450fcb9ec8f693510b00eb2718a24d336b00136b94d6785fc63ff6db98864.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000109", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000109", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [761, 55, 789, 82]\n- T2: [873, 396, 918, 413]\n- T3: [874, 125, 920, 142]\n\nImage 2 (GROUND) labeled targets:\n- T1: [446, 220, 478, 316]\n- T2: [906, 247, 934, 308]\n- T3: [556, 245, 572, 306]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T1 -> T2 -> T3\nC: T1 -> T3 -> T2\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/94a42c416d039b03a48924175f0fe6b51d54b3a84c1a26fee68d14316a486a57.jpg", "reference_inputs/images/2506433c82e3e48ac5f1db08b9f7e3f477f766a7bfcdea59cc70da55f5e29026.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000112", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000112", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [754, 234, 781, 245]\n- T2: [315, 198, 374, 307]\n- T3: [363, 417, 412, 519]\n- T4: [775, 52, 808, 77]\n\nImage 2 (GROUND) labeled targets:\n- T1: [748, 256, 761, 298]\n- T2: [326, 262, 382, 305]\n- T3: [217, 264, 356, 322]\n- T4: [709, 260, 718, 295]\n\nAnswer options:\nA: T3 -> T2 -> T1 -> T4\nB: T3 -> T2 -> T4 -> T1\nC: unable to determine\nD: T3 -> T1 -> T2 -> T4\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/19b97407fa1dc9b3bb12ab0247fd4fa6b09c797c0452b5a47ff0ca4a0408ea8c.jpg", "reference_inputs/images/e071a3d5ffba4f7bb6addc5eae0d5ad70da90edbd8c3be2635e149f0f653c9c0.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000114", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000114", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [100, 124, 147, 145]\n- T2: [81, 514, 124, 538]\n- T3: [92, 368, 131, 383]\n\nImage 2 (GROUND) labeled targets:\n- T1: [523, 242, 542, 306]\n- T2: [38, 253, 56, 309]\n- T3: [210, 250, 231, 308]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T1 -> T2\nC: T1 -> T2 -> T3\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5227f1514bc26367b9a685962255b280ec8c82f911ad7282b2912b36f5b3dc4c.jpg", "reference_inputs/images/b4124de2733458d6b02f2a8d6824132a5a497241fa556c58ee1b573e171b45bc.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000115", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000115", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [942, 43, 992, 66]\n- T2: [327, 553, 345, 575]\n- T3: [652, 54, 673, 81]\n\nImage 2 (GROUND) labeled targets:\n- T1: [774, 267, 781, 293]\n- T2: [256, 242, 276, 306]\n- T3: [598, 266, 605, 294]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T3 -> T2 -> T1\nC: T2 -> T1 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b8792d16dea90b6ef0611eaee2ce16526d9ae336d106646657cea1b42a3c1d7a.jpg", "reference_inputs/images/3968266896f044d15bad291d7598cf1d78a7370419cdf14edbc6a4a4ab52fa2c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000116", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000116", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [162, 274, 183, 287]\n- T2: [100, 519, 138, 545]\n- T3: [183, 486, 216, 513]\n\nImage 2 (GROUND) labeled targets:\n- T1: [988, 238, 1017, 308]\n- T2: [590, 246, 602, 306]\n- T3: [657, 230, 686, 311]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T2 -> T1\nC: T1 -> T3 -> T2\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6af64a6f155a53fd7e9405fd0240f3d88f46f26297b3a634601d63726b1edce1.jpg", "reference_inputs/images/d3a0110b6a67af346162179241e854b0f3917a8527d17d0238e525236b325fc3.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000117", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000117", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [510, 401, 521, 411]\n- T2: [538, 455, 557, 474]\n- T3: [533, 422, 549, 434]\n\nImage 2 (GROUND) labeled targets:\n- T1: [509, 277, 532, 294]\n- T2: [624, 241, 754, 319]\n- T3: [565, 278, 612, 299]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T2 -> T1\nC: T2 -> T1 -> T3\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/df59edf618f1768dd5184af7d6da5fda60eb9c250beb66b6d42ac32de9770ce0.jpg", "reference_inputs/images/cbfef6891d3e74555360b427a6a7211fb069abf7f7d51a894e28baa461754599.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000118", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000118", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [533, 422, 549, 434]\n- T2: [454, 432, 471, 443]\n- T3: [538, 455, 557, 474]\n- T4: [510, 401, 521, 411]\n\nImage 2 (GROUND) labeled targets:\n- T1: [565, 278, 612, 299]\n- T2: [332, 278, 397, 302]\n- T3: [624, 241, 754, 319]\n- T4: [509, 277, 532, 294]\n\nAnswer options:\nA: T3 -> T2 -> T1 -> T4\nB: T3 -> T2 -> T4 -> T1\nC: T3 -> T1 -> T2 -> T4\nD: T2 -> T3 -> T1 -> T4\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/df59edf618f1768dd5184af7d6da5fda60eb9c250beb66b6d42ac32de9770ce0.jpg", "reference_inputs/images/cbfef6891d3e74555360b427a6a7211fb069abf7f7d51a894e28baa461754599.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_20m/000119", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000119", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [398, 87, 454, 201]\n- T2: [812, 295, 846, 313]\n- T3: [504, 364, 554, 476]\n- T4: [766, 0, 795, 29]\n\nImage 2 (GROUND) labeled targets:\n- T1: [439, 279, 477, 297]\n- T2: [746, 262, 757, 296]\n- T3: [506, 269, 550, 304]\n- T4: [650, 267, 657, 292]\n\nAnswer options:\nA: T3 -> T1 -> T2 -> T4\nB: T3 -> T2 -> T4 -> T1\nC: T3 -> T2 -> T1 -> T4\nD: T2 -> T3 -> T1 -> T4\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9b07624335b4fa633b68a1113aa88dc561c5ac1dcdfee58fdfa373ca97d99437.jpg", "reference_inputs/images/67d3c4dfaf5334baa46fdb18857f233075e29e32dcbad3903bb91dcb5950e373.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000001", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [290, 564, 302, 575]\n- T2: [270, 370, 293, 379]\n- T3: [418, 385, 456, 458]\n\nImage 2 (GROUND) labeled targets:\n- T1: [48, 244, 74, 284]\n- T2: [525, 240, 545, 310]\n- T3: [0, 245, 498, 404]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T3 -> T2 -> T1\nC: T2 -> T3 -> T1\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/edbffda0f2b67c463fd64cae867842b22130d1b4b709eeaf66f07cd968060492.jpg", "reference_inputs/images/f387777464b99b654288c0f5691bec7740a4e22e19afc495f2252450b68ef7d1.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000003", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [526, 434, 540, 456]\n- T2: [476, 410, 490, 432]\n- T3: [459, 472, 486, 485]\n- T4: [506, 570, 527, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [536, 261, 560, 279]\n- T2: [460, 267, 479, 290]\n- T3: [414, 281, 466, 294]\n- T4: [492, 274, 568, 311]\n\nAnswer options:\nA: T4 -> T1 -> T3 -> T2\nB: T4 -> T3 -> T1 -> T2\nC: T3 -> T4 -> T1 -> T2\nD: T4 -> T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f5158663b6493a36ff24cee09cee5d0205729b56889b7518818f61b53f9e3b4b.jpg", "reference_inputs/images/b683797c51a05883b7e76e4ff54291fc1eb45bd00558f4d84c85fd38534d3947.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000004", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [459, 472, 486, 485]\n- T2: [476, 410, 490, 432]\n- T3: [526, 434, 540, 456]\n- T4: [506, 570, 527, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [414, 281, 466, 294]\n- T2: [460, 267, 479, 290]\n- T3: [536, 261, 560, 279]\n- T4: [492, 274, 568, 311]\n\nAnswer options:\nA: T4 -> T1 -> T2 -> T3\nB: T4 -> T1 -> T3 -> T2\nC: T1 -> T4 -> T3 -> T2\nD: T4 -> T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f5158663b6493a36ff24cee09cee5d0205729b56889b7518818f61b53f9e3b4b.jpg", "reference_inputs/images/b683797c51a05883b7e76e4ff54291fc1eb45bd00558f4d84c85fd38534d3947.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000005", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [676, 166, 688, 178]\n- T2: [648, 447, 659, 465]\n- T3: [288, 195, 303, 208]\n\nImage 2 (GROUND) labeled targets:\n- T1: [718, 271, 725, 297]\n- T2: [865, 257, 882, 306]\n- T3: [380, 267, 383, 294]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T2 -> T3 -> T1\nC: T3 -> T2 -> T1\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/65ab20dcbba598480d730324f3c9ceffe765f5fccfc0bf381aeb02abec648f0f.jpg", "reference_inputs/images/9c6b3f29f4d9176113b4baac3a280ea580c2ac94190eae813932e337951513d2.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000008", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [446, 367, 479, 440]\n- T2: [697, 205, 716, 215]\n- T3: [507, 283, 539, 356]\n- T4: [678, 289, 690, 303]\n\nImage 2 (GROUND) labeled targets:\n- T1: [254, 235, 423, 333]\n- T2: [782, 255, 797, 298]\n- T3: [502, 261, 572, 314]\n- T4: [829, 243, 846, 301]\n\nAnswer options:\nA: T1 -> T4 -> T3 -> T2\nB: T1 -> T3 -> T4 -> T2\nC: T1 -> T3 -> T2 -> T4\nD: T3 -> T1 -> T4 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c236d9347f55c0c3ec5ba178d174cdfed7be3be0fb69abec27fac1b60b6cc6cf.jpg", "reference_inputs/images/64ecb14e96931fa12d63d7d52e97e6551d9ab8794bec873526ad42f1cf64fd39.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000009", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [275, 566, 294, 575]\n- T2: [288, 502, 311, 514]\n- T3: [76, 430, 110, 443]\n\nImage 2 (GROUND) labeled targets:\n- T1: [0, 242, 18, 312]\n- T2: [102, 247, 142, 314]\n- T3: [397, 260, 408, 300]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T1 -> T3 -> T2\nC: T2 -> T3 -> T1\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7ef8435af79d5039b6ea4551e5399f6861d16b59baa994c4a98a2e33c187dd61.jpg", "reference_inputs/images/a95ddbdca69a13640b9d9595bb317ae802f40659a5f39d4e97c1ceba66babbbb.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000010", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [288, 502, 311, 514]\n- T2: [76, 430, 110, 443]\n- T3: [275, 566, 294, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [102, 247, 142, 314]\n- T2: [397, 260, 408, 300]\n- T3: [0, 242, 18, 312]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: unable to determine\nC: T3 -> T1 -> T2\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7ef8435af79d5039b6ea4551e5399f6861d16b59baa994c4a98a2e33c187dd61.jpg", "reference_inputs/images/a95ddbdca69a13640b9d9595bb317ae802f40659a5f39d4e97c1ceba66babbbb.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000011", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [480, 491, 494, 513]\n- T2: [510, 448, 520, 458]\n- T3: [527, 525, 560, 543]\n\nImage 2 (GROUND) labeled targets:\n- T1: [351, 255, 439, 312]\n- T2: [508, 276, 534, 295]\n- T3: [651, 260, 887, 352]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T3 -> T1 -> T2\nC: T1 -> T2 -> T3\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6321bf478e6b07f7aac4caaad696c33ea7ba3bfde825bc08e5bfdc4c07385130.jpg", "reference_inputs/images/c33e1aeee82405ba47f657e32649e2e4e8d1e73b08d660d0fe811f4195c3ad2e.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000012", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [480, 491, 494, 513]\n- T2: [510, 448, 520, 458]\n- T3: [527, 525, 560, 543]\n\nImage 2 (GROUND) labeled targets:\n- T1: [351, 255, 439, 312]\n- T2: [508, 276, 534, 295]\n- T3: [651, 260, 887, 352]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T2 -> T3 -> T1\nC: T1 -> T3 -> T2\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6321bf478e6b07f7aac4caaad696c33ea7ba3bfde825bc08e5bfdc4c07385130.jpg", "reference_inputs/images/c33e1aeee82405ba47f657e32649e2e4e8d1e73b08d660d0fe811f4195c3ad2e.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000013", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [1002, 441, 1015, 450]\n- T2: [991, 387, 1005, 403]\n- T3: [956, 489, 976, 504]\n\nImage 2 (GROUND) labeled targets:\n- T1: [471, 280, 491, 293]\n- T2: [474, 272, 488, 282]\n- T3: [126, 270, 306, 321]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T3 -> T1 -> T2\nC: T1 -> T3 -> T2\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9dd94905f247910ef62de3c8e5a8092a01734659fe932b42faf400b0acc57c3a.jpg", "reference_inputs/images/170de95388c1beacd301d8b4deddad64ffae0d5a318f315340f7ecf49a4a5c07.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000014", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [991, 387, 1005, 403]\n- T2: [1002, 441, 1015, 450]\n- T3: [956, 489, 976, 504]\n\nImage 2 (GROUND) labeled targets:\n- T1: [474, 272, 488, 282]\n- T2: [471, 280, 491, 293]\n- T3: [126, 270, 306, 321]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T2 -> T1\nC: T3 -> T1 -> T2\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9dd94905f247910ef62de3c8e5a8092a01734659fe932b42faf400b0acc57c3a.jpg", "reference_inputs/images/170de95388c1beacd301d8b4deddad64ffae0d5a318f315340f7ecf49a4a5c07.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000015", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [470, 447, 481, 462]\n- T2: [475, 478, 490, 493]\n- T3: [499, 450, 514, 476]\n\nImage 2 (GROUND) labeled targets:\n- T1: [314, 262, 386, 304]\n- T2: [37, 261, 332, 344]\n- T3: [429, 170, 538, 314]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T3 -> T1 -> T2\nC: T2 -> T1 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6dbfd15899345da323d2701eef481ffd7f42d3c6927e0fa020b0dae7b21106da.jpg", "reference_inputs/images/dd84ec2bafb16bb213460e7a124814e17d0f7d9c46fed3791c34b8ea12eba9b2.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000016", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [499, 450, 514, 476]\n- T2: [470, 447, 481, 462]\n- T3: [475, 478, 490, 493]\n\nImage 2 (GROUND) labeled targets:\n- T1: [429, 170, 538, 314]\n- T2: [314, 262, 386, 304]\n- T3: [37, 261, 332, 344]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T3 -> T2 -> T1\nC: unable to determine\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6dbfd15899345da323d2701eef481ffd7f42d3c6927e0fa020b0dae7b21106da.jpg", "reference_inputs/images/dd84ec2bafb16bb213460e7a124814e17d0f7d9c46fed3791c34b8ea12eba9b2.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000018", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [473, 437, 484, 446]\n- T2: [510, 511, 523, 528]\n- T3: [448, 508, 473, 562]\n\nImage 2 (GROUND) labeled targets:\n- T1: [425, 281, 442, 293]\n- T2: [498, 264, 580, 320]\n- T3: [0, 18, 257, 336]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T2 -> T1\nC: unable to determine\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3b817beb70079de70a053c49a0c92810b0f54beca2a9e32a535454ce3b66b03f.jpg", "reference_inputs/images/34591fc1fa22610d2519ee5814ed778a1c518ffb2f2dfb3ab15e3acf8d937d50.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000019", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [510, 511, 523, 528]\n- T2: [473, 437, 484, 446]\n- T3: [448, 508, 473, 562]\n\nImage 2 (GROUND) labeled targets:\n- T1: [498, 264, 580, 320]\n- T2: [425, 281, 442, 293]\n- T3: [0, 18, 257, 336]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T2 -> T1\nC: T1 -> T3 -> T2\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3b817beb70079de70a053c49a0c92810b0f54beca2a9e32a535454ce3b66b03f.jpg", "reference_inputs/images/34591fc1fa22610d2519ee5814ed778a1c518ffb2f2dfb3ab15e3acf8d937d50.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000020", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [303, 357, 318, 372]\n- T2: [193, 135, 214, 148]\n- T3: [223, 521, 249, 538]\n\nImage 2 (GROUND) labeled targets:\n- T1: [474, 232, 513, 313]\n- T2: [859, 248, 880, 302]\n- T3: [183, 248, 206, 306]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T1 -> T2\nC: T1 -> T2 -> T3\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fb572fefaa415fd470d01e9a6463e9ffa0cc5d87df0fccf0127e204e2ffcf7ec.jpg", "reference_inputs/images/f5962d8efece47c0e8846eaabb2e96a8e06b461376cf4c7a035459a9384da9e9.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000021", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [545, 174, 622, 215]\n- T2: [504, 43, 512, 60]\n- T3: [0, 250, 125, 314]\n\nImage 2 (GROUND) labeled targets:\n- T1: [781, 238, 833, 258]\n- T2: [916, 218, 926, 246]\n- T3: [646, 227, 674, 255]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T1 -> T2 -> T3\nC: T2 -> T1 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9387fb07c1ea73e8176f603aa4bc4630b33ca3cd16a48c5c1865d623fa341a4d.jpg", "reference_inputs/images/a0ccf0f5e696281230600960e2603ce51a994ea07386f34ba1410ded71be1b4c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000022", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [0, 250, 125, 314]\n- T2: [504, 43, 512, 60]\n- T3: [545, 174, 622, 215]\n\nImage 2 (GROUND) labeled targets:\n- T1: [646, 227, 674, 255]\n- T2: [916, 218, 926, 246]\n- T3: [781, 238, 833, 258]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T3 -> T2 -> T1\nC: T3 -> T1 -> T2\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9387fb07c1ea73e8176f603aa4bc4630b33ca3cd16a48c5c1865d623fa341a4d.jpg", "reference_inputs/images/a0ccf0f5e696281230600960e2603ce51a994ea07386f34ba1410ded71be1b4c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000023", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [240, 231, 257, 245]\n- T2: [214, 203, 242, 212]\n- T3: [694, 231, 707, 245]\n- T4: [286, 77, 307, 90]\n\nImage 2 (GROUND) labeled targets:\n- T1: [162, 258, 173, 298]\n- T2: [152, 261, 163, 296]\n- T3: [753, 258, 765, 298]\n- T4: [304, 267, 320, 294]\n\nAnswer options:\nA: T3 -> T1 -> T2 -> T4\nB: T3 -> T2 -> T1 -> T4\nC: T1 -> T3 -> T2 -> T4\nD: T3 -> T1 -> T4 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5b432ec76c85bb5a57956a33adb0353060181098b08c679dfb6e3d264a7c7ff0.jpg", "reference_inputs/images/9a6e5f9f30ee237528c1c8bea12e3e2df67813b2f7eaf7df22f916e035be8caa.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000024", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000024", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [505, 466, 516, 480]\n- T2: [457, 561, 474, 575]\n- T3: [450, 519, 467, 545]\n\nImage 2 (GROUND) labeled targets:\n- T1: [499, 275, 520, 292]\n- T2: [368, 274, 422, 303]\n- T3: [388, 269, 422, 289]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T3 -> T2 -> T1\nC: T2 -> T1 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3bde83c90aa2675fda8182f60cf5101e9a061fcede4f93aed9f93b53dea81928.jpg", "reference_inputs/images/52169b7a0b342f2d63699bcc716256cbbe071a240289254b98442a1483faace4.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000025", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [505, 466, 516, 480]\n- T2: [457, 561, 474, 575]\n- T3: [450, 519, 467, 545]\n\nImage 2 (GROUND) labeled targets:\n- T1: [499, 275, 520, 292]\n- T2: [368, 274, 422, 303]\n- T3: [388, 269, 422, 289]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T3 -> T2 -> T1\nC: T2 -> T3 -> T1\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3bde83c90aa2675fda8182f60cf5101e9a061fcede4f93aed9f93b53dea81928.jpg", "reference_inputs/images/52169b7a0b342f2d63699bcc716256cbbe071a240289254b98442a1483faace4.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000026", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [270, 340, 286, 354]\n- T2: [440, 122, 473, 196]\n- T3: [548, 288, 602, 425]\n\nImage 2 (GROUND) labeled targets:\n- T1: [269, 263, 278, 294]\n- T2: [458, 277, 482, 294]\n- T3: [544, 235, 607, 298]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T3 -> T2 -> T1\nC: unable to determine\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7178e1b2d214b3c18b7b708a46724106443903040c88c0642dfca51bb5d6326b.jpg", "reference_inputs/images/aa247e83cbfa6a55e87f479c5d2e650e3f2266f2b3bb012623eb5ab140aa8602.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000027", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [522, 424, 537, 445]\n- T2: [543, 558, 560, 575]\n- T3: [481, 396, 493, 414]\n\nImage 2 (GROUND) labeled targets:\n- T1: [532, 258, 564, 292]\n- T2: [811, 225, 1023, 362]\n- T3: [460, 267, 479, 290]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T1 -> T2 -> T3\nC: T2 -> T3 -> T1\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f0897f975c4069cd51679c11c42f130c4256ac116bcb15922beda8701b62fa59.jpg", "reference_inputs/images/1b5c7d35d99b4d90fc9acbf22dad919f289e379585dda663df500ef2c2dbccf2.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000028", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [543, 558, 560, 575]\n- T2: [522, 424, 537, 445]\n- T3: [481, 396, 493, 414]\n\nImage 2 (GROUND) labeled targets:\n- T1: [811, 225, 1023, 362]\n- T2: [532, 258, 564, 292]\n- T3: [460, 267, 479, 290]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T2 -> T3\nC: T1 -> T3 -> T2\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f0897f975c4069cd51679c11c42f130c4256ac116bcb15922beda8701b62fa59.jpg", "reference_inputs/images/1b5c7d35d99b4d90fc9acbf22dad919f289e379585dda663df500ef2c2dbccf2.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000029", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [447, 526, 478, 575]\n- T2: [664, 341, 682, 350]\n- T3: [668, 494, 694, 509]\n- T4: [715, 294, 735, 302]\n\nImage 2 (GROUND) labeled targets:\n- T1: [414, 273, 464, 304]\n- T2: [654, 266, 662, 293]\n- T3: [706, 262, 738, 295]\n- T4: [685, 267, 691, 291]\n\nAnswer options:\nA: T1 -> T2 -> T3 -> T4\nB: T1 -> T3 -> T2 -> T4\nC: T1 -> T3 -> T4 -> T2\nD: T3 -> T1 -> T2 -> T4\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a1d1537e3be8850fe6f3702a692d2159724b057ca4725db2e331af7a19f6bdbf.jpg", "reference_inputs/images/fc2f455a3a134b67bdb86f8967c4e888682308e77817c4a1c0ede8e7b7ffb7cb.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000030", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [919, 384, 952, 394]\n- T2: [290, 117, 302, 126]\n- T3: [668, 450, 690, 462]\n\nImage 2 (GROUND) labeled targets:\n- T1: [534, 260, 545, 299]\n- T2: [19, 232, 66, 310]\n- T3: [766, 230, 810, 318]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T2 -> T1\nC: T3 -> T1 -> T2\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ea0b07874c421a30ce3bde9476a6b09e6eeb1f527b7870a89b564f5017a1476a.jpg", "reference_inputs/images/d396ca39d542235fbcc69b5858e5f5a773de1152513842df5f369577757d93ca.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000031", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000031", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [714, 558, 718, 575]\n- T2: [660, 470, 674, 484]\n- T3: [886, 537, 915, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [418, 270, 436, 292]\n- T2: [730, 281, 794, 294]\n- T3: [478, 184, 574, 311]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T1 -> T2\nC: T2 -> T3 -> T1\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f6477f2676e348ed6ac8c5e36ad5d03777705ee4cd07bf1376693b6eee8e1d2d.jpg", "reference_inputs/images/31c6abcf7eb4aa46f118f145da3d0ff14ff0c62da2378150b791ad7d61f0b10b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000032", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [714, 558, 718, 575]\n- T2: [886, 537, 915, 575]\n- T3: [660, 470, 674, 484]\n\nImage 2 (GROUND) labeled targets:\n- T1: [418, 270, 436, 292]\n- T2: [478, 184, 574, 311]\n- T3: [730, 281, 794, 294]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T1 -> T3 -> T2\nC: T1 -> T2 -> T3\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f6477f2676e348ed6ac8c5e36ad5d03777705ee4cd07bf1376693b6eee8e1d2d.jpg", "reference_inputs/images/31c6abcf7eb4aa46f118f145da3d0ff14ff0c62da2378150b791ad7d61f0b10b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000033", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000033", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [330, 448, 343, 463]\n- T2: [287, 222, 302, 236]\n- T3: [222, 90, 247, 104]\n- T4: [58, 530, 86, 546]\n\nImage 2 (GROUND) labeled targets:\n- T1: [0, 240, 10, 318]\n- T2: [607, 242, 642, 310]\n- T3: [827, 245, 853, 303]\n- T4: [200, 266, 210, 301]\n\nAnswer options:\nA: T2 -> T3 -> T1 -> T4\nB: T2 -> T1 -> T4 -> T3\nC: T2 -> T1 -> T3 -> T4\nD: T1 -> T2 -> T3 -> T4\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5503db5f5ae4225d2c6f3add36d2f86fa8ee0bd3d0e0cce90ad74ba26b5a2990.jpg", "reference_inputs/images/ee43fb8141aa4576e3ebdf22ef02f663d6e11cac7f7e49ef76189736638c9d2a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000034", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000034", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [287, 222, 302, 236]\n- T2: [58, 530, 86, 546]\n- T3: [254, 479, 272, 495]\n\nImage 2 (GROUND) labeled targets:\n- T1: [607, 242, 642, 310]\n- T2: [200, 266, 210, 301]\n- T3: [68, 244, 92, 310]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T3 -> T1 -> T2\nC: T1 -> T3 -> T2\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5503db5f5ae4225d2c6f3add36d2f86fa8ee0bd3d0e0cce90ad74ba26b5a2990.jpg", "reference_inputs/images/ee43fb8141aa4576e3ebdf22ef02f663d6e11cac7f7e49ef76189736638c9d2a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000035", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000035", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [479, 398, 492, 417]\n- T2: [507, 518, 524, 534]\n- T3: [466, 447, 490, 458]\n\nImage 2 (GROUND) labeled targets:\n- T1: [454, 266, 477, 290]\n- T2: [477, 268, 601, 327]\n- T3: [399, 280, 460, 295]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T2 -> T3 -> T1\nC: unable to determine\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7ca6b03f8931b1ab6f7da5d8418422538d148f0d692e684c2e49e2790acb8b6c.jpg", "reference_inputs/images/d805d4a2c5ab72663e1777357194e745b5bdfd434697b3d4f6a88ba3b05e17b3.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000036", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000036", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [466, 447, 490, 458]\n- T2: [507, 518, 524, 534]\n- T3: [479, 398, 492, 417]\n\nImage 2 (GROUND) labeled targets:\n- T1: [399, 280, 460, 295]\n- T2: [477, 268, 601, 327]\n- T3: [454, 266, 477, 290]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1 -> T3\nC: T2 -> T3 -> T1\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7ca6b03f8931b1ab6f7da5d8418422538d148f0d692e684c2e49e2790acb8b6c.jpg", "reference_inputs/images/d805d4a2c5ab72663e1777357194e745b5bdfd434697b3d4f6a88ba3b05e17b3.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000037", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000037", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [626, 26, 636, 47]\n- T2: [545, 174, 622, 215]\n- T3: [504, 43, 512, 60]\n\nImage 2 (GROUND) labeled targets:\n- T1: [350, 264, 370, 298]\n- T2: [67, 276, 162, 309]\n- T3: [248, 263, 258, 297]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T1 -> T2 -> T3\nC: T2 -> T1 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9387fb07c1ea73e8176f603aa4bc4630b33ca3cd16a48c5c1865d623fa341a4d.jpg", "reference_inputs/images/16652bd8a0749e2b8903fd22753cff94fa7d4fe455aeb70555cc2ffe81a55955.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000039", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000039", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [386, 266, 418, 333]\n- T2: [188, 251, 210, 262]\n- T3: [736, 230, 757, 239]\n\nImage 2 (GROUND) labeled targets:\n- T1: [359, 275, 406, 300]\n- T2: [178, 262, 188, 294]\n- T3: [744, 262, 754, 294]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T1 -> T2 -> T3\nC: T1 -> T3 -> T2\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/01bec1e6063f2abdaa5a2a82f0cc79873e8527ba96cf4d200ccc3916f0c1bc18.jpg", "reference_inputs/images/502fe2b7bf5d1fe5c44e96178c57cdc8f6cda7fad2cf013c6853ede35a442838.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000042", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000042", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [854, 75, 872, 89]\n- T2: [466, 20, 475, 42]\n- T3: [311, 161, 380, 193]\n\nImage 2 (GROUND) labeled targets:\n- T1: [938, 275, 945, 298]\n- T2: [696, 266, 718, 300]\n- T3: [769, 272, 895, 322]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T2 -> T1 -> T3\nC: T3 -> T2 -> T1\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7d1a04beac0a65e8d5644b0e96e0ffa9971c7899972461738dbebcdb4f5e7b4e.jpg", "reference_inputs/images/b40fb3da6d31f67dfc0380d6881e015176910045bdf7769a44ee54f580242034.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000043", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000043", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [670, 374, 683, 385]\n- T2: [417, 356, 455, 430]\n- T3: [280, 187, 295, 198]\n\nImage 2 (GROUND) labeled targets:\n- T1: [792, 272, 799, 298]\n- T2: [569, 283, 606, 300]\n- T3: [468, 274, 473, 294]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: unable to determine\nC: T1 -> T2 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b4d0eda699256237ed11a4a95c83ee5937f0838e00ca027a698437dc64ed5d94.jpg", "reference_inputs/images/47f9e862659644a60fa5865a905e3616ae64a6438419abb7aae6d257478f5f7c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000044", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000044", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [515, 446, 526, 455]\n- T2: [504, 390, 514, 405]\n- T3: [455, 463, 475, 501]\n\nImage 2 (GROUND) labeled targets:\n- T1: [519, 281, 548, 294]\n- T2: [498, 268, 517, 290]\n- T3: [290, 225, 402, 304]\n\nAnswer options:\nA: unable to determine\nB: T3 -> T1 -> T2\nC: T1 -> T3 -> T2\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a1c511f49696a445fff0f4cd0794fc2a08afe9322f7069f11eeea41c460a318f.jpg", "reference_inputs/images/abcd00b1ecf273170e58d374e635af5b0486c70f398e13ec84205b56af863cca.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000045", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000045", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [504, 390, 514, 405]\n- T2: [455, 463, 475, 501]\n- T3: [515, 446, 526, 455]\n\nImage 2 (GROUND) labeled targets:\n- T1: [498, 268, 517, 290]\n- T2: [290, 225, 402, 304]\n- T3: [519, 281, 548, 294]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T2 -> T1\nC: T2 -> T1 -> T3\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a1c511f49696a445fff0f4cd0794fc2a08afe9322f7069f11eeea41c460a318f.jpg", "reference_inputs/images/abcd00b1ecf273170e58d374e635af5b0486c70f398e13ec84205b56af863cca.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000047", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000047", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [850, 178, 874, 194]\n- T2: [880, 280, 954, 311]\n- T3: [670, 88, 682, 101]\n\nImage 2 (GROUND) labeled targets:\n- T1: [546, 256, 570, 302]\n- T2: [663, 278, 722, 306]\n- T3: [322, 225, 350, 318]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T3 -> T2\nC: T3 -> T1 -> T2\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0fef7bf43103cdecd4e2c7ba577149ffb74ee952c482f428f478d115ae197001.jpg", "reference_inputs/images/9dabade18b21c6aa67c58127d7a894f94f432a931037529764f20ab13e6c9cdc.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000048", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000048", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [850, 178, 874, 194]\n- T2: [670, 88, 682, 101]\n- T3: [880, 280, 954, 311]\n\nImage 2 (GROUND) labeled targets:\n- T1: [546, 256, 570, 302]\n- T2: [322, 225, 350, 318]\n- T3: [663, 278, 722, 306]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T2 -> T3 -> T1\nC: T2 -> T1 -> T3\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0fef7bf43103cdecd4e2c7ba577149ffb74ee952c482f428f478d115ae197001.jpg", "reference_inputs/images/9dabade18b21c6aa67c58127d7a894f94f432a931037529764f20ab13e6c9cdc.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000049", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000049", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [439, 438, 482, 514]\n- T2: [628, 194, 677, 271]\n- T3: [563, 153, 610, 231]\n\nImage 2 (GROUND) labeled targets:\n- T1: [222, 274, 267, 300]\n- T2: [417, 283, 445, 295]\n- T3: [370, 283, 401, 295]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T3 -> T1\nC: T1 -> T3 -> T2\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e5cdc2ffbe31b0bf6dc8b74abe10ba04fd31c30e5a07a413559f73a37a56dd6e.jpg", "reference_inputs/images/8d350c70f2db8cfcc325e55d17672ba8feba9ccb2d80d49bf961be1d03c18840.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000050", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000050", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [802, 566, 837, 575]\n- T2: [723, 475, 738, 490]\n- T3: [790, 392, 801, 407]\n- T4: [750, 374, 762, 387]\n\nImage 2 (GROUND) labeled targets:\n- T1: [550, 224, 630, 302]\n- T2: [436, 282, 453, 292]\n- T3: [522, 274, 534, 288]\n- T4: [484, 276, 495, 287]\n\nAnswer options:\nA: T1 -> T3 -> T2 -> T4\nB: T2 -> T1 -> T3 -> T4\nC: T1 -> T2 -> T3 -> T4\nD: T1 -> T2 -> T4 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/12ba627932e572d5a61af9cd22f432b21d00b9f8a92045064ca84ac488070f7c.jpg", "reference_inputs/images/068981882776608705076f4c4c3e998b66611be3d5507c1608daae3a21fad37c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000051", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000051", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [801, 454, 812, 463]\n- T2: [750, 374, 762, 387]\n- T3: [802, 566, 837, 575]\n- T4: [741, 525, 758, 547]\n\nImage 2 (GROUND) labeled targets:\n- T1: [0, 253, 105, 342]\n- T2: [484, 276, 495, 287]\n- T3: [550, 224, 630, 302]\n- T4: [450, 275, 478, 296]\n\nAnswer options:\nA: T3 -> T4 -> T1 -> T2\nB: T3 -> T4 -> T2 -> T1\nC: T3 -> T1 -> T4 -> T2\nD: T4 -> T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/12ba627932e572d5a61af9cd22f432b21d00b9f8a92045064ca84ac488070f7c.jpg", "reference_inputs/images/068981882776608705076f4c4c3e998b66611be3d5507c1608daae3a21fad37c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000052", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000052", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [235, 70, 250, 99]\n- T2: [242, 197, 265, 208]\n- T3: [674, 340, 686, 355]\n- T4: [548, 499, 581, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [333, 270, 350, 292]\n- T2: [318, 268, 322, 292]\n- T3: [667, 266, 674, 294]\n- T4: [564, 268, 621, 305]\n\nAnswer options:\nA: T4 -> T3 -> T1 -> T2\nB: T3 -> T4 -> T1 -> T2\nC: T4 -> T1 -> T3 -> T2\nD: T4 -> T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2942222141645b774a4b3bd6e5b7422af344083fdde6423f2b414362c1d9f968.jpg", "reference_inputs/images/75bb9f3ac1921d7fed1cce6bc9bbfa9956c3988c4fcf366d0c757f6c48b81399.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000053", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000053", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [434, 22, 445, 40]\n- T2: [334, 243, 396, 306]\n- T3: [338, 70, 354, 89]\n\nImage 2 (GROUND) labeled targets:\n- T1: [585, 258, 598, 296]\n- T2: [178, 271, 310, 307]\n- T3: [470, 263, 482, 294]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1 -> T3\nC: T2 -> T3 -> T1\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/17e8acdee8139296e74054ad46f0b8e4ac3f31e20e58c6ad512a0f49c246c007.jpg", "reference_inputs/images/29fc872b749e944ac734776f6892fe8e4600cf609cbd0a71f1ff2dda5f3dc4e4.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000054", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000054", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [474, 458, 522, 575]\n- T2: [654, 279, 666, 288]\n- T3: [680, 108, 696, 121]\n- T4: [280, 434, 300, 449]\n\nImage 2 (GROUND) labeled targets:\n- T1: [449, 198, 534, 308]\n- T2: [647, 265, 654, 294]\n- T3: [635, 270, 643, 291]\n- T4: [220, 256, 226, 297]\n\nAnswer options:\nA: T4 -> T1 -> T2 -> T3\nB: T1 -> T4 -> T2 -> T3\nC: T1 -> T2 -> T4 -> T3\nD: T1 -> T4 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b54de1266305b9b20f5b628edb6cc3be8bf303f5eaaa4ba4235d674b186ebab7.jpg", "reference_inputs/images/4183a0b6de334d0e851eef8ba64a5005f51b443781882f4bdd69e0315019eef6.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000057", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000057", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [466, 20, 475, 42]\n- T2: [311, 161, 380, 193]\n- T3: [564, 57, 574, 73]\n\nImage 2 (GROUND) labeled targets:\n- T1: [458, 256, 472, 298]\n- T2: [34, 266, 233, 317]\n- T3: [605, 250, 621, 299]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T1 -> T3\nC: T2 -> T3 -> T1\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7d1a04beac0a65e8d5644b0e96e0ffa9971c7899972461738dbebcdb4f5e7b4e.jpg", "reference_inputs/images/e8f1b9d03fe3b3b60044e05053b10d6c5713685735425b089f2fc508c67c0527.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000058", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000058", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [530, 466, 543, 477]\n- T2: [534, 503, 549, 521]\n- T3: [462, 476, 477, 488]\n- T4: [510, 442, 520, 451]\n\nImage 2 (GROUND) labeled targets:\n- T1: [577, 277, 637, 302]\n- T2: [686, 202, 1000, 350]\n- T3: [274, 275, 367, 307]\n- T4: [508, 275, 535, 296]\n\nAnswer options:\nA: T2 -> T3 -> T1 -> T4\nB: unable to determine\nC: T2 -> T1 -> T3 -> T4\nD: T3 -> T2 -> T1 -> T4\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4e33cdd8ea49b40c1dc79aa8a47c3247f0846a5bfc6bf23c11356d4b495d047c.jpg", "reference_inputs/images/e8f86a144ff07fab1b6ed4595caeffbf95d4782f342c62d61c804499bb0065bf.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000059", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000059", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [462, 476, 477, 488]\n- T2: [530, 466, 543, 477]\n- T3: [534, 503, 549, 521]\n- T4: [510, 442, 520, 451]\n\nImage 2 (GROUND) labeled targets:\n- T1: [274, 275, 367, 307]\n- T2: [577, 277, 637, 302]\n- T3: [686, 202, 1000, 350]\n- T4: [508, 275, 535, 296]\n\nAnswer options:\nA: T3 -> T1 -> T2 -> T4\nB: T1 -> T3 -> T2 -> T4\nC: T3 -> T1 -> T4 -> T2\nD: T3 -> T2 -> T1 -> T4\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4e33cdd8ea49b40c1dc79aa8a47c3247f0846a5bfc6bf23c11356d4b495d047c.jpg", "reference_inputs/images/e8f86a144ff07fab1b6ed4595caeffbf95d4782f342c62d61c804499bb0065bf.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000060", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000060", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [528, 463, 538, 476]\n- T2: [502, 413, 514, 432]\n- T3: [485, 433, 494, 442]\n\nImage 2 (GROUND) labeled targets:\n- T1: [548, 276, 576, 295]\n- T2: [495, 263, 518, 290]\n- T3: [459, 282, 481, 291]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T1 -> T3 -> T2\nC: T1 -> T2 -> T3\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b93a5eea86d8ed6ee483ca1841644a860eed2f66bd6ad287b7ff0aa36b80dc9f.jpg", "reference_inputs/images/c0d2a3e08cf989cab00b16bae02eabdb9d196be80eb7ba851f154ec5d4a68f11.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000061", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000061", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [528, 463, 538, 476]\n- T2: [485, 433, 494, 442]\n- T3: [502, 413, 514, 432]\n\nImage 2 (GROUND) labeled targets:\n- T1: [548, 276, 576, 295]\n- T2: [459, 282, 481, 291]\n- T3: [495, 263, 518, 290]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T1 -> T3\nC: T1 -> T3 -> T2\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b93a5eea86d8ed6ee483ca1841644a860eed2f66bd6ad287b7ff0aa36b80dc9f.jpg", "reference_inputs/images/c0d2a3e08cf989cab00b16bae02eabdb9d196be80eb7ba851f154ec5d4a68f11.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000062", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000062", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [215, 425, 242, 436]\n- T2: [279, 226, 299, 237]\n- T3: [474, 253, 528, 387]\n\nImage 2 (GROUND) labeled targets:\n- T1: [238, 266, 251, 294]\n- T2: [358, 270, 362, 291]\n- T3: [486, 249, 519, 295]\n\nAnswer options:\nA: unable to determine\nB: T3 -> T1 -> T2\nC: T1 -> T3 -> T2\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9800dd1985605e5316d25a285f75017cbfaced8522231d9c47b6dd8504e11fe3.jpg", "reference_inputs/images/955cbee518632600ad55301eeb8bda54362a5a47bf884e33762d6db37d7a90c0.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000063", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000063", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [279, 226, 299, 237]\n- T2: [474, 253, 528, 387]\n- T3: [215, 425, 242, 436]\n\nImage 2 (GROUND) labeled targets:\n- T1: [358, 270, 362, 291]\n- T2: [486, 249, 519, 295]\n- T3: [238, 266, 251, 294]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T1 -> T3\nC: T3 -> T2 -> T1\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9800dd1985605e5316d25a285f75017cbfaced8522231d9c47b6dd8504e11fe3.jpg", "reference_inputs/images/955cbee518632600ad55301eeb8bda54362a5a47bf884e33762d6db37d7a90c0.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000064", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000064", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [748, 237, 770, 248]\n- T2: [1, 180, 33, 191]\n- T3: [448, 466, 478, 534]\n- T4: [554, 0, 600, 98]\n\nImage 2 (GROUND) labeled targets:\n- T1: [774, 261, 782, 295]\n- T2: [26, 263, 38, 294]\n- T3: [334, 265, 435, 320]\n- T4: [536, 250, 579, 294]\n\nAnswer options:\nA: T3 -> T1 -> T4 -> T2\nB: T1 -> T3 -> T4 -> T2\nC: T3 -> T1 -> T2 -> T4\nD: T3 -> T4 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6e4af1e6834a23f481da3fb5481f4f060494ac2b58b1035dff9f3ad5ad3f2661.jpg", "reference_inputs/images/1f4e5bfc1dc07dd2fc6ab971db43666d4bee8f303a3194ecb836bf23599c8aa9.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000065", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000065", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [673, 50, 691, 64]\n- T2: [750, 0, 766, 15]\n- T3: [910, 325, 982, 358]\n\nImage 2 (GROUND) labeled targets:\n- T1: [27, 225, 74, 314]\n- T2: [70, 238, 99, 305]\n- T3: [622, 279, 692, 304]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: unable to determine\nC: T1 -> T3 -> T2\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d23def54bc38c6919f30f1cf05d51e32af4daf8d2413d5881837e1e5da81ad0b.jpg", "reference_inputs/images/eb97a3246d862464d2b51b720f73d209b8d3a46bb24e54d9e5a2b762b5dfa240.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000066", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000066", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [673, 50, 691, 64]\n- T2: [910, 325, 982, 358]\n- T3: [822, 108, 848, 122]\n\nImage 2 (GROUND) labeled targets:\n- T1: [27, 225, 74, 314]\n- T2: [622, 279, 692, 304]\n- T3: [354, 251, 371, 302]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T1 -> T2 -> T3\nC: T3 -> T1 -> T2\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d23def54bc38c6919f30f1cf05d51e32af4daf8d2413d5881837e1e5da81ad0b.jpg", "reference_inputs/images/eb97a3246d862464d2b51b720f73d209b8d3a46bb24e54d9e5a2b762b5dfa240.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000068", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000068", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [384, 58, 414, 134]\n- T2: [298, 210, 313, 220]\n- T3: [244, 234, 261, 241]\n\nImage 2 (GROUND) labeled targets:\n- T1: [370, 234, 734, 354]\n- T2: [218, 233, 246, 314]\n- T3: [239, 248, 254, 308]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T2 -> T3\nC: T1 -> T3 -> T2\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/223cbee8041d6f1a55407a61b86a321c281765434511cd82b74d02a28bad5c15.jpg", "reference_inputs/images/5c64e995683b1de21c283c671349dac82351d1b3a64375df3ccdb6bffaedb802.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000070", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000070", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [994, 395, 1023, 413]\n- T2: [849, 467, 922, 536]\n- T3: [347, 224, 476, 344]\n- T4: [898, 347, 924, 358]\n\nImage 2 (GROUND) labeled targets:\n- T1: [657, 253, 678, 302]\n- T2: [349, 272, 470, 317]\n- T3: [34, 245, 103, 297]\n- T4: [503, 258, 514, 298]\n\nAnswer options:\nA: T2 -> T1 -> T4 -> T3\nB: T2 -> T4 -> T1 -> T3\nC: T2 -> T1 -> T3 -> T4\nD: T1 -> T2 -> T4 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9a637268ccc66c4e35c4ed2e56daf521b49a57fefec9c91252cab14f54e236a3.jpg", "reference_inputs/images/6f0f1e07119f53abe8f08c28c43962d83b3debe85de2c64a6e89e21ebb95a02c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000071", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000071", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [994, 395, 1023, 413]\n- T2: [898, 347, 924, 358]\n- T3: [849, 467, 922, 536]\n- T4: [347, 224, 476, 344]\n\nImage 2 (GROUND) labeled targets:\n- T1: [657, 253, 678, 302]\n- T2: [503, 258, 514, 298]\n- T3: [349, 272, 470, 317]\n- T4: [34, 245, 103, 297]\n\nAnswer options:\nA: T4 -> T2 -> T3 -> T1\nB: T2 -> T4 -> T3 -> T1\nC: T4 -> T2 -> T1 -> T3\nD: T4 -> T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9a637268ccc66c4e35c4ed2e56daf521b49a57fefec9c91252cab14f54e236a3.jpg", "reference_inputs/images/6f0f1e07119f53abe8f08c28c43962d83b3debe85de2c64a6e89e21ebb95a02c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000072", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000072", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [430, 544, 458, 575]\n- T2: [502, 414, 515, 433]\n- T3: [516, 504, 532, 518]\n\nImage 2 (GROUND) labeled targets:\n- T1: [0, 63, 244, 329]\n- T2: [495, 264, 517, 290]\n- T3: [523, 278, 572, 301]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T2 -> T3\nC: T1 -> T3 -> T2\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ce00400318ec5a563e2db6cb9d771e702629e1ec0c577a36a04cd47af22eec3e.jpg", "reference_inputs/images/a19f7624e7d3322d2f1c949bac306d833f6743f5a7d12020b7abe157b706f1e0.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000073", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000073", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [289, 522, 307, 538]\n- T2: [382, 265, 421, 338]\n- T3: [257, 397, 274, 407]\n\nImage 2 (GROUND) labeled targets:\n- T1: [58, 245, 100, 314]\n- T2: [544, 258, 945, 358]\n- T3: [390, 247, 408, 309]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T3 -> T1 -> T2\nC: T2 -> T3 -> T1\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e5b411a68a7ef2e6b0baef87b94099e01703403ccf54d36ff3073d57e76ece61.jpg", "reference_inputs/images/6bde287cf09eccf13f1a7282d5f5bbecd74d66d92d3f8b16409fa292a063c87d.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000074", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000074", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [257, 397, 274, 407]\n- T2: [289, 522, 307, 538]\n- T3: [382, 265, 421, 338]\n\nImage 2 (GROUND) labeled targets:\n- T1: [390, 247, 408, 309]\n- T2: [58, 245, 100, 314]\n- T3: [544, 258, 945, 358]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T3 -> T1 -> T2\nC: T3 -> T2 -> T1\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e5b411a68a7ef2e6b0baef87b94099e01703403ccf54d36ff3073d57e76ece61.jpg", "reference_inputs/images/6bde287cf09eccf13f1a7282d5f5bbecd74d66d92d3f8b16409fa292a063c87d.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000075", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000075", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [258, 416, 275, 433]\n- T2: [370, 118, 410, 194]\n- T3: [286, 139, 314, 150]\n\nImage 2 (GROUND) labeled targets:\n- T1: [34, 250, 67, 310]\n- T2: [562, 260, 935, 350]\n- T3: [644, 242, 668, 281]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T3 -> T2 -> T1\nC: T2 -> T1 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/615f44861ee6ee580e557aa2ff2f6a04cff6e3258f6acafed97c60055df847c6.jpg", "reference_inputs/images/b7087cf8dfff20392c5cce16fb1cbc197f4b8a023fd6a1a14e127507670e1033.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000076", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000076", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [254, 274, 271, 284]\n- T2: [302, 148, 318, 158]\n- T3: [329, 242, 342, 252]\n\nImage 2 (GROUND) labeled targets:\n- T1: [425, 242, 454, 308]\n- T2: [730, 234, 758, 311]\n- T3: [489, 222, 520, 316]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T2 -> T1 -> T3\nC: T3 -> T2 -> T1\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ef9daea12006a2e51e5c982a5b280c67b50689e9b2b397d8028d0f4166e33386.jpg", "reference_inputs/images/82e49c092a52ba14ad4d6e963fee05e3b260013b8eddacfdcbdcbc606ff915ef.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000077", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000077", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [38, 148, 69, 166]\n- T2: [42, 540, 72, 564]\n- T3: [239, 158, 262, 170]\n\nImage 2 (GROUND) labeled targets:\n- T1: [598, 263, 615, 298]\n- T2: [146, 266, 166, 301]\n- T3: [646, 242, 668, 306]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T2 -> T1\nC: T1 -> T3 -> T2\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0fef7bf43103cdecd4e2c7ba577149ffb74ee952c482f428f478d115ae197001.jpg", "reference_inputs/images/a6bbb87b287f5401a21be4253d6a9ba6d1b1fb7dd5a4446eec287e86133e81c1.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000078", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000078", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [0, 461, 66, 500]\n- T2: [239, 158, 262, 170]\n- T3: [38, 148, 69, 166]\n\nImage 2 (GROUND) labeled targets:\n- T1: [222, 283, 290, 303]\n- T2: [646, 242, 668, 306]\n- T3: [598, 263, 615, 298]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T2 -> T1 -> T3\nC: T2 -> T3 -> T1\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0fef7bf43103cdecd4e2c7ba577149ffb74ee952c482f428f478d115ae197001.jpg", "reference_inputs/images/a6bbb87b287f5401a21be4253d6a9ba6d1b1fb7dd5a4446eec287e86133e81c1.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000079", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000079", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [674, 78, 686, 93]\n- T2: [686, 553, 702, 574]\n- T3: [286, 262, 310, 271]\n- T4: [719, 456, 734, 468]\n\nImage 2 (GROUND) labeled targets:\n- T1: [635, 270, 641, 292]\n- T2: [962, 226, 994, 310]\n- T3: [284, 266, 304, 294]\n- T4: [865, 248, 881, 299]\n\nAnswer options:\nA: T2 -> T3 -> T4 -> T1\nB: T4 -> T2 -> T3 -> T1\nC: T2 -> T4 -> T1 -> T3\nD: T2 -> T4 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cec587f7f71d59f91723be411e06f634b2c38d94eef8cefcfe273dea4be0b7c4.jpg", "reference_inputs/images/214450fcb9ec8f693510b00eb2718a24d336b00136b94d6785fc63ff6db98864.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000080", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000080", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [286, 262, 310, 271]\n- T2: [686, 553, 702, 574]\n- T3: [719, 456, 734, 468]\n- T4: [490, 310, 525, 382]\n\nImage 2 (GROUND) labeled targets:\n- T1: [284, 266, 304, 294]\n- T2: [962, 226, 994, 310]\n- T3: [865, 248, 881, 299]\n- T4: [478, 265, 529, 301]\n\nAnswer options:\nA: T1 -> T4 -> T3 -> T2\nB: T4 -> T3 -> T1 -> T2\nC: T4 -> T1 -> T2 -> T3\nD: T4 -> T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cec587f7f71d59f91723be411e06f634b2c38d94eef8cefcfe273dea4be0b7c4.jpg", "reference_inputs/images/214450fcb9ec8f693510b00eb2718a24d336b00136b94d6785fc63ff6db98864.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000081", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000081", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [749, 204, 767, 220]\n- T2: [673, 261, 690, 271]\n- T3: [910, 525, 981, 563]\n\nImage 2 (GROUND) labeled targets:\n- T1: [426, 241, 460, 307]\n- T2: [554, 243, 566, 294]\n- T3: [826, 281, 907, 306]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T2 -> T1 -> T3\nC: T1 -> T3 -> T2\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ef9daea12006a2e51e5c982a5b280c67b50689e9b2b397d8028d0f4166e33386.jpg", "reference_inputs/images/57c3f70f35fc88b642076e72a90f518e3004de43237b3e468d97c6500ead1958.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000082", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000082", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [910, 525, 981, 563]\n- T2: [749, 204, 767, 220]\n- T3: [673, 261, 690, 271]\n\nImage 2 (GROUND) labeled targets:\n- T1: [826, 281, 907, 306]\n- T2: [426, 241, 460, 307]\n- T3: [554, 243, 566, 294]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T1 -> T2\nC: T3 -> T2 -> T1\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ef9daea12006a2e51e5c982a5b280c67b50689e9b2b397d8028d0f4166e33386.jpg", "reference_inputs/images/57c3f70f35fc88b642076e72a90f518e3004de43237b3e468d97c6500ead1958.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000083", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000083", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [506, 456, 522, 485]\n- T2: [510, 537, 526, 560]\n- T3: [475, 554, 494, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [494, 239, 539, 298]\n- T2: [497, 247, 615, 334]\n- T3: [0, 171, 350, 426]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: unable to determine\nC: T3 -> T2 -> T1\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5f78e355e5e15fb436335e1dcade9259514f85cdeef39a550afcc8bcc9b80649.jpg", "reference_inputs/images/423edee79517840c130357a8a1a08ab8664be9cec297a266fc992caea0152d92.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000084", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000084", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [475, 554, 494, 575]\n- T2: [510, 537, 526, 560]\n- T3: [506, 456, 522, 485]\n\nImage 2 (GROUND) labeled targets:\n- T1: [0, 171, 350, 426]\n- T2: [497, 247, 615, 334]\n- T3: [494, 239, 539, 298]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T3 -> T2\nC: T1 -> T2 -> T3\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5f78e355e5e15fb436335e1dcade9259514f85cdeef39a550afcc8bcc9b80649.jpg", "reference_inputs/images/423edee79517840c130357a8a1a08ab8664be9cec297a266fc992caea0152d92.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000085", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000085", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [370, 324, 410, 398]\n- T2: [666, 248, 678, 255]\n- T3: [649, 350, 662, 360]\n- T4: [522, 426, 560, 501]\n\nImage 2 (GROUND) labeled targets:\n- T1: [932, 269, 986, 284]\n- T2: [704, 254, 710, 278]\n- T3: [722, 258, 730, 276]\n- T4: [789, 268, 819, 281]\n\nAnswer options:\nA: T3 -> T2 -> T1 -> T4\nB: T2 -> T3 -> T4 -> T1\nC: T2 -> T1 -> T3 -> T4\nD: T2 -> T3 -> T1 -> T4\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5dbeddd1c3c7318355a2bef03ca4c4171d13aa2262ccc2af44a86fbae0d714a3.jpg", "reference_inputs/images/db28015dd827c7a87eca54c8bed03ee4194eed92a3b8f6f40966f1f3a8f4b38c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000086", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000086", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [503, 543, 520, 570]\n- T2: [547, 466, 559, 477]\n- T3: [512, 493, 526, 514]\n\nImage 2 (GROUND) labeled targets:\n- T1: [707, 264, 734, 283]\n- T2: [766, 266, 784, 276]\n- T3: [726, 263, 738, 279]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T3 -> T2\nC: T1 -> T2 -> T3\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b24d57be085adf27a52399b72fc0d8dd66ea00a8c589514551a468c112288fdf.jpg", "reference_inputs/images/5a3abbef29682b89ebc84503a945744821bfb0ddf13046aafbac5b72e33d0641.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000087", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000087", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [547, 466, 559, 477]\n- T2: [503, 543, 520, 570]\n- T3: [512, 493, 526, 514]\n\nImage 2 (GROUND) labeled targets:\n- T1: [766, 266, 784, 276]\n- T2: [707, 264, 734, 283]\n- T3: [726, 263, 738, 279]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T3 -> T1 -> T2\nC: T2 -> T3 -> T1\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b24d57be085adf27a52399b72fc0d8dd66ea00a8c589514551a468c112288fdf.jpg", "reference_inputs/images/5a3abbef29682b89ebc84503a945744821bfb0ddf13046aafbac5b72e33d0641.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000090", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000090", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [557, 446, 567, 460]\n- T2: [562, 550, 579, 575]\n- T3: [598, 485, 610, 501]\n\nImage 2 (GROUND) labeled targets:\n- T1: [411, 274, 441, 294]\n- T2: [0, 210, 334, 399]\n- T3: [500, 264, 544, 301]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T2 -> T1 -> T3\nC: T1 -> T2 -> T3\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/982b4a63b2edbf1c4acd3dca191ce0fee7ef5a177a699d51862a19ce21f26430.jpg", "reference_inputs/images/876e52177c54d2f1a987ee7d6b60aef790c909d21871bccc19c29ad50f3249e4.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000091", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000091", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [546, 489, 562, 505]\n- T2: [562, 550, 579, 575]\n- T3: [557, 446, 567, 460]\n- T4: [598, 485, 610, 501]\n\nImage 2 (GROUND) labeled targets:\n- T1: [331, 277, 393, 302]\n- T2: [0, 210, 334, 399]\n- T3: [411, 274, 441, 294]\n- T4: [500, 264, 544, 301]\n\nAnswer options:\nA: T2 -> T1 -> T4 -> T3\nB: T2 -> T1 -> T3 -> T4\nC: T2 -> T4 -> T1 -> T3\nD: T1 -> T2 -> T4 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/982b4a63b2edbf1c4acd3dca191ce0fee7ef5a177a699d51862a19ce21f26430.jpg", "reference_inputs/images/876e52177c54d2f1a987ee7d6b60aef790c909d21871bccc19c29ad50f3249e4.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000092", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000092", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [481, 134, 529, 272]\n- T2: [612, 11, 644, 80]\n- T3: [542, 549, 578, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [593, 245, 647, 298]\n- T2: [696, 279, 715, 294]\n- T3: [706, 268, 868, 329]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T1 -> T3 -> T2\nC: T2 -> T3 -> T1\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f88fa0fa0cfdf57c7f9670e64a5867d5cc091ec62616cdfb2f509499a9a93c4b.jpg", "reference_inputs/images/11b35dac744ab3219d4eda136f940bcc36e68c83e8dcfb9b037a9b14bc224689.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000093", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000093", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [511, 422, 520, 434]\n- T2: [523, 491, 538, 508]\n- T3: [506, 444, 516, 455]\n\nImage 2 (GROUND) labeled targets:\n- T1: [657, 262, 673, 279]\n- T2: [787, 244, 922, 277]\n- T3: [666, 262, 686, 281]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T2 -> T1\nC: T2 -> T1 -> T3\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ae97a5118e22bef5ce9a41cfa51d0585b4d7e3cc4ffcabb3cb8d764cc31c7e27.jpg", "reference_inputs/images/14c63e7d5653e75ed98311d7c01089858201ac6abe0a8abf88b1d724630e404a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000094", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000094", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [674, 291, 690, 299]\n- T2: [714, 397, 734, 406]\n- T3: [447, 324, 478, 391]\n- T4: [664, 130, 683, 142]\n\nImage 2 (GROUND) labeled targets:\n- T1: [762, 256, 795, 298]\n- T2: [986, 234, 1014, 305]\n- T3: [376, 269, 458, 311]\n- T4: [679, 262, 688, 294]\n\nAnswer options:\nA: T3 -> T2 -> T1 -> T4\nB: T3 -> T2 -> T4 -> T1\nC: T3 -> T1 -> T2 -> T4\nD: T2 -> T3 -> T1 -> T4\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/acba93faab2b2b94bd8f121833dcdc0686ddae44c5819cbb8c7ea4fc08b05082.jpg", "reference_inputs/images/51c4243b98d1a3e1b646c00d0570aefc32d478d1128648c52c2f84bc12b8f8ff.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000095", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000095", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [374, 61, 409, 134]\n- T2: [260, 281, 278, 290]\n- T3: [286, 346, 304, 354]\n\nImage 2 (GROUND) labeled targets:\n- T1: [860, 250, 1023, 318]\n- T2: [446, 245, 462, 306]\n- T3: [310, 240, 333, 312]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T3 -> T1 -> T2\nC: T2 -> T3 -> T1\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e15b5ab4486d18e49579e63edec4137ea576208c7b3ed60f09684c21f8bbd4b2.jpg", "reference_inputs/images/83de54604647f8550f7370570918454c1f86c25ce7ef1fa73fb95060dbcecc4f.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000098", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000098", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [467, 511, 481, 530]\n- T2: [506, 443, 515, 454]\n- T3: [460, 482, 474, 501]\n\nImage 2 (GROUND) labeled targets:\n- T1: [217, 264, 356, 322]\n- T2: [496, 273, 522, 294]\n- T3: [326, 262, 382, 305]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T3 -> T1 -> T2\nC: T1 -> T2 -> T3\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1e06eee690ad51fbd2adeb66f140ab4391418b9db303c7f95d5e7b6904c232c3.jpg", "reference_inputs/images/e071a3d5ffba4f7bb6addc5eae0d5ad70da90edbd8c3be2635e149f0f653c9c0.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000099", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000099", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [467, 511, 481, 530]\n- T2: [460, 482, 474, 501]\n- T3: [506, 443, 515, 454]\n\nImage 2 (GROUND) labeled targets:\n- T1: [217, 264, 356, 322]\n- T2: [326, 262, 382, 305]\n- T3: [496, 273, 522, 294]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T3 -> T2\nC: T1 -> T2 -> T3\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1e06eee690ad51fbd2adeb66f140ab4391418b9db303c7f95d5e7b6904c232c3.jpg", "reference_inputs/images/e071a3d5ffba4f7bb6addc5eae0d5ad70da90edbd8c3be2635e149f0f653c9c0.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000100", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000100", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [550, 386, 558, 402]\n- T2: [594, 529, 609, 547]\n- T3: [600, 464, 611, 475]\n\nImage 2 (GROUND) labeled targets:\n- T1: [504, 273, 518, 288]\n- T2: [530, 276, 562, 298]\n- T3: [562, 278, 579, 292]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T3 -> T2 -> T1\nC: T2 -> T3 -> T1\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f836ca059921239eb42024ba8423dd2007a5475b13086e3b198e74a73afe9fb5.jpg", "reference_inputs/images/6735f98cc46d273f973863ea4bee7b2468bdd114361efff6508e8164bd3036ee.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000105", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000105", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [545, 130, 597, 198]\n- T2: [699, 550, 738, 575]\n- T3: [746, 198, 770, 209]\n\nImage 2 (GROUND) labeled targets:\n- T1: [301, 278, 334, 296]\n- T2: [391, 280, 406, 290]\n- T3: [220, 270, 226, 291]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T3 -> T2\nC: T1 -> T2 -> T3\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e4d2d3e636b4d3b3ff2b02938c80dc6fca66ffd777ec8927b57b34db548cdb62.jpg", "reference_inputs/images/474cb395aa9651abbdae88daf92c2b5ade56e19b9392b0a17f662fedcfd84053.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000106", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000106", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [537, 539, 557, 559]\n- T2: [482, 505, 498, 528]\n- T3: [466, 450, 477, 462]\n\nImage 2 (GROUND) labeled targets:\n- T1: [291, 284, 310, 292]\n- T2: [318, 278, 349, 294]\n- T3: [242, 276, 274, 297]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T3 -> T1 -> T2\nC: T3 -> T2 -> T1\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9a4b8a0ef75644fc6326ef5f14e66451e3488ddd11d148c7e81dea717a438b9b.jpg", "reference_inputs/images/1be0d45fc5eb88503ea3915f0b8fe008a69f91db0a9f5307220b605630c8732b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000107", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000107", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [537, 539, 557, 559]\n- T2: [482, 505, 498, 528]\n- T3: [507, 434, 522, 457]\n- T4: [466, 450, 477, 462]\n\nImage 2 (GROUND) labeled targets:\n- T1: [291, 284, 310, 292]\n- T2: [318, 278, 349, 294]\n- T3: [27, 239, 138, 299]\n- T4: [242, 276, 274, 297]\n\nAnswer options:\nA: T1 -> T2 -> T4 -> T3\nB: T2 -> T1 -> T4 -> T3\nC: unable to determine\nD: T1 -> T4 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9a4b8a0ef75644fc6326ef5f14e66451e3488ddd11d148c7e81dea717a438b9b.jpg", "reference_inputs/images/1be0d45fc5eb88503ea3915f0b8fe008a69f91db0a9f5307220b605630c8732b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000108", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000108", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [510, 550, 536, 575]\n- T2: [711, 509, 728, 526]\n- T3: [680, 322, 694, 330]\n\nImage 2 (GROUND) labeled targets:\n- T1: [507, 274, 538, 297]\n- T2: [680, 269, 686, 293]\n- T3: [620, 271, 625, 290]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T1 -> T2 -> T3\nC: T2 -> T3 -> T1\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5a3888322b97321e7e3fb2d8ff6f31991fad7fb77e5edf490d05b520cea9dfc4.jpg", "reference_inputs/images/95c96102e8a9126a8dc2362e391865f4cccb2be1df7cbe5c3bfa84cbdc0dab96.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000109", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000109", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [469, 450, 481, 460]\n- T2: [460, 490, 473, 506]\n- T3: [531, 470, 543, 484]\n\nImage 2 (GROUND) labeled targets:\n- T1: [445, 282, 465, 291]\n- T2: [417, 279, 442, 294]\n- T3: [543, 275, 566, 293]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T3 -> T2 -> T1\nC: T2 -> T1 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d4add35c59b208cf3073b24e073aa526911f29639774756b227a864960cb28c1.jpg", "reference_inputs/images/9c4e551e68a3a6d6aa8c608c169cf03df9762fc2789534aeba9a6a5bf69ba888.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000110", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000110", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [460, 490, 473, 506]\n- T2: [469, 450, 481, 460]\n- T3: [531, 470, 543, 484]\n- T4: [493, 427, 502, 439]\n\nImage 2 (GROUND) labeled targets:\n- T1: [417, 279, 442, 294]\n- T2: [445, 282, 465, 291]\n- T3: [543, 275, 566, 293]\n- T4: [484, 278, 498, 290]\n\nAnswer options:\nA: T1 -> T3 -> T2 -> T4\nB: T1 -> T3 -> T4 -> T2\nC: T3 -> T1 -> T2 -> T4\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d4add35c59b208cf3073b24e073aa526911f29639774756b227a864960cb28c1.jpg", "reference_inputs/images/9c4e551e68a3a6d6aa8c608c169cf03df9762fc2789534aeba9a6a5bf69ba888.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000111", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000111", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [606, 175, 616, 185]\n- T2: [390, 429, 401, 443]\n- T3: [798, 548, 821, 566]\n\nImage 2 (GROUND) labeled targets:\n- T1: [765, 266, 776, 302]\n- T2: [464, 275, 475, 298]\n- T3: [150, 243, 180, 316]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T2 -> T1\nC: T1 -> T2 -> T3\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/917c46fc5eb1879e3fc37c74683cf256f05979d46d41df406d52e1334f750549.jpg", "reference_inputs/images/0aa9e9877595e7ea42312c409755117570b7d8ce02794676566ebbbedb330209.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000115", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000115", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [386, 540, 446, 575]\n- T2: [58, 530, 86, 546]\n- T3: [330, 448, 343, 463]\n- T4: [302, 360, 324, 367]\n\nImage 2 (GROUND) labeled targets:\n- T1: [178, 258, 485, 379]\n- T2: [509, 265, 514, 298]\n- T3: [741, 234, 754, 273]\n- T4: [938, 235, 988, 310]\n\nAnswer options:\nA: T1 -> T4 -> T3 -> T2\nB: T1 -> T3 -> T2 -> T4\nC: unable to determine\nD: T1 -> T3 -> T4 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5503db5f5ae4225d2c6f3add36d2f86fa8ee0bd3d0e0cce90ad74ba26b5a2990.jpg", "reference_inputs/images/6de9c3585cf47e11898d5f77222fa22a23cf12d738d4e6bd2252109ea1039268.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000116", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000116", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [58, 530, 86, 546]\n- T2: [302, 360, 324, 367]\n- T3: [254, 479, 272, 495]\n\nImage 2 (GROUND) labeled targets:\n- T1: [509, 265, 514, 298]\n- T2: [938, 235, 988, 310]\n- T3: [610, 241, 632, 307]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T2 -> T1 -> T3\nC: T1 -> T2 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5503db5f5ae4225d2c6f3add36d2f86fa8ee0bd3d0e0cce90ad74ba26b5a2990.jpg", "reference_inputs/images/6de9c3585cf47e11898d5f77222fa22a23cf12d738d4e6bd2252109ea1039268.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000117", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000117", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [458, 493, 474, 510]\n- T2: [471, 439, 481, 450]\n- T3: [530, 458, 541, 470]\n- T4: [510, 498, 522, 514]\n\nImage 2 (GROUND) labeled targets:\n- T1: [115, 269, 306, 323]\n- T2: [400, 275, 432, 295]\n- T3: [580, 265, 616, 300]\n- T4: [491, 260, 600, 330]\n\nAnswer options:\nA: T4 -> T1 -> T3 -> T2\nB: T4 -> T1 -> T2 -> T3\nC: T4 -> T3 -> T1 -> T2\nD: T1 -> T4 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d244fd3db364087e2ec3d7b95f0825a29b0801934b47170eb9f4cba1fe921459.jpg", "reference_inputs/images/b719686ac3ce75d89888dbdd46c328c0141d8802f7ef17467032f5f9c3aea3c0.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000118", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000118", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [471, 439, 481, 450]\n- T2: [530, 458, 541, 470]\n- T3: [510, 498, 522, 514]\n- T4: [458, 493, 474, 510]\n\nImage 2 (GROUND) labeled targets:\n- T1: [400, 275, 432, 295]\n- T2: [580, 265, 616, 300]\n- T3: [491, 260, 600, 330]\n- T4: [115, 269, 306, 323]\n\nAnswer options:\nA: T4 -> T3 -> T2 -> T1\nB: T3 -> T4 -> T2 -> T1\nC: T3 -> T4 -> T1 -> T2\nD: T3 -> T2 -> T4 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d244fd3db364087e2ec3d7b95f0825a29b0801934b47170eb9f4cba1fe921459.jpg", "reference_inputs/images/b719686ac3ce75d89888dbdd46c328c0141d8802f7ef17467032f5f9c3aea3c0.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000119", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000119", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [602, 552, 606, 570]\n- T2: [506, 560, 526, 575]\n- T3: [471, 442, 482, 450]\n\nImage 2 (GROUND) labeled targets:\n- T1: [913, 232, 945, 307]\n- T2: [484, 269, 593, 323]\n- T3: [438, 282, 460, 292]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T2 -> T1 -> T3\nC: unable to determine\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e9067dfc8620208fff7788da69543ac530ef3e75011512f0184c3a8e854a456f.jpg", "reference_inputs/images/ce6b08db9a41daef36af00206c758b526312c6cf41bef6347f4d72c299e9eb2b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000121", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000121", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [474, 428, 484, 440]\n- T2: [501, 429, 516, 451]\n- T3: [530, 493, 543, 509]\n\nImage 2 (GROUND) labeled targets:\n- T1: [443, 277, 462, 291]\n- T2: [491, 258, 518, 292]\n- T3: [563, 272, 606, 300]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T1 -> T2\nC: T3 -> T2 -> T1\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cfa18aef7c5f1fd61ebebff7dafc64d1918ac5e9d9bb3839c48b3dd33dba961b.jpg", "reference_inputs/images/34d027dbab11397f568f524e93a1358bb23b7502da274b0c04a1dba07548b837.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000122", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000122", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [530, 493, 543, 509]\n- T2: [474, 428, 484, 440]\n- T3: [501, 429, 516, 451]\n\nImage 2 (GROUND) labeled targets:\n- T1: [563, 272, 606, 300]\n- T2: [443, 277, 462, 291]\n- T3: [491, 258, 518, 292]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2 -> T3\nC: T3 -> T1 -> T2\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cfa18aef7c5f1fd61ebebff7dafc64d1918ac5e9d9bb3839c48b3dd33dba961b.jpg", "reference_inputs/images/34d027dbab11397f568f524e93a1358bb23b7502da274b0c04a1dba07548b837.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000123", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000123", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [674, 291, 690, 299]\n- T2: [715, 82, 735, 94]\n- T3: [714, 397, 734, 406]\n\nImage 2 (GROUND) labeled targets:\n- T1: [797, 231, 848, 318]\n- T2: [264, 234, 283, 307]\n- T3: [1003, 238, 1023, 312]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T1 -> T2\nC: T1 -> T2 -> T3\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/acba93faab2b2b94bd8f121833dcdc0686ddae44c5819cbb8c7ea4fc08b05082.jpg", "reference_inputs/images/0434ee60b659113cb97e064670e8b02e9955eeff7f24ccedb8f970345927aa67.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000124", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000124", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [849, 467, 922, 536]\n- T2: [994, 395, 1023, 413]\n- T3: [347, 224, 476, 344]\n- T4: [898, 347, 924, 358]\n\nImage 2 (GROUND) labeled targets:\n- T1: [726, 282, 777, 297]\n- T2: [888, 270, 902, 293]\n- T3: [558, 268, 577, 293]\n- T4: [804, 271, 810, 292]\n\nAnswer options:\nA: unable to determine\nB: T4 -> T3 -> T1 -> T2\nC: T3 -> T4 -> T1 -> T2\nD: T3 -> T4 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9a637268ccc66c4e35c4ed2e56daf521b49a57fefec9c91252cab14f54e236a3.jpg", "reference_inputs/images/93f8c8c3cf249ce2b1f1c01c90e4fd3362d4265bc084ad006f6dc4e0f549605f.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000125", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000125", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [374, 273, 409, 344]\n- T2: [698, 422, 711, 438]\n- T3: [236, 528, 256, 548]\n\nImage 2 (GROUND) labeled targets:\n- T1: [527, 275, 553, 293]\n- T2: [773, 267, 780, 291]\n- T3: [388, 267, 394, 292]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T2 -> T3 -> T1\nC: T3 -> T2 -> T1\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4ab1ce6d791179a724a0abda241d57be67280c3d399e30581a826bced7e1b7b5.jpg", "reference_inputs/images/b3574f508469643224817c0968375f2dbc9b8fb5695e469f277f5de5349a47a5.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000126", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000126", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [236, 528, 256, 548]\n- T2: [698, 422, 711, 438]\n- T3: [374, 273, 409, 344]\n\nImage 2 (GROUND) labeled targets:\n- T1: [388, 267, 394, 292]\n- T2: [773, 267, 780, 291]\n- T3: [527, 275, 553, 293]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T3 -> T1 -> T2\nC: unable to determine\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4ab1ce6d791179a724a0abda241d57be67280c3d399e30581a826bced7e1b7b5.jpg", "reference_inputs/images/b3574f508469643224817c0968375f2dbc9b8fb5695e469f277f5de5349a47a5.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000127", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000127", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [522, 402, 534, 420]\n- T2: [536, 526, 553, 547]\n- T3: [508, 481, 522, 493]\n\nImage 2 (GROUND) labeled targets:\n- T1: [534, 263, 556, 290]\n- T2: [734, 150, 1023, 390]\n- T3: [497, 276, 554, 305]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T1 -> T2 -> T3\nC: T2 -> T3 -> T1\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e243169b9506053af12d4cec254d3ac2558083486bf9158c6ce0ca07d9a10508.jpg", "reference_inputs/images/8d896736401234e10331aeea3ac1fabfe73c9ad39a25d19bf9916657bd895404.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000128", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000128", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [522, 402, 534, 420]\n- T2: [536, 526, 553, 547]\n- T3: [458, 490, 470, 508]\n- T4: [508, 481, 522, 493]\n\nImage 2 (GROUND) labeled targets:\n- T1: [534, 263, 556, 290]\n- T2: [734, 150, 1023, 390]\n- T3: [234, 263, 342, 311]\n- T4: [497, 276, 554, 305]\n\nAnswer options:\nA: T2 -> T3 -> T1 -> T4\nB: T2 -> T4 -> T3 -> T1\nC: T2 -> T3 -> T4 -> T1\nD: T3 -> T2 -> T4 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e243169b9506053af12d4cec254d3ac2558083486bf9158c6ce0ca07d9a10508.jpg", "reference_inputs/images/8d896736401234e10331aeea3ac1fabfe73c9ad39a25d19bf9916657bd895404.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000129", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000129", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [374, 61, 409, 134]\n- T2: [260, 281, 278, 290]\n- T3: [286, 346, 304, 354]\n\nImage 2 (GROUND) labeled targets:\n- T1: [840, 259, 947, 309]\n- T2: [513, 250, 526, 303]\n- T3: [426, 246, 445, 307]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T2 -> T3 -> T1\nC: T1 -> T3 -> T2\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e15b5ab4486d18e49579e63edec4137ea576208c7b3ed60f09684c21f8bbd4b2.jpg", "reference_inputs/images/53f6413a5aece953d3d3f19040ae4fcc8e6ef179c92e51b1341438a643b7e920.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000130", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000130", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [673, 27, 691, 43]\n- T2: [438, 83, 472, 157]\n- T3: [383, 470, 421, 544]\n\nImage 2 (GROUND) labeled targets:\n- T1: [630, 270, 644, 292]\n- T2: [448, 272, 478, 295]\n- T3: [156, 271, 316, 318]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T3 -> T1 -> T2\nC: unable to determine\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/94afe022a1b3abd64effa4e05540b2eb9a3a7625b0e9302f064304f48566d76b.jpg", "reference_inputs/images/237069dce3d0317d10e096791b5d633b982014b0bd077bcd2d5a32473f3250ae.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000131", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000131", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [530, 466, 542, 477]\n- T2: [507, 482, 522, 495]\n- T3: [535, 524, 550, 545]\n\nImage 2 (GROUND) labeled targets:\n- T1: [566, 278, 618, 301]\n- T2: [492, 275, 557, 307]\n- T3: [770, 197, 1023, 386]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T2 -> T3 -> T1\nC: T3 -> T2 -> T1\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f6f1c52deeda02696e616062d3700b17ffeac1e2d963d9ec5a0d768c6fccd7b7.jpg", "reference_inputs/images/4361054b280cd9d73394edc6e371afc818b989aed9f9327cd8d6243c79133116.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000132", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000132", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [530, 466, 542, 477]\n- T2: [535, 524, 550, 545]\n- T3: [507, 482, 522, 495]\n\nImage 2 (GROUND) labeled targets:\n- T1: [566, 278, 618, 301]\n- T2: [770, 197, 1023, 386]\n- T3: [492, 275, 557, 307]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T2 -> T1 -> T3\nC: T3 -> T1 -> T2\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f6f1c52deeda02696e616062d3700b17ffeac1e2d963d9ec5a0d768c6fccd7b7.jpg", "reference_inputs/images/4361054b280cd9d73394edc6e371afc818b989aed9f9327cd8d6243c79133116.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000133", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000133", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [848, 387, 872, 403]\n- T2: [692, 12, 706, 32]\n- T3: [879, 488, 953, 524]\n- T4: [670, 306, 682, 314]\n\nImage 2 (GROUND) labeled targets:\n- T1: [790, 257, 806, 303]\n- T2: [2, 222, 35, 311]\n- T3: [850, 280, 894, 303]\n- T4: [826, 227, 858, 320]\n\nAnswer options:\nA: T2 -> T4 -> T1 -> T3\nB: unable to determine\nC: T4 -> T2 -> T1 -> T3\nD: T4 -> T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/54fe527ada2d1060e9abe51e500c229043388ec7321da0fcfb22ea61353b10b0.jpg", "reference_inputs/images/794cfbcbe7aebf0fd07de502de177b39301e6da824fb87e6bb1e4aa716ca8342.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000134", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000134", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [692, 12, 706, 32]\n- T2: [670, 306, 682, 314]\n- T3: [651, 81, 663, 97]\n- T4: [848, 387, 872, 403]\n\nImage 2 (GROUND) labeled targets:\n- T1: [2, 222, 35, 311]\n- T2: [826, 227, 858, 320]\n- T3: [86, 208, 129, 320]\n- T4: [790, 257, 806, 303]\n\nAnswer options:\nA: T2 -> T3 -> T4 -> T1\nB: T2 -> T1 -> T3 -> T4\nC: T3 -> T2 -> T1 -> T4\nD: T2 -> T3 -> T1 -> T4\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/54fe527ada2d1060e9abe51e500c229043388ec7321da0fcfb22ea61353b10b0.jpg", "reference_inputs/images/794cfbcbe7aebf0fd07de502de177b39301e6da824fb87e6bb1e4aa716ca8342.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000135", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000135", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [719, 379, 735, 388]\n- T2: [279, 492, 297, 508]\n- T3: [186, 140, 212, 154]\n- T4: [490, 517, 525, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [771, 258, 781, 295]\n- T2: [120, 246, 137, 301]\n- T3: [265, 267, 271, 292]\n- T4: [463, 248, 543, 314]\n\nAnswer options:\nA: T4 -> T2 -> T3 -> T1\nB: T2 -> T4 -> T1 -> T3\nC: T4 -> T2 -> T1 -> T3\nD: T4 -> T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1a0ea3e0c43e1f8333c7b5ac26d86b0d4127ca6debff07ec7d82fe47ce562749.jpg", "reference_inputs/images/c6777d37805bc1e28ba259a064593108c7c27c287b70aceb3538e5e222e83c53.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000136", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000136", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [286, 470, 315, 482]\n- T2: [186, 140, 212, 154]\n- T3: [673, 292, 686, 298]\n\nImage 2 (GROUND) labeled targets:\n- T1: [162, 255, 193, 300]\n- T2: [265, 267, 271, 292]\n- T3: [678, 265, 686, 294]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T3 -> T1 -> T2\nC: T1 -> T3 -> T2\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1a0ea3e0c43e1f8333c7b5ac26d86b0d4127ca6debff07ec7d82fe47ce562749.jpg", "reference_inputs/images/c6777d37805bc1e28ba259a064593108c7c27c287b70aceb3538e5e222e83c53.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000137", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000137", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [678, 370, 689, 384]\n- T2: [502, 350, 539, 425]\n- T3: [674, 36, 690, 51]\n\nImage 2 (GROUND) labeled targets:\n- T1: [692, 262, 702, 295]\n- T2: [498, 278, 548, 300]\n- T3: [612, 273, 626, 290]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T2 -> T1\nC: T1 -> T2 -> T3\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ef325bedc98b528fdeca5d4476bb350fd621ccf27ffa94819def6fde3cedd910.jpg", "reference_inputs/images/42ee3d96da42feea79428db2a7aaa9f59bcc8eab925e72270f906ef0ce0d1823.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000138", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000138", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [756, 551, 760, 569]\n- T2: [471, 440, 482, 454]\n- T3: [527, 496, 540, 513]\n\nImage 2 (GROUND) labeled targets:\n- T1: [955, 262, 966, 294]\n- T2: [457, 278, 471, 290]\n- T3: [535, 278, 557, 293]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T3 -> T2\nC: T2 -> T1 -> T3\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/efb6030f16d7f5b18cfdb64020260858b0e2e3ac87e913fc5a0ab17c0cfbd4fa.jpg", "reference_inputs/images/cc432e0bbc20b1fa31132acb84df4fcdde4420d63cdec092f34992c12a82e331.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000139", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000139", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [527, 496, 540, 513]\n- T2: [471, 440, 482, 454]\n- T3: [756, 551, 760, 569]\n\nImage 2 (GROUND) labeled targets:\n- T1: [535, 278, 557, 293]\n- T2: [457, 278, 471, 290]\n- T3: [955, 262, 966, 294]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T3 -> T2\nC: T1 -> T2 -> T3\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/efb6030f16d7f5b18cfdb64020260858b0e2e3ac87e913fc5a0ab17c0cfbd4fa.jpg", "reference_inputs/images/cc432e0bbc20b1fa31132acb84df4fcdde4420d63cdec092f34992c12a82e331.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000142", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000142", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [215, 206, 235, 217]\n- T2: [386, 63, 424, 139]\n- T3: [258, 208, 274, 218]\n\nImage 2 (GROUND) labeled targets:\n- T1: [338, 248, 358, 306]\n- T2: [429, 257, 843, 360]\n- T3: [305, 246, 326, 309]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T3 -> T2 -> T1\nC: T1 -> T2 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a65c54e78979607a6fb2cb0b3e1f34c6dc46eb994fa50f78920d085334656503.jpg", "reference_inputs/images/c29d2ae0b8e4d26fa8e1494af668e571e75459fc2c61d548d1fdfe35fbebe69b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000143", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000143", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [214, 214, 237, 224]\n- T2: [241, 551, 258, 574]\n- T3: [247, 401, 270, 414]\n\nImage 2 (GROUND) labeled targets:\n- T1: [699, 252, 718, 303]\n- T2: [45, 251, 64, 309]\n- T3: [346, 242, 365, 308]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T2 -> T1\nC: T1 -> T3 -> T2\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d81274f557b0ca07c4521b0e223aee5057051066586632f9e207009e5fe30d47.jpg", "reference_inputs/images/008072f8938f8f1a63f207a5da458f8c1016122eb4926ea5e9575db7ace1132e.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_30m/000144", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000144", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [447, 534, 467, 558]\n- T2: [466, 462, 476, 474]\n- T3: [476, 414, 494, 421]\n\nImage 2 (GROUND) labeled targets:\n- T1: [390, 279, 431, 297]\n- T2: [442, 279, 460, 291]\n- T3: [465, 283, 489, 289]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T1 -> T3\nC: T1 -> T3 -> T2\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/22deebf051013c11ba5cc52fd59354946d20531f925104078bc07df3256d8b9e.jpg", "reference_inputs/images/176edff7c377ecfb0a85bbb5fe6f909debb69e63dd867a4d9757b6b09591aec1.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000001", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [434, 492, 453, 532]\n- T2: [434, 12, 455, 58]\n- T3: [470, 190, 489, 234]\n- T4: [536, 289, 562, 367]\n\nImage 2 (GROUND) labeled targets:\n- T1: [311, 274, 377, 304]\n- T2: [454, 282, 470, 290]\n- T3: [470, 278, 489, 292]\n- T4: [537, 250, 579, 294]\n\nAnswer options:\nA: T1 -> T3 -> T4 -> T2\nB: unable to determine\nC: T1 -> T4 -> T3 -> T2\nD: T4 -> T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1336f727d14e0a43920ba0e565b7a3a66a5ef82e0946e2742228982971657d7b.jpg", "reference_inputs/images/326e309493b74656ab397a5c7fe4a7d782f7c238e5d6a88d3395323169ed9146.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000002", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [511, 294, 532, 337]\n- T2: [508, 30, 526, 72]\n- T3: [570, 162, 589, 203]\n- T4: [574, 458, 596, 503]\n\nImage 2 (GROUND) labeled targets:\n- T1: [341, 290, 366, 305]\n- T2: [442, 290, 489, 310]\n- T3: [310, 292, 335, 308]\n- T4: [261, 295, 295, 304]\n\nAnswer options:\nA: T2 -> T1 -> T3 -> T4\nB: T2 -> T3 -> T4 -> T1\nC: T3 -> T2 -> T1 -> T4\nD: T2 -> T3 -> T1 -> T4\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6335503211ddf6b2368041e62c88cd8c440ca0c151d187f98a89d26e416d6c09.jpg", "reference_inputs/images/606dd5fea012af2756b373ee21b6a9f4a7ff6eca37a4eb356fb07f1cb3476f3e.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000003", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [509, 334, 527, 374]\n- T2: [437, 381, 464, 458]\n- T3: [435, 1, 458, 46]\n\nImage 2 (GROUND) labeled targets:\n- T1: [498, 264, 580, 320]\n- T2: [0, 18, 257, 336]\n- T3: [425, 281, 442, 293]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: unable to determine\nC: T1 -> T2 -> T3\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e4ff7a204c5127ccacc2cdbba88bc5213586c8fae62904c9c188f0a3e8929b50.jpg", "reference_inputs/images/34591fc1fa22610d2519ee5814ed778a1c518ffb2f2dfb3ab15e3acf8d937d50.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000004", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [509, 334, 527, 374]\n- T2: [437, 381, 464, 458]\n- T3: [435, 1, 458, 46]\n\nImage 2 (GROUND) labeled targets:\n- T1: [498, 264, 580, 320]\n- T2: [0, 18, 257, 336]\n- T3: [425, 281, 442, 293]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T1 -> T3 -> T2\nC: T3 -> T1 -> T2\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e4ff7a204c5127ccacc2cdbba88bc5213586c8fae62904c9c188f0a3e8929b50.jpg", "reference_inputs/images/34591fc1fa22610d2519ee5814ed778a1c518ffb2f2dfb3ab15e3acf8d937d50.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000007", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [574, 454, 597, 498]\n- T2: [571, 120, 590, 161]\n- T3: [496, 197, 522, 276]\n- T4: [530, 442, 551, 486]\n\nImage 2 (GROUND) labeled targets:\n- T1: [430, 276, 450, 286]\n- T2: [329, 260, 381, 286]\n- T3: [481, 247, 523, 302]\n- T4: [467, 281, 482, 290]\n\nAnswer options:\nA: T2 -> T3 -> T4 -> T1\nB: T2 -> T3 -> T1 -> T4\nC: T3 -> T2 -> T4 -> T1\nD: T2 -> T4 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f1b1fabf0f694b63c2ad925139ff55b74e5134f92f21fc72456f28796a8d42c4.jpg", "reference_inputs/images/30e6bd80fec7eeb677a4b663c71af085dbdb8ae9f65814a556f247652e14e01b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000009", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [508, 470, 526, 511]\n- T2: [435, 200, 457, 244]\n- T3: [473, 82, 493, 126]\n\nImage 2 (GROUND) labeled targets:\n- T1: [485, 248, 644, 355]\n- T2: [390, 280, 431, 297]\n- T3: [467, 275, 491, 293]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T1 -> T2 -> T3\nC: T2 -> T1 -> T3\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ededbee325399e98678b611d225826acf13d11765882ee4f176ffa431c6d154d.jpg", "reference_inputs/images/64aea7386a93b8d411c31c3c476f955f4bb9d442dea86752dfdbfd3a90eec182.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000010", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [512, 417, 532, 460]\n- T2: [504, 546, 519, 575]\n- T3: [572, 286, 588, 326]\n\nImage 2 (GROUND) labeled targets:\n- T1: [522, 275, 534, 293]\n- T2: [497, 274, 526, 297]\n- T3: [562, 281, 578, 291]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T2 -> T1 -> T3\nC: T2 -> T3 -> T1\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d6d5c7567398c19334c521f3f939a35c6222e35a97340cadca037bc7534d30dd.jpg", "reference_inputs/images/314d01bbfb4d254b346e7f6bc4fb62da813832c57ce8b979e13a5d0aae67e2bd.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000011", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [572, 286, 588, 326]\n- T2: [512, 417, 532, 460]\n- T3: [504, 546, 519, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [562, 281, 578, 291]\n- T2: [522, 275, 534, 293]\n- T3: [497, 274, 526, 297]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T1 -> T3 -> T2\nC: T2 -> T1 -> T3\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d6d5c7567398c19334c521f3f939a35c6222e35a97340cadca037bc7534d30dd.jpg", "reference_inputs/images/314d01bbfb4d254b346e7f6bc4fb62da813832c57ce8b979e13a5d0aae67e2bd.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000012", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [472, 559, 487, 575]\n- T2: [507, 365, 527, 407]\n- T3: [434, 202, 454, 245]\n- T4: [474, 24, 493, 66]\n\nImage 2 (GROUND) labeled targets:\n- T1: [0, 210, 334, 399]\n- T2: [500, 264, 544, 301]\n- T3: [411, 274, 441, 294]\n- T4: [478, 281, 494, 291]\n\nAnswer options:\nA: T1 -> T2 -> T4 -> T3\nB: T1 -> T2 -> T3 -> T4\nC: T2 -> T1 -> T3 -> T4\nD: T1 -> T3 -> T2 -> T4\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5279243a189dac5df5abe787abc68a73bb0ecaf5b22deee4090ef8daf12ab644.jpg", "reference_inputs/images/876e52177c54d2f1a987ee7d6b60aef790c909d21871bccc19c29ad50f3249e4.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000013", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [434, 202, 454, 245]\n- T2: [472, 559, 487, 575]\n- T3: [474, 24, 493, 66]\n\nImage 2 (GROUND) labeled targets:\n- T1: [411, 274, 441, 294]\n- T2: [0, 210, 334, 399]\n- T3: [478, 281, 494, 291]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T3 -> T1 -> T2\nC: T1 -> T3 -> T2\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5279243a189dac5df5abe787abc68a73bb0ecaf5b22deee4090ef8daf12ab644.jpg", "reference_inputs/images/876e52177c54d2f1a987ee7d6b60aef790c909d21871bccc19c29ad50f3249e4.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000016", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [506, 464, 527, 507]\n- T2: [437, 279, 459, 323]\n- T3: [434, 27, 453, 73]\n\nImage 2 (GROUND) labeled targets:\n- T1: [500, 258, 562, 306]\n- T2: [406, 280, 442, 295]\n- T3: [448, 280, 464, 290]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T3 -> T2\nC: T1 -> T2 -> T3\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8bc139be26f110733fffacdf00532d34cea4f956a85c5c19c599c4ebbb253eba.jpg", "reference_inputs/images/4b219c2e1a5ddc8d0638c9babe4ad1646d1f1623be3f4678702dfca40e824bbc.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000017", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [506, 464, 527, 507]\n- T2: [437, 279, 459, 323]\n- T3: [434, 27, 453, 73]\n\nImage 2 (GROUND) labeled targets:\n- T1: [500, 258, 562, 306]\n- T2: [406, 280, 442, 295]\n- T3: [448, 280, 464, 290]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T3 -> T1\nC: T2 -> T1 -> T3\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8bc139be26f110733fffacdf00532d34cea4f956a85c5c19c599c4ebbb253eba.jpg", "reference_inputs/images/4b219c2e1a5ddc8d0638c9babe4ad1646d1f1623be3f4678702dfca40e824bbc.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000019", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [434, 492, 453, 532]\n- T2: [470, 190, 489, 234]\n- T3: [434, 12, 455, 58]\n- T4: [536, 289, 562, 367]\n\nImage 2 (GROUND) labeled targets:\n- T1: [354, 275, 402, 300]\n- T2: [474, 279, 490, 291]\n- T3: [458, 282, 474, 290]\n- T4: [534, 254, 571, 294]\n\nAnswer options:\nA: T4 -> T1 -> T2 -> T3\nB: T1 -> T2 -> T4 -> T3\nC: T1 -> T4 -> T2 -> T3\nD: T1 -> T4 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1336f727d14e0a43920ba0e565b7a3a66a5ef82e0946e2742228982971657d7b.jpg", "reference_inputs/images/0eb73e9b59314fe88798044774dc7b2c640adb41215b0ea735d4ac8fa5c6b73d.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000020", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [435, 322, 457, 366]\n- T2: [542, 435, 562, 478]\n- T3: [473, 206, 493, 249]\n\nImage 2 (GROUND) labeled targets:\n- T1: [438, 282, 461, 292]\n- T2: [547, 274, 573, 294]\n- T3: [482, 278, 498, 290]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T3 -> T1\nC: T3 -> T2 -> T1\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8e830185c444506afc6719687ec1c17624b01db7c88652e33518d29d0f62951b.jpg", "reference_inputs/images/0babe2b74ff3bef3cdac98263fe4524186cb1fd3c2bd927891a0e7a8de627389.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000021", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [536, 412, 568, 498]\n- T2: [444, 562, 472, 575]\n- T3: [434, 255, 464, 331]\n\nImage 2 (GROUND) labeled targets:\n- T1: [529, 265, 552, 290]\n- T2: [432, 282, 474, 292]\n- T3: [466, 270, 483, 289]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T3 -> T2 -> T1\nC: T2 -> T3 -> T1\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7b151ccb6720a410136b56eaa274076a3a78d2dc385f54f1531fc6f92aaf3fae.jpg", "reference_inputs/images/fe8e110d6bb38794d1bebd55d8ac04077ab3b8a2e90fe6f56803526d46685019.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000022", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [536, 412, 568, 498]\n- T2: [434, 255, 464, 331]\n- T3: [444, 562, 472, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [529, 265, 552, 290]\n- T2: [466, 270, 483, 289]\n- T3: [432, 282, 474, 292]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T2 -> T3 -> T1\nC: T1 -> T2 -> T3\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7b151ccb6720a410136b56eaa274076a3a78d2dc385f54f1531fc6f92aaf3fae.jpg", "reference_inputs/images/fe8e110d6bb38794d1bebd55d8ac04077ab3b8a2e90fe6f56803526d46685019.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000026", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [474, 520, 492, 561]\n- T2: [434, 0, 465, 70]\n- T3: [534, 226, 565, 302]\n\nImage 2 (GROUND) labeled targets:\n- T1: [154, 250, 390, 352]\n- T2: [444, 262, 471, 291]\n- T3: [545, 244, 590, 296]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T1 -> T2 -> T3\nC: T3 -> T1 -> T2\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9825ac2cf22175ea4b1eccdc5ca6a7fbc946d54fd5eba1875ac2cbbcef46b1e9.jpg", "reference_inputs/images/f715261c31086ae963af6c95167737fcc81cedadfdf0f8460aaa5262d2a8969f.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000028", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [438, 512, 457, 552]\n- T2: [542, 435, 562, 478]\n- T3: [473, 206, 493, 249]\n\nImage 2 (GROUND) labeled targets:\n- T1: [428, 279, 450, 293]\n- T2: [541, 276, 561, 292]\n- T3: [486, 279, 498, 290]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T3 -> T1\nC: T3 -> T1 -> T2\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8e830185c444506afc6719687ec1c17624b01db7c88652e33518d29d0f62951b.jpg", "reference_inputs/images/07a0ad132afd99207a266320b96b576f8cf8b1518a143c3d17471bc368268383.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000029", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [883, 392, 928, 414]\n- T2: [913, 318, 962, 338]\n- T3: [698, 346, 742, 368]\n\nImage 2 (GROUND) labeled targets:\n- T1: [563, 282, 583, 290]\n- T2: [504, 282, 519, 290]\n- T3: [532, 281, 559, 294]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T2 -> T1\nC: T3 -> T1 -> T2\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bf4e96f4b7d94ff33c01530a446d0874327259742aea607e7efd739a863ebae4.jpg", "reference_inputs/images/912001ea44deb2e90793bbecd6ab0407711e1c2fc6b14c92a7fff3b001b742e5.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000030", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [913, 318, 962, 338]\n- T2: [883, 392, 928, 414]\n- T3: [698, 346, 742, 368]\n\nImage 2 (GROUND) labeled targets:\n- T1: [504, 282, 519, 290]\n- T2: [563, 282, 583, 290]\n- T3: [532, 281, 559, 294]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T2 -> T3 -> T1\nC: T3 -> T2 -> T1\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bf4e96f4b7d94ff33c01530a446d0874327259742aea607e7efd739a863ebae4.jpg", "reference_inputs/images/912001ea44deb2e90793bbecd6ab0407711e1c2fc6b14c92a7fff3b001b742e5.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000031", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000031", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [14, 179, 114, 211]\n- T2: [198, 267, 243, 290]\n- T3: [10, 95, 26, 103]\n\nImage 2 (GROUND) labeled targets:\n- T1: [654, 257, 714, 296]\n- T2: [597, 281, 632, 300]\n- T3: [758, 273, 767, 292]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T3 -> T1\nC: T2 -> T1 -> T3\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7d8c6dc3e4048ac215463ba16b9bf009b62fb5cb1853919ed961264defdf375f.jpg", "reference_inputs/images/d760e180b6d0fe7d5f5740e5ff281233dff08dc94c3c056a4d088a7a0203454d.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000032", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [14, 179, 114, 211]\n- T2: [198, 267, 243, 290]\n- T3: [10, 95, 26, 103]\n\nImage 2 (GROUND) labeled targets:\n- T1: [654, 257, 714, 296]\n- T2: [597, 281, 632, 300]\n- T3: [758, 273, 767, 292]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T2 -> T3 -> T1\nC: T1 -> T3 -> T2\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7d8c6dc3e4048ac215463ba16b9bf009b62fb5cb1853919ed961264defdf375f.jpg", "reference_inputs/images/d760e180b6d0fe7d5f5740e5ff281233dff08dc94c3c056a4d088a7a0203454d.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000034", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000034", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [688, 546, 729, 575]\n- T2: [486, 358, 522, 394]\n- T3: [313, 134, 354, 175]\n\nImage 2 (GROUND) labeled targets:\n- T1: [0, 270, 110, 323]\n- T2: [226, 278, 258, 295]\n- T3: [311, 280, 326, 290]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T1 -> T2 -> T3\nC: T2 -> T1 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1c5457950384d1d8f40f4509497d99dfa6f81e9a2dbba70aa46cd97481aad06a.jpg", "reference_inputs/images/77d355cbe295b4ce12ba03e43795081814bf235837271b4caadc6eb5ded3c1d5.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000035", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000035", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [313, 134, 354, 175]\n- T2: [486, 358, 522, 394]\n- T3: [688, 546, 729, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [311, 280, 326, 290]\n- T2: [226, 278, 258, 295]\n- T3: [0, 270, 110, 323]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T2 -> T1 -> T3\nC: T1 -> T2 -> T3\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1c5457950384d1d8f40f4509497d99dfa6f81e9a2dbba70aa46cd97481aad06a.jpg", "reference_inputs/images/77d355cbe295b4ce12ba03e43795081814bf235837271b4caadc6eb5ded3c1d5.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000037", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000037", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [71, 238, 88, 244]\n- T2: [68, 503, 78, 511]\n- T3: [435, 570, 454, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [863, 270, 871, 292]\n- T2: [548, 272, 559, 294]\n- T3: [0, 213, 283, 355]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T1 -> T2\nC: T2 -> T3 -> T1\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5e605b5e4183648946c7cea4c82bcc77b85956b0b889bb0b887d46beb53628aa.jpg", "reference_inputs/images/1f2a00a0e56882a06b6b2ad1adfb16344a112cd9c178bc762cf53cbf501e6f03.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000038", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000038", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [428, 183, 450, 227]\n- T2: [437, 538, 457, 575]\n- T3: [518, 245, 540, 289]\n\nImage 2 (GROUND) labeled targets:\n- T1: [439, 282, 460, 291]\n- T2: [341, 265, 402, 302]\n- T3: [518, 282, 539, 292]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T2 -> T1 -> T3\nC: T2 -> T3 -> T1\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c0754c3fd9ed92cc477ef99dfe6d1ab619e52fe47f3d8eb86cfd86884958c025.jpg", "reference_inputs/images/121ca0a666b27ee44d92209d4b2bcecda33df73f1f6332b6dff01271e6619a5e.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000039", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000039", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [437, 490, 464, 567]\n- T2: [506, 452, 528, 496]\n- T3: [434, 27, 457, 71]\n\nImage 2 (GROUND) labeled targets:\n- T1: [0, 30, 268, 333]\n- T2: [484, 269, 593, 323]\n- T3: [438, 282, 460, 292]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T3 -> T2\nC: T2 -> T1 -> T3\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e2d3d48257038542e95080942944bed57b0dd925a0658170946298c77bad1501.jpg", "reference_inputs/images/ce6b08db9a41daef36af00206c758b526312c6cf41bef6347f4d72c299e9eb2b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000041", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000041", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [570, 37, 590, 78]\n- T2: [530, 515, 551, 558]\n- T3: [502, 301, 522, 344]\n\nImage 2 (GROUND) labeled targets:\n- T1: [677, 279, 694, 289]\n- T2: [709, 274, 787, 305]\n- T3: [629, 275, 654, 293]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T2 -> T1 -> T3\nC: T1 -> T2 -> T3\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9e6c6109eafac788101cfd3b1b7088160bb23bf14895659d84c525d5f69ff798.jpg", "reference_inputs/images/2b6fff964b0a7b426d3c5321d16ca75da087b49c9e9fb3414269f22cf638b5e2.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000044", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000044", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [493, 260, 518, 337]\n- T2: [539, 523, 558, 569]\n- T3: [456, 344, 478, 388]\n\nImage 2 (GROUND) labeled targets:\n- T1: [482, 245, 524, 296]\n- T2: [629, 251, 784, 328]\n- T3: [409, 278, 454, 298]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T2 -> T3 -> T1\nC: T3 -> T2 -> T1\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/435fb3438a5939da084e7fb0ad694d3be840746dfe99be4dbd160aa60bc58719.jpg", "reference_inputs/images/bb5fb5ea9e5337cd8439330bc758a7b429296eb13e3a974b49d53f06abdc07c5.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000045", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000045", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [456, 344, 478, 388]\n- T2: [493, 260, 518, 337]\n- T3: [539, 523, 558, 569]\n\nImage 2 (GROUND) labeled targets:\n- T1: [463, 282, 483, 291]\n- T2: [496, 265, 517, 290]\n- T3: [545, 277, 570, 294]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T2 -> T3 -> T1\nC: T1 -> T3 -> T2\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/435fb3438a5939da084e7fb0ad694d3be840746dfe99be4dbd160aa60bc58719.jpg", "reference_inputs/images/bcdc1e69c0a36fdd6f81b6dc138bac8948640e1f401b454a6715a6941bd5a4a5.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000046", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000046", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [437, 279, 459, 323]\n- T2: [506, 464, 527, 507]\n- T3: [434, 27, 453, 73]\n\nImage 2 (GROUND) labeled targets:\n- T1: [390, 279, 431, 297]\n- T2: [495, 247, 573, 315]\n- T3: [442, 279, 460, 291]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T2 -> T1 -> T3\nC: T3 -> T2 -> T1\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8bc139be26f110733fffacdf00532d34cea4f956a85c5c19c599c4ebbb253eba.jpg", "reference_inputs/images/176edff7c377ecfb0a85bbb5fe6f909debb69e63dd867a4d9757b6b09591aec1.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000047", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000047", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [506, 464, 527, 507]\n- T2: [437, 279, 459, 323]\n- T3: [434, 27, 453, 73]\n\nImage 2 (GROUND) labeled targets:\n- T1: [495, 247, 573, 315]\n- T2: [390, 279, 431, 297]\n- T3: [442, 279, 460, 291]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T3 -> T2 -> T1\nC: T2 -> T1 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8bc139be26f110733fffacdf00532d34cea4f956a85c5c19c599c4ebbb253eba.jpg", "reference_inputs/images/176edff7c377ecfb0a85bbb5fe6f909debb69e63dd867a4d9757b6b09591aec1.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000048", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000048", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [435, 397, 457, 442]\n- T2: [468, 167, 487, 210]\n- T3: [505, 390, 524, 434]\n\nImage 2 (GROUND) labeled targets:\n- T1: [0, 262, 219, 346]\n- T2: [437, 270, 473, 298]\n- T3: [472, 245, 594, 338]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T1 -> T3 -> T2\nC: T3 -> T2 -> T1\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/88901f1b80fcc6a22add5b0413965bc44db5260cc9fff6217e99c9def73f08ba.jpg", "reference_inputs/images/360617ebd1ddf8668083527a4aeebbdda1341224f852b2297b1aa4e0edd6dbf3.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000049", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000049", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [810, 187, 856, 206]\n- T2: [766, 102, 859, 136]\n- T3: [958, 256, 974, 266]\n\nImage 2 (GROUND) labeled targets:\n- T1: [577, 278, 612, 301]\n- T2: [436, 242, 501, 302]\n- T3: [633, 276, 646, 294]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T2 -> T3 -> T1\nC: unable to determine\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/334da4287271cddf47d6c5e5c4598fa1148a358472a4dead85ffcd9246528b37.jpg", "reference_inputs/images/c5b600c09a6d37cd46165ccf13dda6e2f27b7d463018639951817900fbdef907.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000050", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000050", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [434, 0, 465, 70]\n- T2: [474, 520, 492, 561]\n- T3: [534, 226, 565, 302]\n\nImage 2 (GROUND) labeled targets:\n- T1: [447, 264, 474, 290]\n- T2: [334, 265, 435, 320]\n- T3: [536, 250, 579, 294]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T1 -> T2\nC: T2 -> T1 -> T3\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9825ac2cf22175ea4b1eccdc5ca6a7fbc946d54fd5eba1875ac2cbbcef46b1e9.jpg", "reference_inputs/images/1f4e5bfc1dc07dd2fc6ab971db43666d4bee8f303a3194ecb836bf23599c8aa9.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000051", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000051", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [545, 330, 564, 370]\n- T2: [812, 19, 827, 30]\n- T3: [434, 325, 454, 367]\n\nImage 2 (GROUND) labeled targets:\n- T1: [674, 258, 855, 332]\n- T2: [881, 269, 888, 291]\n- T3: [15, 238, 260, 330]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T3 -> T2\nC: T3 -> T2 -> T1\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/981967d761992e7129c43905aa67c08a3b00f48842f047a8760372f783c45b72.jpg", "reference_inputs/images/43b0a3fe42f2acc793e99dea4dd8389f524b8f3d47f09236481fb1cc1befeb13.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000052", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000052", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [1002, 360, 1023, 379]\n- T2: [957, 502, 973, 513]\n- T3: [530, 146, 551, 191]\n\nImage 2 (GROUND) labeled targets:\n- T1: [505, 267, 554, 306]\n- T2: [243, 262, 266, 296]\n- T3: [650, 282, 667, 290]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T2 -> T1 -> T3\nC: T1 -> T2 -> T3\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d1ee783bb91590ea5128726b99ed4acd64eb536a4dfa2e81f2a4271e019ffe16.jpg", "reference_inputs/images/2c4d8a2bd0304a00d3fcdc28715ad25aedb7c02628a93f4a466fc02ebe7ae0fe.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000053", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000053", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [766, 358, 858, 390]\n- T2: [957, 502, 973, 513]\n- T3: [530, 146, 551, 191]\n\nImage 2 (GROUND) labeled targets:\n- T1: [513, 270, 533, 278]\n- T2: [243, 262, 266, 296]\n- T3: [650, 282, 667, 290]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T1 -> T3 -> T2\nC: T3 -> T1 -> T2\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d1ee783bb91590ea5128726b99ed4acd64eb536a4dfa2e81f2a4271e019ffe16.jpg", "reference_inputs/images/2c4d8a2bd0304a00d3fcdc28715ad25aedb7c02628a93f4a466fc02ebe7ae0fe.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000055", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000055", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [536, 534, 562, 575]\n- T2: [434, 261, 456, 305]\n- T3: [470, 436, 489, 481]\n\nImage 2 (GROUND) labeled targets:\n- T1: [558, 203, 670, 306]\n- T2: [428, 282, 442, 293]\n- T3: [440, 274, 473, 298]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T2 -> T1 -> T3\nC: T1 -> T2 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/93df5d76f25eb5b5a9e6ee01d8da4dc93df5407d5b90f8ac9ed2264eaa9b2d22.jpg", "reference_inputs/images/4d1429f8a559ffa76ff6dd6181113d642a7d6ccac748309b027172a63950138b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000056", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000056", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [437, 296, 456, 338]\n- T2: [534, 536, 553, 575]\n- T3: [518, 2, 540, 46]\n\nImage 2 (GROUND) labeled targets:\n- T1: [374, 269, 419, 298]\n- T2: [680, 214, 1023, 390]\n- T3: [518, 282, 537, 291]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T3 -> T1\nC: T2 -> T1 -> T3\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/81f35adf479a310e3e4657e23a0d8b894b30bb03197b5adc41b6e6f750abcede.jpg", "reference_inputs/images/f3d0a4d31d16c204ff9a76a17bcb947c6cb4b7bbb5cf7f59f6e1d7ad3080d41a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000059", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000059", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [473, 331, 493, 373]\n- T2: [542, 558, 561, 575]\n- T3: [435, 446, 457, 490]\n\nImage 2 (GROUND) labeled targets:\n- T1: [472, 276, 493, 292]\n- T2: [570, 266, 609, 299]\n- T3: [407, 280, 442, 294]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T1 -> T2\nC: T1 -> T2 -> T3\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1741de326d767084dd4e5fd0fc6e3e2bad8b0f6b9c81fdd632056be66516870c.jpg", "reference_inputs/images/ab4684364fe6a38c2026ce521fb39e2527721bca2271eb4e42b0311ffe907e38.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000062", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000062", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [434, 75, 456, 119]\n- T2: [438, 514, 456, 555]\n- T3: [438, 266, 456, 306]\n\nImage 2 (GROUND) labeled targets:\n- T1: [456, 282, 471, 290]\n- T2: [359, 275, 406, 300]\n- T3: [437, 280, 456, 292]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T3 -> T2 -> T1\nC: T2 -> T3 -> T1\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2daef8c47c7bf8ef490e270b16a548536cad0910bb6daa16f77ccca8d04eaa8c.jpg", "reference_inputs/images/502fe2b7bf5d1fe5c44e96178c57cdc8f6cda7fad2cf013c6853ede35a442838.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000064", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000064", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [492, 382, 521, 466]\n- T2: [436, 338, 457, 381]\n- T3: [456, 466, 478, 510]\n\nImage 2 (GROUND) labeled targets:\n- T1: [493, 261, 519, 291]\n- T2: [449, 278, 466, 290]\n- T3: [453, 282, 478, 293]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T3 -> T2 -> T1\nC: T1 -> T3 -> T2\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/dd13bd300738eee426faa5a3b0287c593a83a0111f3fa8ede7064a49048fad82.jpg", "reference_inputs/images/ed2189f53b3f6d1c05bceb524c0005a7508a1ff0431f3713d530ae623c5b2642.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000067", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000067", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [766, 512, 808, 554]\n- T2: [608, 288, 644, 326]\n- T3: [406, 86, 449, 128]\n\nImage 2 (GROUND) labeled targets:\n- T1: [216, 280, 233, 290]\n- T2: [130, 279, 162, 295]\n- T3: [0, 274, 50, 314]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T3 -> T2\nC: T2 -> T3 -> T1\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/74b31aee53bec746cfe75b95269b54832945cc93e1e41e332bd12f1614f13f1e.jpg", "reference_inputs/images/6b5ca664a0bec76fcb4e83634ea93c39a1cd48cb0e2e19cac5000ed55fa74d96.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000068", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000068", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [765, 482, 858, 518]\n- T2: [808, 558, 854, 575]\n- T3: [530, 270, 551, 314]\n\nImage 2 (GROUND) labeled targets:\n- T1: [478, 184, 574, 311]\n- T2: [154, 258, 303, 318]\n- T3: [730, 281, 794, 294]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: unable to determine\nC: T3 -> T1 -> T2\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/03fbec1ea895c3803cbd4ad09f3e7cfdd748ee9c9cd4d6e795535db4b9634076.jpg", "reference_inputs/images/31c6abcf7eb4aa46f118f145da3d0ff14ff0c62da2378150b791ad7d61f0b10b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000069", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000069", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [436, 126, 458, 171]\n- T2: [475, 304, 493, 344]\n- T3: [506, 317, 527, 359]\n\nImage 2 (GROUND) labeled targets:\n- T1: [383, 278, 406, 299]\n- T2: [161, 250, 392, 351]\n- T3: [464, 173, 702, 370]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T2 -> T1\nC: T3 -> T1 -> T2\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d1c99ae970acc5d983e9425d72c5654176ba542ae95b32114c33ffcf0ce51a4c.jpg", "reference_inputs/images/1f14b718b819a632f51cf4ea6d5486cc08d1760dcb059aa48113fb4a35adc9bd.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000070", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000070", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [509, 199, 528, 243]\n- T2: [530, 493, 572, 536]\n- T3: [468, 414, 488, 458]\n\nImage 2 (GROUND) labeled targets:\n- T1: [508, 276, 534, 295]\n- T2: [651, 260, 887, 352]\n- T3: [351, 255, 439, 312]\n\nAnswer options:\nA: unable to determine\nB: T3 -> T2 -> T1\nC: T2 -> T1 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5f7449aa37f81ed4c60e2289ee7e1908e2c8a444221e73b2423da6704b840273.jpg", "reference_inputs/images/c33e1aeee82405ba47f657e32649e2e4e8d1e73b08d660d0fe811f4195c3ad2e.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/city6_ground_aerial_v2_50m/000071", "task_name": "distance_comparison", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000071", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [509, 199, 528, 243]\n- T2: [468, 414, 488, 458]\n- T3: [530, 493, 572, 536]\n\nImage 2 (GROUND) labeled targets:\n- T1: [508, 276, 534, 295]\n- T2: [351, 255, 439, 312]\n- T3: [651, 260, 887, 352]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T2 -> T3 -> T1\nC: T1 -> T2 -> T3\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5f7449aa37f81ed4c60e2289ee7e1908e2c8a444221e73b2423da6704b840273.jpg", "reference_inputs/images/c33e1aeee82405ba47f657e32649e2e4e8d1e73b08d660d0fe811f4195c3ad2e.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000001", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [990, 66, 1023, 176]\n- T2: [680, 113, 705, 130]\n- T3: [830, 338, 892, 445]\n\nImage 2 (GROUND) labeled targets:\n- T1: [303, 306, 418, 342]\n- T2: [69, 302, 126, 431]\n- T3: [607, 326, 784, 372]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T2 -> T1\nC: T2 -> T1 -> T3\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b23b723884fa900f71636d50bacb74594a28a0ad4ec9627cb72ada1841800999.jpg", "reference_inputs/images/99844ffd0082a8c59921a0d522eed8757940efe1d74aca3a8fa296e7efc3337a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000002", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [830, 338, 892, 445]\n- T2: [990, 66, 1023, 176]\n- T3: [680, 113, 705, 130]\n\nImage 2 (GROUND) labeled targets:\n- T1: [607, 326, 784, 372]\n- T2: [303, 306, 418, 342]\n- T3: [69, 302, 126, 431]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T1 -> T2 -> T3\nC: T3 -> T2 -> T1\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b23b723884fa900f71636d50bacb74594a28a0ad4ec9627cb72ada1841800999.jpg", "reference_inputs/images/99844ffd0082a8c59921a0d522eed8757940efe1d74aca3a8fa296e7efc3337a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000004", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [586, 402, 598, 412]\n- T2: [613, 389, 626, 399]\n- T3: [653, 429, 674, 440]\n\nImage 2 (GROUND) labeled targets:\n- T1: [245, 322, 413, 374]\n- T2: [84, 299, 219, 343]\n- T3: [728, 315, 828, 344]\n\nAnswer options:\nA: unable to determine\nB: T3 -> T2 -> T1\nC: T3 -> T1 -> T2\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ef914f0e8dd6ab8d097e08449220a058e4bf7d703d7db775bb65dc43e21a8c4c.jpg", "reference_inputs/images/7a0bf23ec2ab312ed51cb9da293f10c6cf147291121dfec77e840170a2766309.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000005", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [504, 440, 518, 456]\n- T2: [450, 430, 466, 444]\n- T3: [565, 472, 570, 490]\n- T4: [707, 472, 713, 490]\n\nImage 2 (GROUND) labeled targets:\n- T1: [498, 295, 523, 314]\n- T2: [410, 298, 439, 311]\n- T3: [623, 288, 632, 324]\n- T4: [912, 288, 924, 324]\n\nAnswer options:\nA: T3 -> T4 -> T2 -> T1\nB: T4 -> T3 -> T1 -> T2\nC: T3 -> T1 -> T4 -> T2\nD: T3 -> T4 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/532a7bb18fb47c2c7a6afad29eee2e22fe5703751da3632aa370b49a67654041.jpg", "reference_inputs/images/bb1b3b3955a3f2a6491e5176b2d74a97d2bd6453e48df238d57ef60e4f23c296.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000006", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [707, 472, 713, 490]\n- T2: [565, 472, 570, 490]\n- T3: [504, 440, 518, 456]\n- T4: [450, 430, 466, 444]\n\nImage 2 (GROUND) labeled targets:\n- T1: [912, 288, 924, 324]\n- T2: [623, 288, 632, 324]\n- T3: [498, 295, 523, 314]\n- T4: [410, 298, 439, 311]\n\nAnswer options:\nA: T2 -> T3 -> T1 -> T4\nB: T2 -> T1 -> T3 -> T4\nC: T1 -> T2 -> T3 -> T4\nD: T2 -> T1 -> T4 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/532a7bb18fb47c2c7a6afad29eee2e22fe5703751da3632aa370b49a67654041.jpg", "reference_inputs/images/bb1b3b3955a3f2a6491e5176b2d74a97d2bd6453e48df238d57ef60e4f23c296.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000007", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [348, 418, 367, 437]\n- T2: [570, 90, 590, 110]\n- T3: [546, 505, 556, 533]\n\nImage 2 (GROUND) labeled targets:\n- T1: [338, 292, 353, 339]\n- T2: [555, 293, 560, 319]\n- T3: [559, 290, 572, 353]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T2 -> T1\nC: T1 -> T3 -> T2\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/faf5c975d2ec136c30cf96efcb52dbaabe98daa7b1c8995bf160f483c0b5d4bc.jpg", "reference_inputs/images/1ae237aad839b445a89cc0e8f1fdecf1e406c23485f9ed36443a2be1a3304a8c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000008", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [628, 309, 645, 323]\n- T2: [970, 125, 1023, 224]\n- T3: [654, 159, 670, 174]\n\nImage 2 (GROUND) labeled targets:\n- T1: [674, 300, 743, 510]\n- T2: [358, 311, 465, 345]\n- T3: [141, 330, 194, 461]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T1 -> T3 -> T2\nC: T2 -> T1 -> T3\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9d4861e27c7af67f311d1114d48865240984cf9383313830a45ab9c71394e6ae.jpg", "reference_inputs/images/b28e4de0f5b59dbe34d9977ae7924f1d6f127b56999d871fee4297beb2c3d38a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000009", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [424, 469, 462, 488]\n- T2: [528, 438, 562, 454]\n- T3: [358, 514, 362, 533]\n- T4: [392, 538, 443, 569]\n\nImage 2 (GROUND) labeled targets:\n- T1: [437, 291, 473, 305]\n- T2: [530, 290, 563, 298]\n- T3: [352, 289, 358, 309]\n- T4: [387, 298, 441, 314]\n\nAnswer options:\nA: T4 -> T3 -> T2 -> T1\nB: T3 -> T4 -> T1 -> T2\nC: T4 -> T1 -> T3 -> T2\nD: T4 -> T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/17b62a74f74c1ea060246e106301a4bc1105d82173d30d693bf2f9d1dd4a700a.jpg", "reference_inputs/images/c050fcbf6a2fbb7ae73cec88f53be285420ba47b9cda22afa8774910d8ad59db.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000010", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [424, 469, 462, 488]\n- T2: [651, 565, 658, 575]\n- T3: [358, 514, 362, 533]\n- T4: [528, 438, 562, 454]\n\nImage 2 (GROUND) labeled targets:\n- T1: [437, 291, 473, 305]\n- T2: [658, 289, 666, 314]\n- T3: [352, 289, 358, 309]\n- T4: [530, 290, 563, 298]\n\nAnswer options:\nA: T3 -> T2 -> T1 -> T4\nB: T2 -> T1 -> T3 -> T4\nC: T2 -> T3 -> T1 -> T4\nD: T2 -> T3 -> T4 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/17b62a74f74c1ea060246e106301a4bc1105d82173d30d693bf2f9d1dd4a700a.jpg", "reference_inputs/images/c050fcbf6a2fbb7ae73cec88f53be285420ba47b9cda22afa8774910d8ad59db.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000011", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [506, 547, 512, 573]\n- T2: [715, 438, 753, 460]\n- T3: [720, 458, 765, 486]\n- T4: [822, 542, 831, 568]\n\nImage 2 (GROUND) labeled targets:\n- T1: [125, 290, 154, 338]\n- T2: [518, 293, 553, 304]\n- T3: [524, 296, 550, 310]\n- T4: [655, 287, 666, 320]\n\nAnswer options:\nA: T1 -> T3 -> T4 -> T2\nB: T1 -> T4 -> T3 -> T2\nC: T1 -> T4 -> T2 -> T3\nD: T4 -> T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d5bfe9a4746a03c92cbeee7518807d18504fafcbc4319a64d05fe881f42e9aab.jpg", "reference_inputs/images/daafd56b7f9b0584e99977588f276d6d9f561cd5540ec6e91fc46c16fc19f642.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000012", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [506, 547, 512, 573]\n- T2: [822, 542, 831, 568]\n- T3: [720, 458, 765, 486]\n- T4: [728, 484, 775, 504]\n\nImage 2 (GROUND) labeled targets:\n- T1: [125, 290, 154, 338]\n- T2: [655, 287, 666, 320]\n- T3: [524, 296, 550, 310]\n- T4: [534, 294, 566, 313]\n\nAnswer options:\nA: T1 -> T2 -> T3 -> T4\nB: T1 -> T2 -> T4 -> T3\nC: T1 -> T4 -> T2 -> T3\nD: T2 -> T1 -> T4 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d5bfe9a4746a03c92cbeee7518807d18504fafcbc4319a64d05fe881f42e9aab.jpg", "reference_inputs/images/daafd56b7f9b0584e99977588f276d6d9f561cd5540ec6e91fc46c16fc19f642.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000013", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [562, 420, 584, 434]\n- T2: [574, 510, 589, 539]\n- T3: [507, 122, 518, 142]\n\nImage 2 (GROUND) labeled targets:\n- T1: [562, 294, 596, 338]\n- T2: [598, 290, 617, 355]\n- T3: [508, 286, 516, 317]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T2 -> T3 -> T1\nC: T2 -> T1 -> T3\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/98588555a9656d7cfc9f63266d2fcf2b6eb3064482278c14ddac2c450e16d2c2.jpg", "reference_inputs/images/c32b20010a9f85d7f75228c727ce166db866aa2ad664200bc62120323d137ad8.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000014", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [518, 407, 532, 417]\n- T2: [429, 444, 455, 456]\n- T3: [538, 342, 547, 353]\n\nImage 2 (GROUND) labeled targets:\n- T1: [529, 302, 565, 322]\n- T2: [0, 326, 217, 386]\n- T3: [552, 280, 566, 295]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T1 -> T3 -> T2\nC: T2 -> T3 -> T1\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3af649bc1ded32eb4c23a0be47a0841d61cf1666558c6fd8369d70e79125362d.jpg", "reference_inputs/images/462d3dff450dfd5c907f20d1d934b7066d1a5c8b3121c61e4aef9b8e13af0603.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000015", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [538, 342, 547, 353]\n- T2: [518, 407, 532, 417]\n- T3: [429, 444, 455, 456]\n\nImage 2 (GROUND) labeled targets:\n- T1: [552, 280, 566, 295]\n- T2: [529, 302, 565, 322]\n- T3: [0, 326, 217, 386]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T1 -> T2\nC: T2 -> T1 -> T3\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3af649bc1ded32eb4c23a0be47a0841d61cf1666558c6fd8369d70e79125362d.jpg", "reference_inputs/images/462d3dff450dfd5c907f20d1d934b7066d1a5c8b3121c61e4aef9b8e13af0603.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000016", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [654, 179, 674, 197]\n- T2: [301, 252, 361, 360]\n- T3: [486, 373, 534, 482]\n\nImage 2 (GROUND) labeled targets:\n- T1: [726, 287, 731, 311]\n- T2: [538, 302, 574, 319]\n- T3: [668, 298, 722, 325]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T2 -> T1 -> T3\nC: T2 -> T3 -> T1\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d9e96895d6c2797f50c2aea2b79ee35dc34e37c5d5b54b575d4db8cd5807d9e2.jpg", "reference_inputs/images/7e335e03c825200cce7a9e42c6bbfaa0e960ed5feec24233d18da53e6340c06c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000017", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [249, 483, 260, 500]\n- T2: [181, 503, 190, 523]\n- T3: [198, 479, 202, 497]\n\nImage 2 (GROUND) labeled targets:\n- T1: [897, 290, 907, 307]\n- T2: [764, 289, 771, 303]\n- T3: [840, 289, 848, 303]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T1 -> T2 -> T3\nC: unable to determine\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d9984a287ff5447a4720a69174962625e9d679ce734ad6a6ccd16e9ae8fa2a56.jpg", "reference_inputs/images/40cf4580edadcb9d01aed3a17b06aad52789138e141da5f4c764ac23632deaba.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000018", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [398, 519, 450, 550]\n- T2: [529, 502, 573, 521]\n- T3: [422, 474, 463, 494]\n- T4: [534, 548, 593, 573]\n\nImage 2 (GROUND) labeled targets:\n- T1: [399, 291, 450, 310]\n- T2: [525, 294, 554, 306]\n- T3: [438, 291, 474, 305]\n- T4: [532, 298, 570, 313]\n\nAnswer options:\nA: T1 -> T4 -> T2 -> T3\nB: T4 -> T2 -> T1 -> T3\nC: T4 -> T1 -> T2 -> T3\nD: T4 -> T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8d623e668ea8bf1fb46683001b988f91917220a05ed970bf04e08492689ac461.jpg", "reference_inputs/images/f598fb5a843da633d25b3a73385b1f3542e3af3c5ae2199248979d89ad1ff582.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000019", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [389, 548, 442, 575]\n- T2: [530, 442, 564, 458]\n- T3: [354, 522, 359, 542]\n- T4: [529, 502, 573, 521]\n\nImage 2 (GROUND) labeled targets:\n- T1: [390, 298, 442, 314]\n- T2: [530, 290, 562, 298]\n- T3: [354, 289, 360, 308]\n- T4: [525, 294, 554, 306]\n\nAnswer options:\nA: T1 -> T4 -> T3 -> T2\nB: T1 -> T3 -> T2 -> T4\nC: T1 -> T3 -> T4 -> T2\nD: T3 -> T1 -> T4 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8d623e668ea8bf1fb46683001b988f91917220a05ed970bf04e08492689ac461.jpg", "reference_inputs/images/f598fb5a843da633d25b3a73385b1f3542e3af3c5ae2199248979d89ad1ff582.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000021", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [970, 539, 1014, 571]\n- T2: [617, 378, 818, 506]\n- T3: [171, 350, 284, 419]\n\nImage 2 (GROUND) labeled targets:\n- T1: [403, 288, 414, 325]\n- T2: [531, 276, 586, 331]\n- T3: [589, 307, 614, 319]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T1 -> T3 -> T2\nC: T2 -> T3 -> T1\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/01886ce0440f5aca0e712fc42566fbbda12378686622992519e2b014b4fe3e5b.jpg", "reference_inputs/images/2de10aa040bcdf1ae947d053dc6f1971086bd5f033b6425794a494edb23eb40d.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000022", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [917, 389, 942, 396]\n- T2: [794, 387, 820, 397]\n- T3: [802, 405, 833, 414]\n\nImage 2 (GROUND) labeled targets:\n- T1: [661, 296, 679, 304]\n- T2: [742, 295, 782, 314]\n- T3: [938, 300, 985, 310]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T3 -> T2 -> T1\nC: T2 -> T1 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/46ca6a46d8e89f5f6d2288e2ff737e8f34222d083d61b4f925c0d2b9633da384.jpg", "reference_inputs/images/356e298d4b7742b6b7c826b061b76cb8db7acd128030789aca33d2c107501fd6.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000023", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [794, 387, 820, 397]\n- T2: [802, 405, 833, 414]\n- T3: [917, 389, 942, 396]\n\nImage 2 (GROUND) labeled targets:\n- T1: [742, 295, 782, 314]\n- T2: [938, 300, 985, 310]\n- T3: [661, 296, 679, 304]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T1 -> T3 -> T2\nC: T1 -> T2 -> T3\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/46ca6a46d8e89f5f6d2288e2ff737e8f34222d083d61b4f925c0d2b9633da384.jpg", "reference_inputs/images/356e298d4b7742b6b7c826b061b76cb8db7acd128030789aca33d2c107501fd6.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000024", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000024", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [280, 244, 333, 350]\n- T2: [652, 388, 676, 406]\n- T3: [137, 164, 171, 178]\n\nImage 2 (GROUND) labeled targets:\n- T1: [96, 314, 261, 380]\n- T2: [829, 291, 860, 395]\n- T3: [102, 290, 122, 340]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T3 -> T1\nC: T2 -> T1 -> T3\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1a01ed403e1d284e43fa05e564f7484c77989d75c0705a6d46e99f6f062bb7ce.jpg", "reference_inputs/images/4c030ac1994474c88f320c0a101f7b988a4282a18b3f7adb0be9b0a96bdc2f49.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000026", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [588, 525, 594, 546]\n- T2: [611, 438, 631, 453]\n- T3: [581, 511, 586, 534]\n- T4: [591, 406, 605, 418]\n\nImage 2 (GROUND) labeled targets:\n- T1: [714, 294, 729, 350]\n- T2: [690, 298, 711, 312]\n- T3: [677, 290, 691, 342]\n- T4: [626, 292, 648, 304]\n\nAnswer options:\nA: T3 -> T1 -> T2 -> T4\nB: T1 -> T3 -> T2 -> T4\nC: T1 -> T2 -> T3 -> T4\nD: T1 -> T3 -> T4 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5decbba8daa7c5df5a7fff7c056b20766744bd23a17ecce6a98368f20a0ea9b9.jpg", "reference_inputs/images/b85a6b29f399b960fda0517b94c5f99319b190ff4a324961395b1f408498e61b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000027", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [611, 438, 631, 453]\n- T2: [581, 511, 586, 534]\n- T3: [588, 525, 594, 546]\n- T4: [438, 552, 444, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [690, 298, 711, 312]\n- T2: [677, 290, 691, 342]\n- T3: [714, 294, 729, 350]\n- T4: [271, 290, 293, 374]\n\nAnswer options:\nA: T4 -> T2 -> T3 -> T1\nB: T4 -> T3 -> T1 -> T2\nC: T4 -> T3 -> T2 -> T1\nD: T3 -> T4 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5decbba8daa7c5df5a7fff7c056b20766744bd23a17ecce6a98368f20a0ea9b9.jpg", "reference_inputs/images/b85a6b29f399b960fda0517b94c5f99319b190ff4a324961395b1f408498e61b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000028", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [479, 444, 505, 455]\n- T2: [646, 422, 666, 432]\n- T3: [559, 356, 573, 372]\n\nImage 2 (GROUND) labeled targets:\n- T1: [392, 308, 493, 334]\n- T2: [807, 302, 866, 318]\n- T3: [582, 278, 602, 298]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T2 -> T1 -> T3\nC: T2 -> T3 -> T1\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9b77a6fdcbcd164e42d030afdba126226416d464d2bb4b234b73eec459bc4446.jpg", "reference_inputs/images/9454bfbf127a6ef586c8cbd21a88516a04c24bf9b4a4d2d70da2bdcf1d16e104.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000029", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [646, 422, 666, 432]\n- T2: [479, 444, 505, 455]\n- T3: [559, 356, 573, 372]\n\nImage 2 (GROUND) labeled targets:\n- T1: [807, 302, 866, 318]\n- T2: [392, 308, 493, 334]\n- T3: [582, 278, 602, 298]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T1 -> T3\nC: T2 -> T3 -> T1\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9b77a6fdcbcd164e42d030afdba126226416d464d2bb4b234b73eec459bc4446.jpg", "reference_inputs/images/9454bfbf127a6ef586c8cbd21a88516a04c24bf9b4a4d2d70da2bdcf1d16e104.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000030", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [842, 422, 877, 433]\n- T2: [582, 426, 614, 437]\n- T3: [583, 442, 614, 482]\n- T4: [826, 399, 856, 410]\n\nImage 2 (GROUND) labeled targets:\n- T1: [385, 295, 410, 310]\n- T2: [2, 330, 157, 390]\n- T3: [597, 226, 1023, 418]\n- T4: [280, 291, 305, 300]\n\nAnswer options:\nA: T3 -> T2 -> T1 -> T4\nB: T3 -> T1 -> T2 -> T4\nC: T2 -> T3 -> T1 -> T4\nD: T3 -> T2 -> T4 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8502f61f181f5d74610587b6ad6b6b5a8d2308a26aaf20c07336a3dfbc0b0ca5.jpg", "reference_inputs/images/efa5658689482f7351dfb2d3e76efc3166d2ac78bdd9cedd302f88a92851f48f.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000031", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000031", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [826, 399, 856, 410]\n- T2: [582, 426, 614, 437]\n- T3: [842, 422, 877, 433]\n- T4: [583, 442, 614, 482]\n\nImage 2 (GROUND) labeled targets:\n- T1: [280, 291, 305, 300]\n- T2: [2, 330, 157, 390]\n- T3: [385, 295, 410, 310]\n- T4: [597, 226, 1023, 418]\n\nAnswer options:\nA: T4 -> T3 -> T2 -> T1\nB: T4 -> T2 -> T1 -> T3\nC: T4 -> T2 -> T3 -> T1\nD: T2 -> T4 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8502f61f181f5d74610587b6ad6b6b5a8d2308a26aaf20c07336a3dfbc0b0ca5.jpg", "reference_inputs/images/efa5658689482f7351dfb2d3e76efc3166d2ac78bdd9cedd302f88a92851f48f.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000032", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [202, 113, 314, 198]\n- T2: [950, 339, 987, 354]\n- T3: [820, 496, 938, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [82, 289, 113, 303]\n- T2: [218, 290, 222, 310]\n- T3: [42, 301, 90, 320]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T3 -> T1 -> T2\nC: T2 -> T3 -> T1\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/73152e00523860cf2ab782d32fe55c4a2d58139de4733b092fbc4707466ee32b.jpg", "reference_inputs/images/e5fa50935381e5c23b51f22b66a0923b5bc960a91acf1d6afb5864b7f013b278.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000033", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000033", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [330, 538, 338, 562]\n- T2: [530, 486, 571, 503]\n- T3: [408, 504, 456, 530]\n- T4: [383, 531, 440, 562]\n\nImage 2 (GROUND) labeled targets:\n- T1: [950, 298, 983, 347]\n- T2: [847, 306, 895, 326]\n- T3: [805, 309, 876, 339]\n- T4: [975, 325, 1023, 351]\n\nAnswer options:\nA: T1 -> T3 -> T4 -> T2\nB: T4 -> T1 -> T3 -> T2\nC: T1 -> T4 -> T3 -> T2\nD: T1 -> T4 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/45161aa0655e931e7924983dd2f4d431231d378aad482002b7e8acb1c3c8ed8a.jpg", "reference_inputs/images/dce70d33d8f9602df8bca856a63b807f674d48cc5394e78f6422eb83339193aa.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000034", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000034", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [537, 515, 588, 542]\n- T2: [330, 538, 338, 562]\n- T3: [530, 486, 571, 503]\n- T4: [408, 504, 456, 530]\n\nImage 2 (GROUND) labeled targets:\n- T1: [996, 302, 1023, 332]\n- T2: [950, 298, 983, 347]\n- T3: [847, 306, 895, 326]\n- T4: [805, 309, 876, 339]\n\nAnswer options:\nA: T2 -> T4 -> T1 -> T3\nB: T2 -> T1 -> T4 -> T3\nC: T2 -> T1 -> T3 -> T4\nD: T1 -> T2 -> T4 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/45161aa0655e931e7924983dd2f4d431231d378aad482002b7e8acb1c3c8ed8a.jpg", "reference_inputs/images/dce70d33d8f9602df8bca856a63b807f674d48cc5394e78f6422eb83339193aa.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000035", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000035", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [502, 378, 516, 398]\n- T2: [588, 374, 608, 379]\n- T3: [467, 388, 478, 399]\n\nImage 2 (GROUND) labeled targets:\n- T1: [659, 260, 771, 332]\n- T2: [834, 298, 873, 311]\n- T3: [550, 296, 600, 334]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T1 -> T3 -> T2\nC: unable to determine\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5bc11b7f2e38ac137619a4800f8fe2d0896a00695683ac0192ea35c976912819.jpg", "reference_inputs/images/bfc1fb5fe7cb11654b62cb5be6043c83f36ee5eba7f398484c2053fbf88b30c7.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000036", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000036", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [499, 403, 518, 430]\n- T2: [495, 549, 519, 575]\n- T3: [404, 510, 428, 538]\n\nImage 2 (GROUND) labeled targets:\n- T1: [502, 279, 516, 295]\n- T2: [502, 294, 517, 304]\n- T3: [438, 293, 453, 302]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T3 -> T1\nC: T3 -> T2 -> T1\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ab388bd18f6508e0a4495070aec8d68e3d050596d6b4de98d726f89bf804c871.jpg", "reference_inputs/images/49a8d354e7a01418072c7c697722506dd87cf39e7bd223bd9a07d8e47d694c7a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000038", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000038", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [415, 73, 527, 138]\n- T2: [185, 222, 299, 286]\n- T3: [204, 525, 278, 575]\n- T4: [129, 133, 168, 148]\n\nImage 2 (GROUND) labeled targets:\n- T1: [446, 311, 498, 329]\n- T2: [396, 310, 463, 337]\n- T3: [587, 333, 666, 364]\n- T4: [341, 302, 348, 327]\n\nAnswer options:\nA: T1 -> T2 -> T3 -> T4\nB: T1 -> T4 -> T2 -> T3\nC: T1 -> T2 -> T4 -> T3\nD: T2 -> T1 -> T4 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c1fe12231e3c6505f4561f7b59d589c5fc02488b59c32ec3725a2de26e313288.jpg", "reference_inputs/images/12612aa50c4da7e27c053652a3c4e25219e5ca62e5385dfabd28088fd37f4ec1.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000039", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000039", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [785, 497, 846, 575]\n- T2: [694, 450, 717, 468]\n- T3: [609, 208, 639, 223]\n\nImage 2 (GROUND) labeled targets:\n- T1: [739, 301, 849, 334]\n- T2: [649, 291, 669, 321]\n- T3: [567, 289, 582, 311]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T1 -> T2 -> T3\nC: unable to determine\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f3f337c634a9cc8fb5161273331a95970b7f1312fe185c12bb1a6b550e321af6.jpg", "reference_inputs/images/cb10762cc79fb403ed3d060779fd177fdbc256d05c9dc19842a408604ffd0524.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000040", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000040", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [785, 497, 846, 575]\n- T2: [694, 450, 717, 468]\n- T3: [609, 208, 639, 223]\n\nImage 2 (GROUND) labeled targets:\n- T1: [739, 301, 849, 334]\n- T2: [649, 291, 669, 321]\n- T3: [567, 289, 582, 311]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T1 -> T2\nC: T3 -> T2 -> T1\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f3f337c634a9cc8fb5161273331a95970b7f1312fe185c12bb1a6b550e321af6.jpg", "reference_inputs/images/cb10762cc79fb403ed3d060779fd177fdbc256d05c9dc19842a408604ffd0524.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000043", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000043", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [749, 466, 757, 483]\n- T2: [648, 510, 682, 540]\n- T3: [599, 534, 610, 556]\n\nImage 2 (GROUND) labeled targets:\n- T1: [146, 288, 157, 325]\n- T2: [82, 326, 245, 375]\n- T3: [222, 308, 290, 440]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T2 -> T1 -> T3\nC: T3 -> T2 -> T1\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4b34f4cb71a2d71d19dd3baacb4efbffba03c9ad7a50932e6b3432ed3969757b.jpg", "reference_inputs/images/5c973ec6b63e4c1d2c149201ffa5cd3b1388a2ea9967a81b72eefd098be5952a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000044", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000044", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [648, 510, 682, 540]\n- T2: [599, 534, 610, 556]\n- T3: [749, 466, 757, 483]\n\nImage 2 (GROUND) labeled targets:\n- T1: [82, 326, 245, 375]\n- T2: [222, 308, 290, 440]\n- T3: [146, 288, 157, 325]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T1 -> T2 -> T3\nC: unable to determine\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4b34f4cb71a2d71d19dd3baacb4efbffba03c9ad7a50932e6b3432ed3969757b.jpg", "reference_inputs/images/5c973ec6b63e4c1d2c149201ffa5cd3b1388a2ea9967a81b72eefd098be5952a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000045", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000045", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [694, 376, 720, 397]\n- T2: [653, 330, 676, 346]\n- T3: [506, 0, 620, 55]\n\nImage 2 (GROUND) labeled targets:\n- T1: [798, 294, 814, 346]\n- T2: [761, 293, 783, 352]\n- T3: [254, 331, 362, 391]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T1 -> T2\nC: T2 -> T1 -> T3\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a86b10af6f61a2902a8ff2549f52ec7d4696fe04c8625701878839ce4cffcbbc.jpg", "reference_inputs/images/b222eeecc0a0b3660e5632c382b7f4bdbff21bc456f06d0165822026551dda65.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000046", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000046", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [332, 432, 347, 445]\n- T2: [347, 413, 362, 422]\n- T3: [298, 453, 334, 494]\n\nImage 2 (GROUND) labeled targets:\n- T1: [740, 296, 826, 328]\n- T2: [927, 325, 1023, 369]\n- T3: [641, 273, 697, 311]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T1 -> T3 -> T2\nC: T2 -> T1 -> T3\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2d9c0d21686be8efdd9a86bf727dc62b7458f9c9b66e2b26748d02e03988a651.jpg", "reference_inputs/images/696db644fc0e650348f4be016e8e5fd9c181eee2de39d8b0f2f71ba4a8fde652.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000047", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000047", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [347, 413, 362, 422]\n- T2: [298, 453, 334, 494]\n- T3: [332, 432, 347, 445]\n\nImage 2 (GROUND) labeled targets:\n- T1: [927, 325, 1023, 369]\n- T2: [641, 273, 697, 311]\n- T3: [740, 296, 826, 328]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T2 -> T1\nC: T2 -> T1 -> T3\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2d9c0d21686be8efdd9a86bf727dc62b7458f9c9b66e2b26748d02e03988a651.jpg", "reference_inputs/images/696db644fc0e650348f4be016e8e5fd9c181eee2de39d8b0f2f71ba4a8fde652.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000048", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000048", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [595, 383, 618, 389]\n- T2: [511, 385, 533, 392]\n- T3: [698, 382, 721, 390]\n- T4: [462, 400, 474, 414]\n\nImage 2 (GROUND) labeled targets:\n- T1: [717, 298, 778, 311]\n- T2: [529, 298, 568, 313]\n- T3: [977, 294, 1023, 310]\n- T4: [313, 296, 379, 336]\n\nAnswer options:\nA: T4 -> T1 -> T2 -> T3\nB: T4 -> T2 -> T3 -> T1\nC: T4 -> T2 -> T1 -> T3\nD: T2 -> T4 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fc020544619452bceeb4c01d861680999eff0a57983a98dd5efea5e57ebef25d.jpg", "reference_inputs/images/6c1faeec1b212385a6473aba22bc10edfcccad8ee4a4d132b5fa6116fc167924.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000049", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000049", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [698, 382, 721, 390]\n- T2: [702, 370, 723, 378]\n- T3: [595, 383, 618, 389]\n- T4: [462, 400, 474, 414]\n\nImage 2 (GROUND) labeled targets:\n- T1: [977, 294, 1023, 310]\n- T2: [921, 290, 946, 297]\n- T3: [717, 298, 778, 311]\n- T4: [313, 296, 379, 336]\n\nAnswer options:\nA: T4 -> T1 -> T3 -> T2\nB: T4 -> T3 -> T2 -> T1\nC: T4 -> T3 -> T1 -> T2\nD: T3 -> T4 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fc020544619452bceeb4c01d861680999eff0a57983a98dd5efea5e57ebef25d.jpg", "reference_inputs/images/6c1faeec1b212385a6473aba22bc10edfcccad8ee4a4d132b5fa6116fc167924.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000050", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000050", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [577, 432, 592, 444]\n- T2: [522, 481, 541, 504]\n- T3: [454, 511, 459, 534]\n\nImage 2 (GROUND) labeled targets:\n- T1: [590, 294, 609, 304]\n- T2: [526, 295, 550, 314]\n- T3: [433, 287, 440, 316]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T2 -> T3 -> T1\nC: T3 -> T1 -> T2\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fb34b53b679d1237735335f190243505fad3e06e30ee7b8151f738312973ddd5.jpg", "reference_inputs/images/fb83154e0a62656e6b5c69979f54dd7b956389860903523b5d68cc091995ba8d.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000051", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000051", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [454, 511, 459, 534]\n- T2: [685, 494, 689, 514]\n- T3: [577, 432, 592, 444]\n\nImage 2 (GROUND) labeled targets:\n- T1: [433, 287, 440, 316]\n- T2: [742, 286, 749, 312]\n- T3: [590, 294, 609, 304]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T2 -> T1 -> T3\nC: T3 -> T1 -> T2\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fb34b53b679d1237735335f190243505fad3e06e30ee7b8151f738312973ddd5.jpg", "reference_inputs/images/fb83154e0a62656e6b5c69979f54dd7b956389860903523b5d68cc091995ba8d.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000052", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000052", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [887, 494, 938, 522]\n- T2: [711, 530, 766, 557]\n- T3: [798, 428, 832, 439]\n\nImage 2 (GROUND) labeled targets:\n- T1: [1006, 291, 1023, 307]\n- T2: [819, 292, 875, 310]\n- T3: [896, 293, 951, 303]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T1 -> T3 -> T2\nC: T2 -> T3 -> T1\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f459bf256a51aa8a9babcda85af13e41bdfe56751b2ca2d2cae1fc8e8bcc5e15.jpg", "reference_inputs/images/5343fb3594b5630345448d98043996f02515530e2eddc6c3b7565939086db183.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000053", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000053", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [798, 428, 832, 439]\n- T2: [711, 530, 766, 557]\n- T3: [887, 494, 938, 522]\n\nImage 2 (GROUND) labeled targets:\n- T1: [896, 293, 951, 303]\n- T2: [819, 292, 875, 310]\n- T3: [1006, 291, 1023, 307]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T1 -> T2 -> T3\nC: T3 -> T2 -> T1\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f459bf256a51aa8a9babcda85af13e41bdfe56751b2ca2d2cae1fc8e8bcc5e15.jpg", "reference_inputs/images/5343fb3594b5630345448d98043996f02515530e2eddc6c3b7565939086db183.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000054", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000054", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [871, 98, 908, 118]\n- T2: [631, 329, 646, 338]\n- T3: [781, 534, 806, 560]\n\nImage 2 (GROUND) labeled targets:\n- T1: [419, 291, 426, 320]\n- T2: [330, 292, 334, 313]\n- T3: [172, 294, 180, 318]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T1 -> T3 -> T2\nC: T2 -> T1 -> T3\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/96282862df9fed1e54bdafbd1cefb080486775ae3b35c0d1cf0a80741f19ab67.jpg", "reference_inputs/images/6c19efe1630f6eca1709b8bf989300cf0ebc534fdb36b20b60aeb358ef605f59.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000055", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000055", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [871, 98, 908, 118]\n- T2: [575, 0, 589, 22]\n- T3: [658, 364, 678, 379]\n\nImage 2 (GROUND) labeled targets:\n- T1: [419, 291, 426, 320]\n- T2: [493, 290, 498, 311]\n- T3: [306, 291, 311, 314]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T2 -> T1\nC: T2 -> T3 -> T1\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/96282862df9fed1e54bdafbd1cefb080486775ae3b35c0d1cf0a80741f19ab67.jpg", "reference_inputs/images/6c19efe1630f6eca1709b8bf989300cf0ebc534fdb36b20b60aeb358ef605f59.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000056", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000056", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [970, 370, 1023, 468]\n- T2: [626, 551, 643, 575]\n- T3: [797, 118, 853, 217]\n- T4: [654, 409, 669, 424]\n\nImage 2 (GROUND) labeled targets:\n- T1: [888, 305, 985, 333]\n- T2: [638, 288, 662, 338]\n- T3: [679, 299, 722, 318]\n- T4: [632, 295, 642, 326]\n\nAnswer options:\nA: T2 -> T4 -> T3 -> T1\nB: T2 -> T4 -> T1 -> T3\nC: T4 -> T2 -> T1 -> T3\nD: T2 -> T1 -> T4 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b8200a86bbc7cdf8a4e6f13a73c77702bd6d6aef0e6249858945696f82d3c305.jpg", "reference_inputs/images/8e90901750b16872fdaa935ea78ac05d94b56c4f0c1df5921fade7be28e20261.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000057", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000057", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [970, 370, 1023, 468]\n- T2: [626, 551, 643, 575]\n- T3: [654, 409, 669, 424]\n\nImage 2 (GROUND) labeled targets:\n- T1: [888, 305, 985, 333]\n- T2: [638, 288, 662, 338]\n- T3: [632, 295, 642, 326]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T1 -> T2\nC: T3 -> T2 -> T1\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b8200a86bbc7cdf8a4e6f13a73c77702bd6d6aef0e6249858945696f82d3c305.jpg", "reference_inputs/images/8e90901750b16872fdaa935ea78ac05d94b56c4f0c1df5921fade7be28e20261.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000058", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000058", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [961, 496, 1023, 575]\n- T2: [652, 297, 760, 382]\n- T3: [769, 289, 875, 368]\n\nImage 2 (GROUND) labeled targets:\n- T1: [665, 377, 790, 423]\n- T2: [0, 360, 202, 465]\n- T3: [157, 335, 383, 410]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T2 -> T3 -> T1\nC: T3 -> T2 -> T1\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0671ae8ee169440e6627d92e84b2ba53644dd471a3dc9097fb5bbae57d101527.jpg", "reference_inputs/images/dc50adabea5c473de95da4c3e3bf5ddb25880f2dd43b3eb3b1e235dc6a75a193.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000059", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000059", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [769, 289, 875, 368]\n- T2: [652, 297, 760, 382]\n- T3: [961, 496, 1023, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [157, 335, 383, 410]\n- T2: [0, 360, 202, 465]\n- T3: [665, 377, 790, 423]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T1 -> T2 -> T3\nC: T2 -> T1 -> T3\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0671ae8ee169440e6627d92e84b2ba53644dd471a3dc9097fb5bbae57d101527.jpg", "reference_inputs/images/dc50adabea5c473de95da4c3e3bf5ddb25880f2dd43b3eb3b1e235dc6a75a193.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000060", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000060", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [445, 429, 459, 443]\n- T2: [410, 519, 436, 553]\n- T3: [506, 455, 524, 474]\n\nImage 2 (GROUND) labeled targets:\n- T1: [526, 294, 548, 310]\n- T2: [572, 310, 646, 361]\n- T3: [657, 294, 705, 320]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T2 -> T3 -> T1\nC: T1 -> T2 -> T3\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/eecc5fd6027c2498737cfd1fb08d36e7c6fe4d6927d160f242cd9a1edbb2dc0e.jpg", "reference_inputs/images/0270e6c49c8feef3c5b6875c7c99e75229547dceb165bee3d8a659b59ba4fe54.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000061", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000061", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [445, 429, 459, 443]\n- T2: [506, 455, 524, 474]\n- T3: [410, 519, 436, 553]\n\nImage 2 (GROUND) labeled targets:\n- T1: [526, 294, 548, 310]\n- T2: [657, 294, 705, 320]\n- T3: [572, 310, 646, 361]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T2 -> T1 -> T3\nC: T3 -> T1 -> T2\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/eecc5fd6027c2498737cfd1fb08d36e7c6fe4d6927d160f242cd9a1edbb2dc0e.jpg", "reference_inputs/images/0270e6c49c8feef3c5b6875c7c99e75229547dceb165bee3d8a659b59ba4fe54.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000063", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000063", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [675, 533, 737, 575]\n- T2: [685, 470, 710, 486]\n- T3: [561, 484, 566, 502]\n- T4: [602, 439, 619, 454]\n\nImage 2 (GROUND) labeled targets:\n- T1: [738, 261, 882, 337]\n- T2: [723, 296, 737, 309]\n- T3: [573, 286, 582, 310]\n- T4: [619, 294, 640, 305]\n\nAnswer options:\nA: T1 -> T3 -> T2 -> T4\nB: T3 -> T1 -> T2 -> T4\nC: T1 -> T2 -> T3 -> T4\nD: T1 -> T3 -> T4 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/01986be9f391684280a129ee31aa9ba31926cad35759a3b95020608ed7dfb573.jpg", "reference_inputs/images/15efa56c21d2bc6e235bb1a3ccf4483e7a48fdd00089930beb82826e589d1667.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000064", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000064", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [675, 533, 737, 575]\n- T2: [685, 470, 710, 486]\n- T3: [602, 439, 619, 454]\n- T4: [561, 484, 566, 502]\n\nImage 2 (GROUND) labeled targets:\n- T1: [738, 261, 882, 337]\n- T2: [723, 296, 737, 309]\n- T3: [619, 294, 640, 305]\n- T4: [573, 286, 582, 310]\n\nAnswer options:\nA: T1 -> T4 -> T2 -> T3\nB: T1 -> T4 -> T3 -> T2\nC: unable to determine\nD: T1 -> T2 -> T4 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/01986be9f391684280a129ee31aa9ba31926cad35759a3b95020608ed7dfb573.jpg", "reference_inputs/images/15efa56c21d2bc6e235bb1a3ccf4483e7a48fdd00089930beb82826e589d1667.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000067", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000067", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [969, 90, 1016, 114]\n- T2: [270, 298, 363, 407]\n- T3: [70, 303, 106, 315]\n\nImage 2 (GROUND) labeled targets:\n- T1: [732, 325, 739, 348]\n- T2: [524, 334, 599, 385]\n- T3: [221, 270, 254, 340]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T2 -> T3 -> T1\nC: T2 -> T1 -> T3\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2f5b9c5803ca4db366c311011221b5a28998dfc282cf3612aa0248beaeaaa7a9.jpg", "reference_inputs/images/2b0085602fd1cf33b6065a19d446355c48f9ee3e442de955cf7b10c5227f648e.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000068", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000068", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [969, 90, 1016, 114]\n- T2: [70, 303, 106, 315]\n- T3: [270, 298, 363, 407]\n\nImage 2 (GROUND) labeled targets:\n- T1: [732, 325, 739, 348]\n- T2: [221, 270, 254, 340]\n- T3: [524, 334, 599, 385]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T3 -> T1\nC: T3 -> T1 -> T2\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2f5b9c5803ca4db366c311011221b5a28998dfc282cf3612aa0248beaeaaa7a9.jpg", "reference_inputs/images/2b0085602fd1cf33b6065a19d446355c48f9ee3e442de955cf7b10c5227f648e.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000069", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000069", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [861, 530, 870, 552]\n- T2: [903, 560, 909, 575]\n- T3: [651, 518, 685, 550]\n\nImage 2 (GROUND) labeled targets:\n- T1: [877, 302, 891, 333]\n- T2: [949, 309, 959, 333]\n- T3: [954, 326, 1023, 370]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T1 -> T2\nC: T3 -> T2 -> T1\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/87d5cebc84b97e0be052460d2ca2b4ada12d3b9a1a0a465e6c4d5e9d259bbbbc.jpg", "reference_inputs/images/a6f0dc1acff8a9f1bda04ce3966e51b2cf5fe48703546cf8bfefc2aeb8f55b32.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000071", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000071", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [639, 418, 658, 430]\n- T2: [608, 438, 625, 452]\n- T3: [488, 406, 506, 414]\n\nImage 2 (GROUND) labeled targets:\n- T1: [578, 293, 611, 305]\n- T2: [666, 297, 689, 306]\n- T3: [514, 301, 541, 314]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: unable to determine\nC: T2 -> T3 -> T1\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4eb701386ccf91f87d8153ba6786d69689da1ca369180d1c62781fc6f8b641d2.jpg", "reference_inputs/images/d0eeb6b43019302b751b886be04ae065f56c73947ae476b6c614edaf1be9df61.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000072", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000072", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [608, 438, 625, 452]\n- T2: [488, 406, 506, 414]\n- T3: [639, 418, 658, 430]\n\nImage 2 (GROUND) labeled targets:\n- T1: [666, 297, 689, 306]\n- T2: [514, 301, 541, 314]\n- T3: [578, 293, 611, 305]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T1 -> T3 -> T2\nC: T3 -> T1 -> T2\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4eb701386ccf91f87d8153ba6786d69689da1ca369180d1c62781fc6f8b641d2.jpg", "reference_inputs/images/d0eeb6b43019302b751b886be04ae065f56c73947ae476b6c614edaf1be9df61.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000073", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000073", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [532, 472, 573, 492]\n- T2: [527, 536, 578, 558]\n- T3: [409, 484, 454, 505]\n\nImage 2 (GROUND) labeled targets:\n- T1: [594, 294, 634, 307]\n- T2: [633, 299, 665, 314]\n- T3: [494, 298, 518, 309]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T2 -> T3 -> T1\nC: T2 -> T1 -> T3\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0051393cd63d511e7658334ca57f89ab965af6108304cd2d8b82d0ed36c77437.jpg", "reference_inputs/images/6dd7fca2b6d17dd9c99effce28ce8aa2c7b8f274bbff06780d240f09f540dca2.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000074", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000074", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [527, 536, 578, 558]\n- T2: [399, 518, 450, 546]\n- T3: [409, 484, 454, 505]\n\nImage 2 (GROUND) labeled targets:\n- T1: [633, 299, 665, 314]\n- T2: [507, 295, 546, 312]\n- T3: [494, 298, 518, 309]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T1 -> T3\nC: T1 -> T3 -> T2\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0051393cd63d511e7658334ca57f89ab965af6108304cd2d8b82d0ed36c77437.jpg", "reference_inputs/images/6dd7fca2b6d17dd9c99effce28ce8aa2c7b8f274bbff06780d240f09f540dca2.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000075", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000075", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [336, 266, 355, 285]\n- T2: [393, 384, 415, 401]\n- T3: [297, 9, 320, 42]\n\nImage 2 (GROUND) labeled targets:\n- T1: [747, 289, 757, 326]\n- T2: [728, 293, 748, 337]\n- T3: [814, 290, 822, 317]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T1 -> T3\nC: T2 -> T3 -> T1\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fb55ae6fe0a5e87f2a974b92788564c804ef2b673d2f3f9e41ec8f55f4f06477.jpg", "reference_inputs/images/8acb44882ad014a1644004c081aaf072d078f577d9277171e12f3dedf544739b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000076", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000076", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [491, 393, 506, 400]\n- T2: [595, 418, 610, 430]\n- T3: [624, 402, 639, 413]\n- T4: [626, 489, 654, 514]\n\nImage 2 (GROUND) labeled targets:\n- T1: [475, 294, 504, 303]\n- T2: [666, 297, 701, 313]\n- T3: [702, 290, 736, 307]\n- T4: [886, 307, 1023, 376]\n\nAnswer options:\nA: T4 -> T3 -> T2 -> T1\nB: T4 -> T2 -> T3 -> T1\nC: T2 -> T4 -> T3 -> T1\nD: T4 -> T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a05c6e1fc084f2c02448041fed5d2848f797979541b7617a3dcd203089276c02.jpg", "reference_inputs/images/a64a4eeffc4e041cfee0750c7ca09cfaaa10ab0dd50029a956576eef779531a2.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000077", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000077", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [624, 402, 639, 413]\n- T2: [491, 393, 506, 400]\n- T3: [626, 489, 654, 514]\n- T4: [595, 418, 610, 430]\n\nImage 2 (GROUND) labeled targets:\n- T1: [702, 290, 736, 307]\n- T2: [475, 294, 504, 303]\n- T3: [886, 307, 1023, 376]\n- T4: [666, 297, 701, 313]\n\nAnswer options:\nA: T4 -> T3 -> T1 -> T2\nB: T3 -> T1 -> T4 -> T2\nC: T3 -> T4 -> T2 -> T1\nD: T3 -> T4 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a05c6e1fc084f2c02448041fed5d2848f797979541b7617a3dcd203089276c02.jpg", "reference_inputs/images/a64a4eeffc4e041cfee0750c7ca09cfaaa10ab0dd50029a956576eef779531a2.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000078", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000078", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [602, 423, 618, 436]\n- T2: [658, 435, 679, 450]\n- T3: [575, 398, 589, 407]\n- T4: [629, 465, 654, 485]\n\nImage 2 (GROUND) labeled targets:\n- T1: [664, 294, 698, 311]\n- T2: [774, 295, 828, 314]\n- T3: [614, 294, 630, 303]\n- T4: [758, 306, 839, 330]\n\nAnswer options:\nA: unable to determine\nB: T4 -> T2 -> T1 -> T3\nC: T4 -> T1 -> T2 -> T3\nD: T4 -> T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c391ab20637d039794ce863c1b9db7c92b6d6325ce8db865c9cab7217923f7de.jpg", "reference_inputs/images/fa38e654e7d24f439480ca85078b43de575e08da51ef24adf0cb83b923b47e43.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000079", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000079", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [602, 423, 618, 436]\n- T2: [629, 465, 654, 485]\n- T3: [658, 435, 679, 450]\n- T4: [575, 398, 589, 407]\n\nImage 2 (GROUND) labeled targets:\n- T1: [664, 294, 698, 311]\n- T2: [758, 306, 839, 330]\n- T3: [774, 295, 828, 314]\n- T4: [614, 294, 630, 303]\n\nAnswer options:\nA: T2 -> T1 -> T3 -> T4\nB: unable to determine\nC: T3 -> T2 -> T1 -> T4\nD: T2 -> T3 -> T1 -> T4\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c391ab20637d039794ce863c1b9db7c92b6d6325ce8db865c9cab7217923f7de.jpg", "reference_inputs/images/fa38e654e7d24f439480ca85078b43de575e08da51ef24adf0cb83b923b47e43.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000082", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000082", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [710, 520, 769, 542]\n- T2: [510, 532, 566, 558]\n- T3: [600, 470, 630, 485]\n\nImage 2 (GROUND) labeled targets:\n- T1: [796, 302, 868, 321]\n- T2: [546, 303, 587, 322]\n- T3: [624, 298, 665, 310]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T2 -> T1 -> T3\nC: T1 -> T3 -> T2\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/dabcddb2c5f31308ba63bf2401e624fa4873357724a5c22c4e346a03c97f202d.jpg", "reference_inputs/images/c871a2b6bab048eb4995aff931bf7060b4d846a70080dfeef714fc030b01d5d7.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000084", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000084", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [506, 437, 519, 453]\n- T2: [434, 470, 454, 490]\n- T3: [567, 475, 571, 493]\n- T4: [454, 424, 467, 437]\n\nImage 2 (GROUND) labeled targets:\n- T1: [503, 289, 522, 302]\n- T2: [413, 294, 438, 311]\n- T3: [585, 286, 590, 310]\n- T4: [442, 291, 458, 303]\n\nAnswer options:\nA: T3 -> T2 -> T4 -> T1\nB: T2 -> T3 -> T1 -> T4\nC: T3 -> T2 -> T1 -> T4\nD: T3 -> T1 -> T2 -> T4\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/58535b94fbff0387d3d936599a30966b8283f1f79f82b2557b1e6cb15cb992b7.jpg", "reference_inputs/images/d364b0e3541250a939545692adbf2b7f5dd4eda4eed7a82b53526764b5620522.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000085", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000085", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [723, 558, 730, 575]\n- T2: [362, 516, 366, 538]\n- T3: [522, 475, 543, 497]\n- T4: [519, 426, 533, 440]\n\nImage 2 (GROUND) labeled targets:\n- T1: [598, 288, 605, 320]\n- T2: [179, 286, 187, 313]\n- T3: [408, 297, 436, 310]\n- T4: [438, 291, 454, 302]\n\nAnswer options:\nA: T1 -> T2 -> T3 -> T4\nB: T2 -> T1 -> T3 -> T4\nC: T1 -> T3 -> T2 -> T4\nD: T1 -> T2 -> T4 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/07ca824b99b314bd94d57eda78a95cebe85ce30bc8e7d8698d800d83e704a7c3.jpg", "reference_inputs/images/1075720d899aa88c5aa14cb7eeb20344e150199624b9bde5c85be648db3d26cb.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000086", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000086", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [362, 516, 366, 538]\n- T2: [723, 558, 730, 575]\n- T3: [605, 474, 626, 493]\n- T4: [519, 426, 533, 440]\n\nImage 2 (GROUND) labeled targets:\n- T1: [179, 286, 187, 313]\n- T2: [598, 288, 605, 320]\n- T3: [505, 291, 528, 308]\n- T4: [438, 291, 454, 302]\n\nAnswer options:\nA: T2 -> T1 -> T3 -> T4\nB: unable to determine\nC: T1 -> T2 -> T3 -> T4\nD: T2 -> T1 -> T4 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/07ca824b99b314bd94d57eda78a95cebe85ce30bc8e7d8698d800d83e704a7c3.jpg", "reference_inputs/images/1075720d899aa88c5aa14cb7eeb20344e150199624b9bde5c85be648db3d26cb.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000087", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000087", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [645, 227, 662, 259]\n- T2: [118, 394, 155, 410]\n- T3: [130, 158, 165, 177]\n\nImage 2 (GROUND) labeled targets:\n- T1: [770, 290, 782, 320]\n- T2: [375, 290, 386, 325]\n- T3: [440, 287, 447, 314]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T2 -> T1\nC: T2 -> T1 -> T3\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/41821988d9f59431e844993ef642182ec744c51695610e657e3aa61b90f80b96.jpg", "reference_inputs/images/5b5e53f763930b5c8ada4a79985b9f6114cb0b759d93796acd02cb2dc9a44557.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000088", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000088", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [424, 489, 449, 513]\n- T2: [501, 509, 522, 538]\n- T3: [507, 401, 518, 412]\n\nImage 2 (GROUND) labeled targets:\n- T1: [527, 294, 546, 303]\n- T2: [598, 292, 618, 306]\n- T3: [553, 287, 572, 296]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T3 -> T1\nC: T2 -> T1 -> T3\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/879e112e284022f2692d0d7d519fe264dcfb695bd5472fce5af5345ec094d295.jpg", "reference_inputs/images/1e7b3515bbc18e19e874005bb680a85b86b524dbafd95cdbb9a85308de0ac895.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000089", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000089", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [507, 197, 518, 213]\n- T2: [707, 266, 737, 279]\n- T3: [838, 86, 880, 110]\n\nImage 2 (GROUND) labeled targets:\n- T1: [510, 287, 516, 310]\n- T2: [625, 290, 632, 304]\n- T3: [674, 286, 682, 307]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T3 -> T1\nC: T2 -> T1 -> T3\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/38b14bfd106225a2e22f062250eb7d51bddb58c2ae67d4a7ab3f274540ea8a91.jpg", "reference_inputs/images/57ae96560ad6020664e5a95327c1992dcf7337d3980ecf843a71447b788d5bfd.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000090", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000090", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [401, 382, 414, 395]\n- T2: [887, 546, 935, 574]\n- T3: [538, 361, 582, 459]\n- T4: [903, 22, 948, 47]\n\nImage 2 (GROUND) labeled targets:\n- T1: [251, 288, 262, 319]\n- T2: [636, 289, 655, 330]\n- T3: [334, 302, 382, 326]\n- T4: [583, 289, 592, 308]\n\nAnswer options:\nA: T3 -> T2 -> T1 -> T4\nB: T2 -> T3 -> T1 -> T4\nC: T2 -> T1 -> T3 -> T4\nD: T2 -> T3 -> T4 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/71627e38503f8fa008a5c48ecbf7cb63719a3f00c0377de6c1e90fbd4fd13671.jpg", "reference_inputs/images/18af91d51cab25cd269148c24de3e6c462001bcbe4cdf2ad120569133edaefec.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000091", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000091", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [371, 435, 393, 446]\n- T2: [624, 535, 671, 575]\n- T3: [621, 518, 659, 546]\n- T4: [510, 510, 534, 534]\n\nImage 2 (GROUND) labeled targets:\n- T1: [461, 334, 557, 391]\n- T2: [811, 298, 872, 312]\n- T3: [760, 302, 815, 317]\n- T4: [805, 308, 887, 325]\n\nAnswer options:\nA: T1 -> T4 -> T2 -> T3\nB: T1 -> T3 -> T4 -> T2\nC: T4 -> T1 -> T3 -> T2\nD: T1 -> T4 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d76465ef7fbd5baaf0fe865c9d44135d8dba217611268bbefb6586e6993147b2.jpg", "reference_inputs/images/c52d3ed15c8ad4699b4bc938831c5b06416414b65cdf6d09847efe920d0a1925.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000092", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000092", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [371, 435, 393, 446]\n- T2: [621, 518, 659, 546]\n- T3: [624, 535, 671, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [461, 334, 557, 391]\n- T2: [760, 302, 815, 317]\n- T3: [811, 298, 872, 312]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T1 -> T3 -> T2\nC: T3 -> T2 -> T1\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d76465ef7fbd5baaf0fe865c9d44135d8dba217611268bbefb6586e6993147b2.jpg", "reference_inputs/images/c52d3ed15c8ad4699b4bc938831c5b06416414b65cdf6d09847efe920d0a1925.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000093", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000093", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [649, 530, 682, 550]\n- T2: [809, 134, 870, 243]\n- T3: [570, 431, 584, 451]\n\nImage 2 (GROUND) labeled targets:\n- T1: [656, 293, 690, 335]\n- T2: [689, 301, 739, 318]\n- T3: [561, 287, 572, 327]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T2 -> T3\nC: T1 -> T3 -> T2\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/05b6f20e4d9f1346db76bff920938e567e5ec7a0b5c20003a050d61e4e631ef7.jpg", "reference_inputs/images/153c42acc503b7f49f9717f56313e3f98e87cf0cd70405ca91151993eafe3310.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000094", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000094", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [791, 480, 797, 498]\n- T2: [626, 474, 650, 497]\n- T3: [570, 484, 581, 500]\n\nImage 2 (GROUND) labeled targets:\n- T1: [612, 291, 621, 325]\n- T2: [558, 318, 741, 379]\n- T3: [972, 330, 1023, 455]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T3 -> T1 -> T2\nC: T2 -> T3 -> T1\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1b772583e746028a242e2d1e0176a99834e2d1175de12585fea52f2211d89320.jpg", "reference_inputs/images/38526620ed5a3af876796ff5433e7cce2e3ad95683e15a623a09116642efa690.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000095", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000095", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [818, 500, 823, 516]\n- T2: [791, 480, 797, 498]\n- T3: [570, 484, 581, 500]\n\nImage 2 (GROUND) labeled targets:\n- T1: [682, 298, 690, 324]\n- T2: [612, 291, 621, 325]\n- T3: [972, 330, 1023, 455]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T2 -> T3 -> T1\nC: T1 -> T3 -> T2\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1b772583e746028a242e2d1e0176a99834e2d1175de12585fea52f2211d89320.jpg", "reference_inputs/images/38526620ed5a3af876796ff5433e7cce2e3ad95683e15a623a09116642efa690.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000096", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000096", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [510, 428, 543, 439]\n- T2: [433, 463, 454, 489]\n- T3: [630, 424, 662, 434]\n\nImage 2 (GROUND) labeled targets:\n- T1: [534, 301, 576, 318]\n- T2: [250, 299, 346, 352]\n- T3: [774, 300, 816, 315]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T1 -> T3 -> T2\nC: T2 -> T3 -> T1\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/99ff36e6b012edd7fa9d4c7b1bc65614497e52512fc40b1cf1cc29158b33f267.jpg", "reference_inputs/images/cb0db9da20c0b9cc4c4332bedcaa62bc9fc99c8659420cd44020cb244ade6d10.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000097", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000097", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [242, 550, 271, 575]\n- T2: [575, 550, 590, 574]\n- T3: [618, 353, 634, 369]\n\nImage 2 (GROUND) labeled targets:\n- T1: [325, 287, 332, 310]\n- T2: [561, 294, 567, 320]\n- T3: [573, 286, 578, 312]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T1 -> T3 -> T2\nC: T1 -> T2 -> T3\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/40d0ce2fd4a56089ea8837e202fe35b54cc6d6a9d77031cfb4521ab135148a32.jpg", "reference_inputs/images/f00be5f5431d5f216408980ab71683bd8405197facf8cec2f1e73396ba223dd5.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000098", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000098", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [575, 550, 590, 574]\n- T2: [618, 353, 634, 369]\n- T3: [242, 550, 271, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [561, 294, 567, 320]\n- T2: [573, 286, 578, 312]\n- T3: [325, 287, 332, 310]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T1 -> T2 -> T3\nC: T2 -> T3 -> T1\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/40d0ce2fd4a56089ea8837e202fe35b54cc6d6a9d77031cfb4521ab135148a32.jpg", "reference_inputs/images/f00be5f5431d5f216408980ab71683bd8405197facf8cec2f1e73396ba223dd5.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000099", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000099", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [700, 43, 723, 70]\n- T2: [557, 519, 590, 540]\n- T3: [434, 283, 456, 294]\n\nImage 2 (GROUND) labeled targets:\n- T1: [263, 294, 271, 312]\n- T2: [353, 290, 363, 308]\n- T3: [423, 290, 428, 311]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T1 -> T2 -> T3\nC: T3 -> T1 -> T2\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ad6fe422531a93b9485098c6c3e4bc860f8aeb03907c9f838ade835d139050c2.jpg", "reference_inputs/images/d2fdc4683ecf48ad9dd5cf6d1739017f93cacc73846c598a25716dd0416a30a8.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000100", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000100", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [502, 459, 517, 475]\n- T2: [588, 374, 608, 379]\n- T3: [308, 530, 314, 553]\n- T4: [347, 505, 354, 525]\n\nImage 2 (GROUND) labeled targets:\n- T1: [650, 300, 690, 320]\n- T2: [670, 291, 696, 297]\n- T3: [297, 289, 316, 357]\n- T4: [394, 290, 410, 338]\n\nAnswer options:\nA: T3 -> T4 -> T2 -> T1\nB: T3 -> T4 -> T1 -> T2\nC: T3 -> T1 -> T4 -> T2\nD: T4 -> T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5bc11b7f2e38ac137619a4800f8fe2d0896a00695683ac0192ea35c976912819.jpg", "reference_inputs/images/ef1c9e15657d46d314b74cd595a8bf5a6aa2686d5fedec13ec8318fc99ba7f6c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000101", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000101", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [391, 533, 421, 569]\n- T2: [439, 441, 454, 455]\n- T3: [347, 505, 354, 525]\n- T4: [308, 530, 314, 553]\n\nImage 2 (GROUND) labeled targets:\n- T1: [551, 313, 633, 375]\n- T2: [526, 298, 550, 314]\n- T3: [394, 290, 410, 338]\n- T4: [297, 289, 316, 357]\n\nAnswer options:\nA: T1 -> T4 -> T3 -> T2\nB: T4 -> T1 -> T3 -> T2\nC: T1 -> T3 -> T4 -> T2\nD: T1 -> T4 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5bc11b7f2e38ac137619a4800f8fe2d0896a00695683ac0192ea35c976912819.jpg", "reference_inputs/images/ef1c9e15657d46d314b74cd595a8bf5a6aa2686d5fedec13ec8318fc99ba7f6c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000102", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000102", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [739, 0, 842, 76]\n- T2: [917, 504, 962, 535]\n- T3: [965, 14, 1003, 59]\n\nImage 2 (GROUND) labeled targets:\n- T1: [279, 321, 436, 402]\n- T2: [998, 294, 1018, 350]\n- T3: [420, 297, 444, 342]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T2 -> T3\nC: unable to determine\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1aedfd6862af453cdd58f1227e627eccd860b652ccd30b1767d50432e1f9439a.jpg", "reference_inputs/images/9d7dd89769e249a4eddd5467312146be0428ac4007260c0ea7d8f0137e5d7c62.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000103", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000103", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [739, 0, 842, 76]\n- T2: [731, 166, 840, 269]\n- T3: [917, 504, 962, 535]\n\nImage 2 (GROUND) labeled targets:\n- T1: [279, 321, 436, 402]\n- T2: [608, 310, 835, 409]\n- T3: [998, 294, 1018, 350]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: unable to determine\nC: T2 -> T3 -> T1\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1aedfd6862af453cdd58f1227e627eccd860b652ccd30b1767d50432e1f9439a.jpg", "reference_inputs/images/9d7dd89769e249a4eddd5467312146be0428ac4007260c0ea7d8f0137e5d7c62.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000104", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000104", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [639, 58, 666, 81]\n- T2: [498, 139, 546, 247]\n- T3: [618, 486, 635, 511]\n\nImage 2 (GROUND) labeled targets:\n- T1: [605, 287, 609, 316]\n- T2: [505, 299, 539, 327]\n- T3: [656, 294, 673, 352]\n\nAnswer options:\nA: unable to determine\nB: T3 -> T1 -> T2\nC: T3 -> T2 -> T1\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/927a6db3097148ac7af46cf66cc1d6625a7eb825690a7f4628d4b84905873f8a.jpg", "reference_inputs/images/839e7480594c685cc455a57a7f79881f7053bafcfd91d1204a3a1244770b77d3.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000105", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000105", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [491, 519, 512, 543]\n- T2: [526, 198, 536, 214]\n- T3: [554, 248, 568, 267]\n\nImage 2 (GROUND) labeled targets:\n- T1: [438, 317, 502, 515]\n- T2: [529, 306, 542, 361]\n- T3: [570, 290, 587, 348]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T2 -> T1\nC: T3 -> T1 -> T2\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/82de66c33aa7a0590d806c77253770c7eab69ef7bfc0111dd4211431220f31f4.jpg", "reference_inputs/images/58b047aeb9b58dd0ab84674467ce22bda3cef16ba28ca4600778e7ba9afda82a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000106", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000106", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [633, 458, 649, 475]\n- T2: [918, 0, 951, 17]\n- T3: [958, 455, 1000, 473]\n\nImage 2 (GROUND) labeled targets:\n- T1: [51, 295, 75, 346]\n- T2: [203, 287, 213, 312]\n- T3: [385, 293, 403, 322]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T3 -> T2 -> T1\nC: T1 -> T3 -> T2\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fce96307898600f8aa5eac561a501e327b9732a3917600d3ba991883de1c00b7.jpg", "reference_inputs/images/23d606339b44ef900e4f4b8fd734f19f809273819df9d7e7aac8b402b5896432.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000107", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000107", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [585, 400, 596, 410]\n- T2: [594, 421, 611, 433]\n- T3: [694, 473, 724, 490]\n\nImage 2 (GROUND) labeled targets:\n- T1: [600, 290, 615, 300]\n- T2: [615, 294, 638, 303]\n- T3: [765, 299, 817, 314]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T2 -> T1\nC: unable to determine\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/09e4d63e55bd5f22dd2cec7666efaac4ee4d83856b562dc1c9726bfdc0600797.jpg", "reference_inputs/images/8c822065e6a4a3137d25a6abc18bb3c2c267e21f7d088cad1a3a17dbee7b5c2a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000108", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000108", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [187, 358, 218, 373]\n- T2: [196, 450, 314, 539]\n- T3: [620, 218, 637, 234]\n- T4: [769, 289, 875, 368]\n\nImage 2 (GROUND) labeled targets:\n- T1: [101, 294, 118, 346]\n- T2: [213, 302, 360, 349]\n- T3: [305, 287, 310, 304]\n- T4: [400, 294, 449, 311]\n\nAnswer options:\nA: T2 -> T1 -> T4 -> T3\nB: T1 -> T2 -> T3 -> T4\nC: T2 -> T1 -> T3 -> T4\nD: T2 -> T3 -> T1 -> T4\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0671ae8ee169440e6627d92e84b2ba53644dd471a3dc9097fb5bbae57d101527.jpg", "reference_inputs/images/cd09321292f0788f78323c26abbf46a241c86ce743f184ba8aece6a49f533751.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000109", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000109", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [196, 450, 314, 539]\n- T2: [620, 218, 637, 234]\n- T3: [100, 486, 213, 564]\n\nImage 2 (GROUND) labeled targets:\n- T1: [213, 302, 360, 349]\n- T2: [305, 287, 310, 304]\n- T3: [116, 323, 309, 383]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T3 -> T2 -> T1\nC: T1 -> T2 -> T3\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0671ae8ee169440e6627d92e84b2ba53644dd471a3dc9097fb5bbae57d101527.jpg", "reference_inputs/images/cd09321292f0788f78323c26abbf46a241c86ce743f184ba8aece6a49f533751.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000110", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000110", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [255, 545, 286, 575]\n- T2: [156, 530, 162, 552]\n- T3: [335, 484, 358, 506]\n\nImage 2 (GROUND) labeled targets:\n- T1: [408, 292, 447, 318]\n- T2: [362, 288, 368, 307]\n- T3: [532, 290, 562, 309]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T2 -> T3\nC: T1 -> T3 -> T2\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1b51b7b6b84ff1a9d8e198eaf5e87d3a7a4fa6656b294051d6264f679cc47f33.jpg", "reference_inputs/images/d31f007add278cf876ac86158736bcb1f97f1402eaef053294693069319ac645.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000111", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000111", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [459, 446, 478, 464]\n- T2: [214, 83, 244, 110]\n- T3: [302, 187, 330, 201]\n\nImage 2 (GROUND) labeled targets:\n- T1: [631, 292, 639, 330]\n- T2: [562, 290, 567, 311]\n- T3: [583, 291, 589, 315]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T2 -> T1 -> T3\nC: T1 -> T2 -> T3\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a4968ce1b55bf0be69a034655155c8cadb8cd1afd0efe8ff7198b89dd500688a.jpg", "reference_inputs/images/3a6c37ec5821918390e2d7fe5f5eced40254021cdba472145ae816c701d4959b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000112", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000112", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [930, 247, 1023, 303]\n- T2: [52, 546, 100, 575]\n- T3: [114, 122, 222, 178]\n\nImage 2 (GROUND) labeled targets:\n- T1: [474, 294, 496, 301]\n- T2: [580, 289, 586, 309]\n- T3: [394, 300, 422, 311]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T2 -> T3 -> T1\nC: T3 -> T1 -> T2\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b31cbbae1559d1aa0a7a38ebb3cbe9cf1fc7511cd0a6c9b152204e9b2d52ad7d.jpg", "reference_inputs/images/23d32f7a6776881bab4902b18c060313973797d4d8e2e4d0cb1a413c41e67a4c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000113", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000113", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [573, 410, 586, 422]\n- T2: [634, 547, 667, 575]\n- T3: [594, 462, 615, 481]\n- T4: [534, 506, 556, 534]\n\nImage 2 (GROUND) labeled targets:\n- T1: [578, 289, 594, 300]\n- T2: [661, 301, 705, 322]\n- T3: [605, 290, 631, 307]\n- T4: [519, 290, 564, 314]\n\nAnswer options:\nA: T4 -> T2 -> T3 -> T1\nB: T2 -> T4 -> T3 -> T1\nC: T2 -> T3 -> T4 -> T1\nD: T2 -> T4 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7ff9a063ea7e900809d71200db9a76b02c6c06d8d46e1304fb4b04a2199158b9.jpg", "reference_inputs/images/ffeabae965ca87a9a76fa6420a55de21d573a30f19c479c2e2f9f7fc91d2e546.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000114", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000114", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [534, 506, 556, 534]\n- T2: [634, 547, 667, 575]\n- T3: [573, 410, 586, 422]\n\nImage 2 (GROUND) labeled targets:\n- T1: [519, 290, 564, 314]\n- T2: [661, 301, 705, 322]\n- T3: [578, 289, 594, 300]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T3 -> T2 -> T1\nC: T1 -> T2 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7ff9a063ea7e900809d71200db9a76b02c6c06d8d46e1304fb4b04a2199158b9.jpg", "reference_inputs/images/ffeabae965ca87a9a76fa6420a55de21d573a30f19c479c2e2f9f7fc91d2e546.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000115", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000115", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [543, 466, 594, 575]\n- T2: [366, 307, 383, 328]\n- T3: [725, 182, 780, 288]\n\nImage 2 (GROUND) labeled targets:\n- T1: [485, 294, 521, 325]\n- T2: [383, 290, 390, 312]\n- T3: [568, 294, 598, 312]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T3 -> T1 -> T2\nC: T1 -> T3 -> T2\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3d91c45a797db089579e59426df5d2694d63f16a52b3e2a02525e4ff3e3856ec.jpg", "reference_inputs/images/419abca19db99a735386f5f2dcc0ec76522897e379bf2f4f77a39ef13eb421b8.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000116", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000116", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [951, 452, 1002, 478]\n- T2: [834, 462, 875, 477]\n- T3: [850, 431, 886, 443]\n- T4: [968, 498, 1023, 522]\n\nImage 2 (GROUND) labeled targets:\n- T1: [523, 293, 558, 304]\n- T2: [402, 296, 449, 310]\n- T3: [447, 290, 475, 303]\n- T4: [539, 295, 574, 313]\n\nAnswer options:\nA: T4 -> T2 -> T3 -> T1\nB: T4 -> T2 -> T1 -> T3\nC: T2 -> T4 -> T1 -> T3\nD: T4 -> T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/73b0591df8bd42450b092703a7faf5032294f338d53e2fc9e85ddb18f4a0ef0f.jpg", "reference_inputs/images/af00de34a0a8156b25fa9cd1398e7cfd58c1674afe33b86509ac463390783e8b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000117", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000117", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [726, 518, 733, 540]\n- T2: [968, 498, 1023, 522]\n- T3: [834, 462, 875, 477]\n- T4: [951, 452, 1002, 478]\n\nImage 2 (GROUND) labeled targets:\n- T1: [127, 289, 146, 339]\n- T2: [539, 295, 574, 313]\n- T3: [402, 296, 449, 310]\n- T4: [523, 293, 558, 304]\n\nAnswer options:\nA: T1 -> T2 -> T3 -> T4\nB: T1 -> T2 -> T4 -> T3\nC: T2 -> T1 -> T3 -> T4\nD: T1 -> T3 -> T2 -> T4\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/73b0591df8bd42450b092703a7faf5032294f338d53e2fc9e85ddb18f4a0ef0f.jpg", "reference_inputs/images/af00de34a0a8156b25fa9cd1398e7cfd58c1674afe33b86509ac463390783e8b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000118", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000118", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [517, 559, 522, 575]\n- T2: [900, 558, 974, 575]\n- T3: [966, 506, 974, 525]\n\nImage 2 (GROUND) labeled targets:\n- T1: [442, 294, 446, 312]\n- T2: [258, 314, 363, 368]\n- T3: [693, 310, 760, 469]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T3 -> T1 -> T2\nC: T2 -> T3 -> T1\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e4aac62165cb28afcbd439a83cee130a5bc480fe9cb15115ea14cabc005d8d78.jpg", "reference_inputs/images/52f9f64e14c06f5f784d017fd8cb72625c111c2ded139cc3bbe2085137c9c586.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000119", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000119", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [517, 559, 522, 575]\n- T2: [966, 506, 974, 525]\n- T3: [900, 558, 974, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [442, 294, 446, 312]\n- T2: [693, 310, 760, 469]\n- T3: [258, 314, 363, 368]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T1 -> T3 -> T2\nC: T1 -> T2 -> T3\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e4aac62165cb28afcbd439a83cee130a5bc480fe9cb15115ea14cabc005d8d78.jpg", "reference_inputs/images/52f9f64e14c06f5f784d017fd8cb72625c111c2ded139cc3bbe2085137c9c586.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000120", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000120", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [649, 333, 667, 346]\n- T2: [695, 122, 721, 140]\n- T3: [658, 562, 674, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [441, 293, 458, 350]\n- T2: [217, 290, 241, 342]\n- T3: [733, 299, 751, 351]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T1 -> T2 -> T3\nC: T2 -> T1 -> T3\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b93ab76ddb5a25b217c5c906a07d5efb3a35760afac1f9e5f0eaea7e0b6e0003.jpg", "reference_inputs/images/09c75aba58c003ce40bb681fe7b9397e8cc38c0610726570eef95c39e242f617.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000121", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000121", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [695, 122, 721, 140]\n- T2: [653, 66, 677, 90]\n- T3: [658, 562, 674, 575]\n- T4: [138, 486, 174, 510]\n\nImage 2 (GROUND) labeled targets:\n- T1: [608, 287, 618, 310]\n- T2: [583, 286, 587, 309]\n- T3: [649, 292, 667, 330]\n- T4: [222, 288, 231, 325]\n\nAnswer options:\nA: T4 -> T3 -> T1 -> T2\nB: T3 -> T4 -> T1 -> T2\nC: T3 -> T4 -> T2 -> T1\nD: T3 -> T1 -> T4 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b93ab76ddb5a25b217c5c906a07d5efb3a35760afac1f9e5f0eaea7e0b6e0003.jpg", "reference_inputs/images/30142c91b4a377fd8bb01708b286c070258fd5c4a84e7771a9bec6c4b1def4c8.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000122", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000122", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [658, 562, 674, 575]\n- T2: [649, 333, 667, 346]\n- T3: [138, 486, 174, 510]\n\nImage 2 (GROUND) labeled targets:\n- T1: [649, 292, 667, 330]\n- T2: [602, 288, 608, 317]\n- T3: [222, 288, 231, 325]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T2 -> T1 -> T3\nC: T2 -> T3 -> T1\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b93ab76ddb5a25b217c5c906a07d5efb3a35760afac1f9e5f0eaea7e0b6e0003.jpg", "reference_inputs/images/30142c91b4a377fd8bb01708b286c070258fd5c4a84e7771a9bec6c4b1def4c8.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000123", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000123", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [959, 277, 998, 302]\n- T2: [382, 374, 482, 476]\n- T3: [722, 19, 830, 126]\n\nImage 2 (GROUND) labeled targets:\n- T1: [838, 290, 854, 321]\n- T2: [388, 296, 488, 336]\n- T3: [631, 291, 680, 316]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T1 -> T3\nC: unable to determine\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/593aba1a5f50088dcad24abfed915ae977e9337ecee2638c883220b5be6434b8.jpg", "reference_inputs/images/c78a0378bceef9047b848116311ccaa5e549a203c8ea138f175b0306b5e4c7b7.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000124", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000124", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [382, 374, 482, 476]\n- T2: [730, 417, 839, 520]\n- T3: [959, 277, 998, 302]\n\nImage 2 (GROUND) labeled targets:\n- T1: [388, 296, 488, 336]\n- T2: [735, 296, 813, 342]\n- T3: [838, 290, 854, 321]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T3 -> T1 -> T2\nC: T2 -> T1 -> T3\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/593aba1a5f50088dcad24abfed915ae977e9337ecee2638c883220b5be6434b8.jpg", "reference_inputs/images/c78a0378bceef9047b848116311ccaa5e549a203c8ea138f175b0306b5e4c7b7.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000127", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000127", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [632, 475, 654, 494]\n- T2: [730, 490, 767, 510]\n- T3: [644, 430, 676, 442]\n- T4: [763, 553, 815, 575]\n\nImage 2 (GROUND) labeled targets:\n- T1: [869, 307, 954, 324]\n- T2: [833, 302, 889, 315]\n- T3: [610, 306, 662, 320]\n- T4: [979, 300, 1023, 316]\n\nAnswer options:\nA: T1 -> T3 -> T2 -> T4\nB: T3 -> T1 -> T2 -> T4\nC: T3 -> T1 -> T4 -> T2\nD: T3 -> T2 -> T1 -> T4\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d4e07913a5a14259f4b28417bcd90f7478a358c22747ce6cf3e6cfa47b710b39.jpg", "reference_inputs/images/6aeff7f4a3fe7015e7da04d5c9d75664c3247c67da7bc0db28bb73528a27a828.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000128", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000128", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [730, 490, 767, 510]\n- T2: [730, 502, 775, 534]\n- T3: [763, 553, 815, 575]\n- T4: [644, 430, 676, 442]\n\nImage 2 (GROUND) labeled targets:\n- T1: [833, 302, 889, 315]\n- T2: [886, 298, 953, 312]\n- T3: [979, 300, 1023, 316]\n- T4: [610, 306, 662, 320]\n\nAnswer options:\nA: T3 -> T1 -> T2 -> T4\nB: T3 -> T2 -> T1 -> T4\nC: T3 -> T2 -> T4 -> T1\nD: T2 -> T3 -> T1 -> T4\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d4e07913a5a14259f4b28417bcd90f7478a358c22747ce6cf3e6cfa47b710b39.jpg", "reference_inputs/images/6aeff7f4a3fe7015e7da04d5c9d75664c3247c67da7bc0db28bb73528a27a828.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000129", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000129", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [870, 362, 906, 374]\n- T2: [561, 158, 581, 173]\n- T3: [898, 470, 935, 488]\n\nImage 2 (GROUND) labeled targets:\n- T1: [862, 289, 875, 331]\n- T2: [543, 290, 564, 318]\n- T3: [990, 294, 1009, 345]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: unable to determine\nC: T3 -> T2 -> T1\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4f1d592425f43ddd44ef33551b00627e3bfa2b20112e736d3202b2b1e08e70c9.jpg", "reference_inputs/images/5aa2727bedc35e1a6443c7568f420d63beeb311a931ccbf27122807cd271e2ca.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000130", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000130", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [561, 158, 581, 173]\n- T2: [898, 470, 935, 488]\n- T3: [870, 362, 906, 374]\n\nImage 2 (GROUND) labeled targets:\n- T1: [543, 290, 564, 318]\n- T2: [990, 294, 1009, 345]\n- T3: [862, 289, 875, 331]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: unable to determine\nC: T1 -> T3 -> T2\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4f1d592425f43ddd44ef33551b00627e3bfa2b20112e736d3202b2b1e08e70c9.jpg", "reference_inputs/images/5aa2727bedc35e1a6443c7568f420d63beeb311a931ccbf27122807cd271e2ca.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000131", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000131", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [302, 187, 330, 201]\n- T2: [214, 83, 244, 110]\n- T3: [459, 446, 478, 464]\n\nImage 2 (GROUND) labeled targets:\n- T1: [768, 298, 798, 357]\n- T2: [809, 291, 825, 341]\n- T3: [394, 310, 437, 457]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T1 -> T2\nC: T3 -> T2 -> T1\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a4968ce1b55bf0be69a034655155c8cadb8cd1afd0efe8ff7198b89dd500688a.jpg", "reference_inputs/images/725dd9e99787074d704eb4359760886aa6d4349bd16a59719544d615d334bba5.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000132", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000132", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [820, 307, 879, 415]\n- T2: [551, 437, 562, 453]\n- T3: [569, 170, 585, 187]\n\nImage 2 (GROUND) labeled targets:\n- T1: [866, 300, 1022, 358]\n- T2: [573, 302, 587, 362]\n- T3: [562, 286, 574, 327]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T1 -> T3\nC: unable to determine\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d0048a3c1c294547ebe61642ac6915062a9d96f63aae94ff790904dbd788c32e.jpg", "reference_inputs/images/8f4940264acc56f4ce551a90fee909e68ecf5da8ccb78f8c6385b3774c7c049d.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000133", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000133", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [749, 367, 805, 476]\n- T2: [879, 118, 920, 139]\n- T3: [236, 168, 338, 214]\n\nImage 2 (GROUND) labeled targets:\n- T1: [501, 302, 543, 338]\n- T2: [585, 288, 590, 314]\n- T3: [159, 300, 236, 320]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T1 -> T3\nC: T1 -> T3 -> T2\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/644ca97d9b48e632d07cc5fb56a4168f0cd1c94ce6d868d31853df66563b8285.jpg", "reference_inputs/images/0d1249b0b489795184fdf483552ed60b7f003460f9096becd497208470f58342.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000134", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000134", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [562, 420, 584, 434]\n- T2: [574, 510, 589, 539]\n- T3: [574, 19, 589, 49]\n\nImage 2 (GROUND) labeled targets:\n- T1: [551, 292, 578, 328]\n- T2: [576, 289, 590, 337]\n- T3: [545, 287, 550, 310]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T1 -> T2 -> T3\nC: T2 -> T3 -> T1\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/98588555a9656d7cfc9f63266d2fcf2b6eb3064482278c14ddac2c450e16d2c2.jpg", "reference_inputs/images/b6533713d8ccd6c7ed91c22c2decd479e9eb3d09d24937b933b92a291f6e5430.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000136", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000136", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [519, 423, 534, 436]\n- T2: [619, 508, 647, 537]\n- T3: [671, 484, 680, 502]\n\nImage 2 (GROUND) labeled targets:\n- T1: [83, 294, 110, 303]\n- T2: [169, 295, 197, 313]\n- T3: [250, 287, 254, 306]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T2 -> T3 -> T1\nC: T1 -> T2 -> T3\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2af14fbe1a42008fa850bbe6da4cb1076d10c39e20f9ac899fb560071177a9cf.jpg", "reference_inputs/images/8d920b98306faae1dd570d511d985769d0571056126f4c467bc883416cf04161.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000137", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000137", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [716, 217, 742, 229]\n- T2: [0, 147, 34, 162]\n- T3: [151, 440, 192, 461]\n\nImage 2 (GROUND) labeled targets:\n- T1: [453, 287, 460, 310]\n- T2: [1, 290, 17, 315]\n- T3: [12, 289, 39, 327]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T3 -> T1 -> T2\nC: T1 -> T3 -> T2\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/62b21a2b843cd73f959a114c90ba7721502bba41770522608b3cf2779d5083c2.jpg", "reference_inputs/images/ba1dd72c489a349f7b0ce72aff1a61fec69dba3647c2dca0d283dd44fc578db2.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000138", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000138", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [716, 217, 742, 229]\n- T2: [0, 147, 34, 162]\n- T3: [151, 440, 192, 461]\n\nImage 2 (GROUND) labeled targets:\n- T1: [453, 287, 460, 310]\n- T2: [1, 290, 17, 315]\n- T3: [12, 289, 39, 327]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: unable to determine\nC: T1 -> T2 -> T3\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/62b21a2b843cd73f959a114c90ba7721502bba41770522608b3cf2779d5083c2.jpg", "reference_inputs/images/ba1dd72c489a349f7b0ce72aff1a61fec69dba3647c2dca0d283dd44fc578db2.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000139", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000139", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [744, 453, 771, 479]\n- T2: [132, 455, 166, 475]\n- T3: [561, 126, 626, 238]\n- T4: [295, 390, 356, 500]\n\nImage 2 (GROUND) labeled targets:\n- T1: [736, 288, 742, 306]\n- T2: [493, 291, 498, 310]\n- T3: [620, 297, 642, 306]\n- T4: [558, 296, 579, 311]\n\nAnswer options:\nA: T2 -> T1 -> T4 -> T3\nB: T4 -> T2 -> T1 -> T3\nC: T2 -> T4 -> T1 -> T3\nD: T2 -> T4 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5c0e4b7fd28d747f1c4c10d28e50b31f60ce5a6fbd644a44f8b265ae00c19b94.jpg", "reference_inputs/images/9a8ae0b3d983add7ddedd2d91a1213471f9aed9b1ff95a593c93278990174473.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000140", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000140", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [132, 455, 166, 475]\n- T2: [295, 390, 356, 500]\n- T3: [561, 126, 626, 238]\n\nImage 2 (GROUND) labeled targets:\n- T1: [493, 291, 498, 310]\n- T2: [558, 296, 579, 311]\n- T3: [620, 297, 642, 306]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T2 -> T3 -> T1\nC: T3 -> T1 -> T2\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5c0e4b7fd28d747f1c4c10d28e50b31f60ce5a6fbd644a44f8b265ae00c19b94.jpg", "reference_inputs/images/9a8ae0b3d983add7ddedd2d91a1213471f9aed9b1ff95a593c93278990174473.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000141", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000141", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [551, 180, 563, 193]\n- T2: [654, 370, 669, 383]\n- T3: [821, 51, 881, 162]\n\nImage 2 (GROUND) labeled targets:\n- T1: [535, 292, 541, 314]\n- T2: [624, 294, 631, 323]\n- T3: [681, 296, 702, 315]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T1 -> T3 -> T2\nC: T2 -> T1 -> T3\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c2ec65c1192f0977926ea22641ec5ce0c2381a4a2709c4e60b82c90ba52c3b99.jpg", "reference_inputs/images/71c1d97e3384db44f94c215ca7826cfdc15574bf16d173d6c6f24d4872a16d89.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000142", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000142", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [905, 522, 969, 546]\n- T2: [908, 495, 960, 514]\n- T3: [988, 471, 1017, 489]\n\nImage 2 (GROUND) labeled targets:\n- T1: [285, 312, 466, 382]\n- T2: [386, 309, 466, 334]\n- T3: [514, 296, 573, 319]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T3 -> T1\nC: T1 -> T3 -> T2\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2a6d422c5d7113659c40116522d55551f55e3fce579ed0dcd074766c3c123762.jpg", "reference_inputs/images/5275ce1621856d5c511eee6aa53f4b0032a9973d2584e679a486449739a29142.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000143", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000143", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [905, 522, 969, 546]\n- T2: [988, 471, 1017, 489]\n- T3: [908, 495, 960, 514]\n\nImage 2 (GROUND) labeled targets:\n- T1: [285, 312, 466, 382]\n- T2: [514, 296, 573, 319]\n- T3: [386, 309, 466, 334]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T1 -> T2 -> T3\nC: T2 -> T1 -> T3\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2a6d422c5d7113659c40116522d55551f55e3fce579ed0dcd074766c3c123762.jpg", "reference_inputs/images/5275ce1621856d5c511eee6aa53f4b0032a9973d2584e679a486449739a29142.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000144", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000144", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [621, 401, 634, 411]\n- T2: [569, 480, 579, 496]\n- T3: [450, 489, 454, 506]\n- T4: [594, 414, 607, 426]\n\nImage 2 (GROUND) labeled targets:\n- T1: [654, 291, 675, 298]\n- T2: [599, 289, 629, 317]\n- T3: [404, 290, 412, 321]\n- T4: [622, 292, 645, 304]\n\nAnswer options:\nA: T3 -> T4 -> T2 -> T1\nB: T2 -> T3 -> T4 -> T1\nC: T3 -> T2 -> T1 -> T4\nD: T3 -> T2 -> T4 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/39f792258ee5da7ad110e4a0028f5a09df7eaabf2742743885a78cb2856015d7.jpg", "reference_inputs/images/702dd7ba3d6815564455245f3199e57483d587890f24e75a7c679226bfe87823.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000145", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000145", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [517, 375, 537, 388]\n- T2: [578, 470, 590, 494]\n- T3: [928, 494, 1023, 552]\n- T4: [589, 146, 602, 166]\n\nImage 2 (GROUND) labeled targets:\n- T1: [518, 294, 534, 328]\n- T2: [576, 294, 585, 336]\n- T3: [939, 312, 1023, 344]\n- T4: [558, 286, 565, 314]\n\nAnswer options:\nA: T2 -> T1 -> T3 -> T4\nB: T1 -> T2 -> T3 -> T4\nC: T2 -> T3 -> T1 -> T4\nD: T2 -> T1 -> T4 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/40f61ea9edb503112601c62b3f52b58bfbfc3755589baafaa08d713e3b776977.jpg", "reference_inputs/images/12ff10e2b317b9f7c0aff4ccda2d4142e0e5b16c61dab7dfee15d3580d0725a9.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000146", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000146", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [115, 372, 223, 429]\n- T2: [589, 146, 602, 166]\n- T3: [928, 494, 1023, 552]\n\nImage 2 (GROUND) labeled targets:\n- T1: [173, 307, 274, 331]\n- T2: [558, 286, 565, 314]\n- T3: [939, 312, 1023, 344]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T3 -> T1\nC: T2 -> T1 -> T3\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/40f61ea9edb503112601c62b3f52b58bfbfc3755589baafaa08d713e3b776977.jpg", "reference_inputs/images/12ff10e2b317b9f7c0aff4ccda2d4142e0e5b16c61dab7dfee15d3580d0725a9.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000147", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000147", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [699, 314, 722, 329]\n- T2: [278, 227, 288, 238]\n- T3: [600, 538, 621, 565]\n\nImage 2 (GROUND) labeled targets:\n- T1: [275, 294, 283, 319]\n- T2: [536, 293, 544, 321]\n- T3: [374, 290, 380, 312]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1 -> T3\nC: T2 -> T3 -> T1\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7b9a15fc20b7d171680b09bbfcd007ee5591003dc770b19a61842feef5fce354.jpg", "reference_inputs/images/3521a15d80c0751979b2e5fc8e26a0832da82eaf6d059b8b8657531196675665.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000148", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000148", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [738, 475, 843, 575]\n- T2: [737, 174, 834, 266]\n- T3: [958, 531, 998, 558]\n\nImage 2 (GROUND) labeled targets:\n- T1: [918, 337, 1023, 386]\n- T2: [403, 330, 572, 402]\n- T3: [842, 299, 874, 344]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T2 -> T3 -> T1\nC: T3 -> T2 -> T1\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7d1435db3185a85d834e261e7135d080ba24244be700995c1b8b272af1cd25ac.jpg", "reference_inputs/images/e166c23d8949500a17c9f6e160875b5ed05aecf935baae0f424f04e755e82240.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000149", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000149", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [738, 475, 843, 575]\n- T2: [958, 531, 998, 558]\n- T3: [737, 174, 834, 266]\n\nImage 2 (GROUND) labeled targets:\n- T1: [918, 337, 1023, 386]\n- T2: [842, 299, 874, 344]\n- T3: [403, 330, 572, 402]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T3 -> T2\nC: T3 -> T2 -> T1\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7d1435db3185a85d834e261e7135d080ba24244be700995c1b8b272af1cd25ac.jpg", "reference_inputs/images/e166c23d8949500a17c9f6e160875b5ed05aecf935baae0f424f04e755e82240.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000151", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000151", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [286, 187, 342, 296]\n- T2: [101, 386, 136, 409]\n- T3: [505, 185, 550, 283]\n- T4: [20, 438, 64, 456]\n\nImage 2 (GROUND) labeled targets:\n- T1: [523, 297, 548, 309]\n- T2: [439, 289, 445, 310]\n- T3: [618, 295, 643, 309]\n- T4: [394, 287, 402, 311]\n\nAnswer options:\nA: T3 -> T1 -> T2 -> T4\nB: T3 -> T2 -> T1 -> T4\nC: T3 -> T1 -> T4 -> T2\nD: T1 -> T3 -> T2 -> T4\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/13e458bd345b9533939925519d14eefa64247eac736299c9c7b2cfa86943f1c2.jpg", "reference_inputs/images/778acd66ce6b79ac291b2d9c2a18217f97c9cf2ed7c14781bd1bcf9246f8b465.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000152", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000152", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [870, 362, 906, 374]\n- T2: [898, 470, 935, 488]\n- T3: [575, 263, 588, 273]\n\nImage 2 (GROUND) labeled targets:\n- T1: [663, 294, 674, 315]\n- T2: [786, 298, 806, 352]\n- T3: [565, 314, 654, 575]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T3 -> T1 -> T2\nC: T3 -> T2 -> T1\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4f1d592425f43ddd44ef33551b00627e3bfa2b20112e736d3202b2b1e08e70c9.jpg", "reference_inputs/images/a04ee1aa01447531e8a5ff396360422102cee8a5d040a61a205db0f05731279d.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000153", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000153", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [891, 556, 900, 575]\n- T2: [635, 475, 661, 498]\n- T3: [719, 499, 754, 526]\n\nImage 2 (GROUND) labeled targets:\n- T1: [809, 289, 822, 306]\n- T2: [617, 291, 636, 302]\n- T3: [685, 292, 711, 304]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T2 -> T3\nC: T1 -> T3 -> T2\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/514d7c7966819c35c4c7bde0dc82fe9428118e4707a02c31cf6b36f610b2ad77.jpg", "reference_inputs/images/7e07668d7a6b94a053a99a2a815b4722d357061700a25e036f0f4e03b45484e0.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000154", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000154", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [719, 499, 754, 526]\n- T2: [891, 556, 900, 575]\n- T3: [635, 475, 661, 498]\n\nImage 2 (GROUND) labeled targets:\n- T1: [685, 292, 711, 304]\n- T2: [809, 289, 822, 306]\n- T3: [617, 291, 636, 302]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T1 -> T3 -> T2\nC: T1 -> T2 -> T3\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/514d7c7966819c35c4c7bde0dc82fe9428118e4707a02c31cf6b36f610b2ad77.jpg", "reference_inputs/images/7e07668d7a6b94a053a99a2a815b4722d357061700a25e036f0f4e03b45484e0.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000155", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000155", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [517, 375, 537, 388]\n- T2: [578, 470, 590, 494]\n- T3: [589, 146, 602, 166]\n\nImage 2 (GROUND) labeled targets:\n- T1: [522, 308, 562, 392]\n- T2: [730, 314, 774, 468]\n- T3: [594, 288, 607, 338]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T3 -> T1\nC: T2 -> T1 -> T3\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/40f61ea9edb503112601c62b3f52b58bfbfc3755589baafaa08d713e3b776977.jpg", "reference_inputs/images/c1af4763b62bb1c4e18c6f2872b14ca427b73e8e71ef4b9b97e5add7b49f9e43.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000156", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000156", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [151, 436, 183, 457]\n- T2: [314, 361, 370, 470]\n- T3: [155, 302, 186, 314]\n\nImage 2 (GROUND) labeled targets:\n- T1: [518, 296, 537, 346]\n- T2: [675, 316, 822, 384]\n- T3: [619, 293, 636, 335]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T2 -> T1 -> T3\nC: unable to determine\nD: T1 -> T2 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c97a9f8f7de6a094a1f743d885b8fb4e23676ff22a3149bfbf2818ff7fc8494b.jpg", "reference_inputs/images/607bbcd38675e22b718f62ece52a9b3e0e1298b2e309ebcd14d1fcd3700efcec.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000159", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000159", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [137, 164, 171, 178]\n- T2: [90, 282, 130, 296]\n- T3: [117, 471, 154, 497]\n\nImage 2 (GROUND) labeled targets:\n- T1: [250, 288, 261, 320]\n- T2: [162, 288, 174, 327]\n- T3: [16, 290, 43, 349]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T2 -> T1 -> T3\nC: T3 -> T1 -> T2\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1a01ed403e1d284e43fa05e564f7484c77989d75c0705a6d46e99f6f062bb7ce.jpg", "reference_inputs/images/570950fea3c5b9a9401fd164a88933b77a227eb1e8d7870537c987807cdfe0c0.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000160", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000160", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [359, 558, 366, 575]\n- T2: [52, 534, 87, 575]\n- T3: [147, 478, 172, 498]\n\nImage 2 (GROUND) labeled targets:\n- T1: [616, 294, 625, 328]\n- T2: [343, 292, 378, 314]\n- T3: [454, 290, 481, 306]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T1 -> T2 -> T3\nC: T1 -> T3 -> T2\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/dd01ae03a3c96e543bbcc23433cb23a95262a873ac133ed71a8b9ef474ac7a8a.jpg", "reference_inputs/images/d0e875f4efbc3ef538530c67ca5ab47489c9be3c8a833dd216b294b767f2e0f6.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000161", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000161", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [538, 361, 582, 459]\n- T2: [401, 382, 414, 395]\n- T3: [360, 305, 382, 325]\n- T4: [887, 546, 935, 574]\n\nImage 2 (GROUND) labeled targets:\n- T1: [289, 306, 354, 337]\n- T2: [192, 288, 206, 327]\n- T3: [202, 288, 210, 323]\n- T4: [673, 290, 703, 345]\n\nAnswer options:\nA: T1 -> T4 -> T2 -> T3\nB: T4 -> T2 -> T1 -> T3\nC: T4 -> T1 -> T2 -> T3\nD: T4 -> T1 -> T3 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/71627e38503f8fa008a5c48ecbf7cb63719a3f00c0377de6c1e90fbd4fd13671.jpg", "reference_inputs/images/d3558a936a65613cbbec820d11b51cc2ac958af87155f8146abf0217b9691467.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000162", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000162", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [146, 320, 175, 332]\n- T2: [132, 455, 166, 475]\n- T3: [295, 390, 356, 500]\n\nImage 2 (GROUND) labeled targets:\n- T1: [610, 291, 625, 326]\n- T2: [543, 294, 558, 335]\n- T3: [686, 308, 782, 356]\n\nAnswer options:\nA: T1 -> T3 -> T2\nB: T2 -> T3 -> T1\nC: T3 -> T2 -> T1\nD: T3 -> T1 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5c0e4b7fd28d747f1c4c10d28e50b31f60ce5a6fbd644a44f8b265ae00c19b94.jpg", "reference_inputs/images/26d845ab04d097c45765f43e5aa8a16f916542a3ee6a6a0303c397d41e78000c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000163", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000163", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [547, 94, 646, 143]\n- T2: [752, 436, 774, 454]\n- T3: [173, 236, 202, 253]\n\nImage 2 (GROUND) labeled targets:\n- T1: [531, 297, 596, 325]\n- T2: [880, 295, 903, 355]\n- T3: [198, 292, 210, 331]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T3 -> T2 -> T1\nC: T2 -> T1 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ff808de7f4377d64ed2599cf41e244006e312f326c271da6d10d4ea92c0320fb.jpg", "reference_inputs/images/570aa2e1dcd552d6dd226263519545311ebc7a1e09b440e10a71bb754cf35d99.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000164", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000164", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Ground Agent, observing the scene from the ground view.\nYour aerial teammate provides an aerial view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3, T4.\nDetermine the order of the labeled targets from nearest to farthest relative to your own ground-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nChoose the best near-to-far ranking among the listed options.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [70, 303, 106, 315]\n- T2: [614, 148, 709, 258]\n- T3: [270, 298, 363, 407]\n- T4: [969, 90, 1016, 114]\n\nImage 2 (GROUND) labeled targets:\n- T1: [490, 296, 506, 333]\n- T2: [752, 302, 795, 318]\n- T3: [655, 310, 705, 334]\n- T4: [830, 287, 837, 307]\n\nAnswer options:\nA: T3 -> T1 -> T2 -> T4\nB: T1 -> T3 -> T2 -> T4\nC: T1 -> T2 -> T3 -> T4\nD: T1 -> T3 -> T4 -> T2\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2f5b9c5803ca4db366c311011221b5a28998dfc282cf3612aa0248beaeaaa7a9.jpg", "reference_inputs/images/a54b9a361aecb04345c29fc9d533f9763dfc3b55579870f0ad6398d05d7a8e84.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "distance_comparison/town1_ground_aerial_v2/000165", "task_name": "distance_comparison", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/distance_comparison/town1_ground_aerial_v2/all_questions.json", "question_id": "000165", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Distance Comparison.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThe same shared targets are marked with the same labels in both views: T1, T2, T3.\nDetermine the order of the labeled targets from nearest to farthest relative to your own aerial-view position.\nUse both views jointly, together with perspective, scene layout, object scale, and context.\nIf the relative ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n\nImage 1 (AERIAL) labeled targets:\n- T1: [614, 148, 709, 258]\n- T2: [70, 303, 106, 315]\n- T3: [969, 90, 1016, 114]\n\nImage 2 (GROUND) labeled targets:\n- T1: [752, 302, 795, 318]\n- T2: [490, 296, 506, 333]\n- T3: [830, 287, 837, 307]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T2 -> T1 -> T3\nC: T1 -> T3 -> T2\nD: unable to determine\n\nTask: choose the option that gives the correct near-to-far ranking of the labeled targets relative to your position.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2f5b9c5803ca4db366c311011221b5a28998dfc282cf3612aa0248beaeaaa7a9.jpg", "reference_inputs/images/a54b9a361aecb04345c29fc9d533f9763dfc3b55579870f0ad6398d05d7a8e84.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000001", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G4, G2, G1, G6, G3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [398, 562, 403, 575] (category: pedestrian)\n- A2: [156, 423, 176, 436] (category: vehicle)\n- A3: [178, 453, 197, 472] (category: vehicle)\n- A4: [246, 453, 277, 500] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [522, 275, 531, 287] (category: vehicle)\n- G4: [470, 278, 489, 292] (category: vehicle)\n- G2: [845, 242, 878, 302] (category: pedestrian)\n- G1: [537, 250, 579, 294] (category: vehicle)\n- G6: [454, 282, 470, 290] (category: vehicle)\n- G3: [508, 278, 522, 290] (category: vehicle)\n\nAnswer options:\nA: A1->G3, A2->G1, A3->G2, A4->G5\nB: A1->G2, A2->G5, A3->G6, A4->G1\nC: A1->G3, A2->G6, A3->G5, A4->G1\nD: A1->G2, A2->G6, A3->G4, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/65bc64984b2063a131f48fe526cf6a7b07da1d78c38097687558bd599f26a609.jpg", "reference_inputs/images/326e309493b74656ab397a5c7fe4a7d782f7c238e5d6a88d3395323169ed9146.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000002", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A3, A2, A5, A4, A1, A6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [537, 250, 579, 294] (category: vehicle)\n- G2: [470, 278, 489, 292] (category: vehicle)\n- G3: [454, 282, 470, 290] (category: vehicle)\n- G4: [845, 242, 878, 302] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A3: [168, 394, 182, 407] (category: vehicle)\n- A2: [246, 453, 277, 500] (category: vehicle)\n- A5: [218, 411, 231, 424] (category: vehicle)\n- A4: [178, 453, 197, 472] (category: vehicle)\n- A1: [398, 562, 403, 575] (category: pedestrian)\n- A6: [156, 423, 176, 436] (category: vehicle)\n\nAnswer options:\nA: G1->A4, G2->A6, G3->A5, G4->A1\nB: G1->A2, G2->A4, G3->A6, G4->A1\nC: G1->A2, G2->A6, G3->A4, G4->A3\nD: G1->A2, G2->A1, G3->A4, G4->A5\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/326e309493b74656ab397a5c7fe4a7d782f7c238e5d6a88d3395323169ed9146.jpg", "reference_inputs/images/65bc64984b2063a131f48fe526cf6a7b07da1d78c38097687558bd599f26a609.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000005", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G1, G3, G4, G2, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [475, 412, 490, 423] (category: vehicle)\n- A2: [534, 448, 550, 465] (category: vehicle)\n- A3: [632, 510, 637, 531] (category: pedestrian)\n- A4: [300, 546, 306, 564] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [476, 282, 490, 290] (category: vehicle)\n- G1: [291, 270, 302, 292] (category: pedestrian)\n- G3: [534, 280, 550, 290] (category: vehicle)\n- G4: [307, 275, 314, 292] (category: pedestrian)\n- G2: [628, 272, 633, 291] (category: pedestrian)\n- G6: [706, 268, 712, 292] (category: pedestrian)\n\nAnswer options:\nA: A1->G6, A2->G4, A3->G5, A4->G2\nB: A1->G3, A2->G4, A3->G5, A4->G6\nC: A1->G5, A2->G3, A3->G2, A4->G4\nD: A1->G5, A2->G2, A3->G4, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a4cf592a72e7c039e89f2ce68a4f098467de07ca020316c2c156372cb9a1cfc5.jpg", "reference_inputs/images/9762b261fdab1ae82f5704a59390f7a7a5f789c680e4cf37c3c11ba8d2d2fc0f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000006", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A3, A4, A2, A5, A1, A6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [500, 270, 517, 289] (category: vehicle)\n- G2: [628, 272, 633, 291] (category: pedestrian)\n- G3: [307, 275, 314, 292] (category: pedestrian)\n- G4: [476, 282, 490, 290] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A3: [632, 510, 637, 531] (category: pedestrian)\n- A4: [300, 546, 306, 564] (category: pedestrian)\n- A2: [499, 388, 517, 412] (category: vehicle)\n- A5: [470, 390, 481, 402] (category: vehicle)\n- A1: [195, 548, 205, 574] (category: pedestrian)\n- A6: [475, 412, 490, 423] (category: vehicle)\n\nAnswer options:\nA: G1->A4, G2->A5, G3->A1, G4->A3\nB: G1->A6, G2->A1, G3->A5, G4->A3\nC: G1->A2, G2->A3, G3->A4, G4->A6\nD: G1->A1, G2->A2, G3->A3, G4->A6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9762b261fdab1ae82f5704a59390f7a7a5f789c680e4cf37c3c11ba8d2d2fc0f.jpg", "reference_inputs/images/a4cf592a72e7c039e89f2ce68a4f098467de07ca020316c2c156372cb9a1cfc5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000007", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G1, G6, G5, G2, G3, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [480, 355, 492, 372] (category: vehicle)\n- A2: [406, 557, 439, 575] (category: vehicle)\n- A3: [446, 446, 463, 464] (category: vehicle)\n- A4: [466, 391, 490, 399] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G1: [469, 271, 486, 289] (category: vehicle)\n- G6: [169, 258, 178, 282] (category: pedestrian)\n- G5: [381, 274, 420, 298] (category: vehicle)\n- G2: [0, 257, 188, 352] (category: vehicle)\n- G3: [446, 282, 480, 290] (category: vehicle)\n- G4: [485, 233, 753, 396] (category: vehicle)\n\nAnswer options:\nA: A1->G1, A2->G2, A3->G5, A4->G3\nB: A1->G3, A2->G1, A3->G4, A4->G6\nC: A1->G4, A2->G6, A3->G3, A4->G1\nD: A1->G3, A2->G5, A3->G6, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/54a45b98db6cf9dcae184341b11c3ada9efb63b5c8c0222a21ea269ead7746f1.jpg", "reference_inputs/images/22e82f3f5da3ca8e0a41507973a1054fc81a0436bd364afb7118546fe3ec6147.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000009", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G3, G2, G1, G6, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [510, 466, 526, 485] (category: vehicle)\n- A2: [468, 399, 480, 408] (category: vehicle)\n- A3: [641, 464, 644, 482] (category: pedestrian)\n- A4: [430, 458, 464, 520] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [926, 244, 946, 302] (category: pedestrian)\n- G3: [696, 261, 707, 294] (category: pedestrian)\n- G2: [502, 270, 558, 309] (category: vehicle)\n- G1: [13, 152, 344, 330] (category: vehicle)\n- G6: [434, 282, 454, 292] (category: vehicle)\n- G4: [452, 278, 497, 297] (category: vehicle)\n\nAnswer options:\nA: A1->G4, A2->G6, A3->G5, A4->G3\nB: A1->G2, A2->G6, A3->G5, A4->G1\nC: A1->G1, A2->G2, A3->G3, A4->G6\nD: A1->G4, A2->G1, A3->G2, A4->G3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/09d6bbd65d5060eae93487145b4ecd97a46eec4da32909caff0ec56456d840de.jpg", "reference_inputs/images/b44ca937189d4b9529c968b0f3bea457bb9b34d3c8f1a0061bfd35bda0a95e8c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000010", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A5, A6, A2, A1, A3, A4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [926, 244, 946, 302] (category: pedestrian)\n- G2: [434, 282, 454, 292] (category: vehicle)\n- G3: [502, 270, 558, 309] (category: vehicle)\n- G4: [13, 152, 344, 330] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A5: [641, 464, 644, 482] (category: pedestrian)\n- A6: [530, 398, 539, 407] (category: vehicle)\n- A2: [468, 399, 480, 408] (category: vehicle)\n- A1: [509, 427, 522, 442] (category: vehicle)\n- A3: [510, 466, 526, 485] (category: vehicle)\n- A4: [430, 458, 464, 520] (category: vehicle)\n\nAnswer options:\nA: G1->A1, G2->A5, G3->A6, G4->A2\nB: G1->A5, G2->A2, G3->A3, G4->A4\nC: G1->A5, G2->A1, G3->A2, G4->A6\nD: G1->A3, G2->A1, G3->A5, G4->A2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b44ca937189d4b9529c968b0f3bea457bb9b34d3c8f1a0061bfd35bda0a95e8c.jpg", "reference_inputs/images/09d6bbd65d5060eae93487145b4ecd97a46eec4da32909caff0ec56456d840de.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000012", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A1, A2, A6, A4, A5, A3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [398, 214, 440, 347] (category: pedestrian)\n- G2: [0, 250, 310, 358] (category: vehicle)\n- G3: [440, 251, 458, 304] (category: pedestrian)\n- G4: [539, 235, 560, 311] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A1: [322, 230, 372, 331] (category: vehicle)\n- A2: [162, 121, 205, 139] (category: pedestrian)\n- A6: [505, 172, 526, 186] (category: pedestrian)\n- A4: [785, 432, 826, 449] (category: pedestrian)\n- A5: [10, 205, 58, 222] (category: pedestrian)\n- A3: [328, 180, 351, 194] (category: pedestrian)\n\nAnswer options:\nA: G1->A3, G2->A6, G3->A4, G4->A5\nB: G1->A1, G2->A4, G3->A3, G4->A5\nC: G1->A3, G2->A1, G3->A5, G4->A2\nD: G1->A5, G2->A2, G3->A4, G4->A6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7bb72823dd1ff767bcfa0af18ffdf886134eb00fa75108eafdef971bd5ab925b.jpg", "reference_inputs/images/0270d8b538279fc1cc2d42738a8e13de18ac9681608ba8b8b64e1d8f6abef7bd.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000013", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G1, G2, G3, G6, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [501, 376, 516, 396] (category: vehicle)\n- A2: [646, 557, 651, 575] (category: pedestrian)\n- A3: [517, 459, 537, 474] (category: vehicle)\n- A4: [395, 494, 444, 575] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [498, 267, 518, 290] (category: vehicle)\n- G1: [705, 261, 714, 296] (category: pedestrian)\n- G2: [673, 261, 682, 295] (category: pedestrian)\n- G3: [520, 280, 554, 296] (category: vehicle)\n- G6: [862, 241, 884, 306] (category: pedestrian)\n- G4: [210, 206, 374, 310] (category: vehicle)\n\nAnswer options:\nA: A1->G5, A2->G6, A3->G3, A4->G4\nB: A1->G2, A2->G1, A3->G4, A4->G6\nC: A1->G3, A2->G4, A3->G1, A4->G5\nD: A1->G5, A2->G6, A3->G4, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d0c7a1e600af97e4214a01d83668ccb9038b75752431c9384f9b153978205fed.jpg", "reference_inputs/images/6c5bcc96e9ec88143a2af460a2c36c62a89566b5824bb951b130bc5915679c3b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000014", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G3, G6, G4, G1, G2, G5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [100, 519, 138, 545] (category: pedestrian)\n- A2: [183, 486, 216, 513] (category: pedestrian)\n- A3: [162, 274, 183, 287] (category: pedestrian)\n- A4: [312, 259, 365, 370] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G3: [218, 233, 246, 314] (category: pedestrian)\n- G6: [239, 248, 254, 308] (category: pedestrian)\n- G4: [370, 234, 734, 354] (category: vehicle)\n- G1: [930, 272, 938, 292] (category: pedestrian)\n- G2: [815, 274, 867, 294] (category: vehicle)\n- G5: [574, 241, 594, 268] (category: pedestrian)\n\nAnswer options:\nA: A1->G6, A2->G3, A3->G5, A4->G4\nB: A1->G2, A2->G3, A3->G5, A4->G1\nC: A1->G2, A2->G1, A3->G5, A4->G3\nD: A1->G4, A2->G3, A3->G1, A4->G2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6af64a6f155a53fd7e9405fd0240f3d88f46f26297b3a634601d63726b1edce1.jpg", "reference_inputs/images/5c64e995683b1de21c283c671349dac82351d1b3a64375df3ccdb6bffaedb802.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000015", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A1, A6, A2, A3, A5, A4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [574, 241, 594, 268] (category: pedestrian)\n- G2: [239, 248, 254, 308] (category: pedestrian)\n- G3: [218, 233, 246, 314] (category: pedestrian)\n- G4: [370, 234, 734, 354] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A1: [162, 274, 183, 287] (category: pedestrian)\n- A6: [312, 259, 365, 370] (category: vehicle)\n- A2: [100, 519, 138, 545] (category: pedestrian)\n- A3: [183, 486, 216, 513] (category: pedestrian)\n- A5: [496, 132, 554, 243] (category: vehicle)\n- A4: [754, 55, 779, 87] (category: pedestrian)\n\nAnswer options:\nA: G1->A2, G2->A1, G3->A4, G4->A6\nB: G1->A2, G2->A5, G3->A1, G4->A3\nC: G1->A1, G2->A2, G3->A3, G4->A6\nD: G1->A2, G2->A6, G3->A5, G4->A3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5c64e995683b1de21c283c671349dac82351d1b3a64375df3ccdb6bffaedb802.jpg", "reference_inputs/images/6af64a6f155a53fd7e9405fd0240f3d88f46f26297b3a634601d63726b1edce1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000016", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G4, G5, G1, G6, G2, G3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [52, 508, 118, 529] (category: vehicle)\n- A2: [170, 463, 176, 479] (category: pedestrian)\n- A3: [385, 505, 432, 525] (category: vehicle)\n- A4: [89, 470, 94, 487] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G4: [742, 280, 762, 304] (category: pedestrian)\n- G5: [686, 279, 691, 301] (category: pedestrian)\n- G1: [596, 287, 632, 303] (category: vehicle)\n- G6: [483, 269, 505, 294] (category: vehicle)\n- G2: [360, 269, 378, 293] (category: pedestrian)\n- G3: [842, 279, 1023, 356] (category: vehicle)\n\nAnswer options:\nA: A1->G4, A2->G5, A3->G2, A4->G6\nB: A1->G5, A2->G4, A3->G6, A4->G2\nC: A1->G1, A2->G4, A3->G3, A4->G5\nD: A1->G2, A2->G1, A3->G3, A4->G5\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7585f41fe0b84cff321aa7f677c0cca23c67433ce4c183eb9baafe39b873f4d3.jpg", "reference_inputs/images/b829ddf4c587e79a67ffe16fe25873fe19448bfe353db00f3602571b6511c0a3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000017", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A4, A5, A6, A3, A1, A2. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [596, 287, 632, 303] (category: vehicle)\n- G2: [842, 279, 1023, 356] (category: vehicle)\n- G3: [742, 280, 762, 304] (category: pedestrian)\n- G4: [686, 279, 691, 301] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A4: [52, 508, 118, 529] (category: vehicle)\n- A5: [550, 470, 556, 488] (category: pedestrian)\n- A6: [385, 505, 432, 525] (category: vehicle)\n- A3: [712, 524, 773, 549] (category: vehicle)\n- A1: [89, 470, 94, 487] (category: pedestrian)\n- A2: [170, 463, 176, 479] (category: pedestrian)\n\nAnswer options:\nA: G1->A4, G2->A3, G3->A2, G4->A6\nB: G1->A1, G2->A5, G3->A4, G4->A6\nC: G1->A3, G2->A6, G3->A4, G4->A2\nD: G1->A4, G2->A6, G3->A2, G4->A1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b829ddf4c587e79a67ffe16fe25873fe19448bfe353db00f3602571b6511c0a3.jpg", "reference_inputs/images/7585f41fe0b84cff321aa7f677c0cca23c67433ce4c183eb9baafe39b873f4d3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000018", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G1, G4, G3, G6, G2. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [542, 463, 560, 480] (category: vehicle)\n- A2: [522, 372, 537, 390] (category: vehicle)\n- A3: [404, 543, 438, 575] (category: vehicle)\n- A4: [482, 354, 493, 370] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [523, 273, 534, 288] (category: vehicle)\n- G1: [440, 274, 473, 298] (category: vehicle)\n- G4: [507, 275, 527, 292] (category: vehicle)\n- G3: [538, 281, 552, 290] (category: vehicle)\n- G6: [428, 282, 442, 293] (category: vehicle)\n- G2: [482, 275, 494, 287] (category: vehicle)\n\nAnswer options:\nA: A1->G1, A2->G4, A3->G6, A4->G2\nB: A1->G4, A2->G5, A3->G1, A4->G2\nC: A1->G1, A2->G3, A3->G4, A4->G6\nD: A1->G3, A2->G5, A3->G6, A4->G2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2dc7235177eab99e00758787c60e4230c1c1263feedc79745fdcd2eae1d0ed46.jpg", "reference_inputs/images/4d1429f8a559ffa76ff6dd6181113d642a7d6ccac748309b027172a63950138b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000020", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G2, G4, G1, G3, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [414, 325, 460, 426] (category: vehicle)\n- A2: [753, 274, 786, 287] (category: pedestrian)\n- A3: [822, 434, 857, 452] (category: pedestrian)\n- A4: [106, 422, 143, 453] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [178, 259, 190, 297] (category: pedestrian)\n- G2: [503, 279, 517, 290] (category: vehicle)\n- G4: [779, 257, 791, 296] (category: pedestrian)\n- G1: [675, 266, 698, 294] (category: pedestrian)\n- G3: [635, 268, 641, 292] (category: pedestrian)\n- G6: [434, 275, 472, 300] (category: vehicle)\n\nAnswer options:\nA: A1->G1, A2->G3, A3->G6, A4->G2\nB: A1->G3, A2->G4, A3->G5, A4->G1\nC: A1->G6, A2->G1, A3->G4, A4->G5\nD: A1->G6, A2->G4, A3->G3, A4->G5\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ab2188839eca588e14ececd553c05ab309b3db806f153f7d78edfdf9250a16d7.jpg", "reference_inputs/images/72cd0a8adb1640bf26b9f9f7b03503233379e3d16e3754adda473802969f4c0f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000021", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G3, G1, G2, G4, G5, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [550, 470, 556, 488] (category: pedestrian)\n- A2: [712, 524, 773, 549] (category: vehicle)\n- A3: [385, 505, 432, 525] (category: vehicle)\n- A4: [52, 508, 118, 529] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G3: [216, 280, 233, 290] (category: vehicle)\n- G1: [261, 274, 278, 288] (category: vehicle)\n- G2: [346, 274, 353, 290] (category: pedestrian)\n- G4: [0, 274, 50, 314] (category: vehicle)\n- G5: [82, 271, 88, 290] (category: pedestrian)\n- G6: [130, 279, 162, 295] (category: vehicle)\n\nAnswer options:\nA: A1->G5, A2->G3, A3->G6, A4->G4\nB: A1->G1, A2->G5, A3->G3, A4->G2\nC: A1->G5, A2->G2, A3->G3, A4->G6\nD: A1->G4, A2->G3, A3->G2, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7585f41fe0b84cff321aa7f677c0cca23c67433ce4c183eb9baafe39b873f4d3.jpg", "reference_inputs/images/6b5ca664a0bec76fcb4e83634ea93c39a1cd48cb0e2e19cac5000ed55fa74d96.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000022", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A4, A2, A6, A1, A5, A3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [82, 271, 88, 290] (category: pedestrian)\n- G2: [216, 280, 233, 290] (category: vehicle)\n- G3: [0, 274, 50, 314] (category: vehicle)\n- G4: [130, 279, 162, 295] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A4: [385, 505, 432, 525] (category: vehicle)\n- A2: [89, 470, 94, 487] (category: pedestrian)\n- A6: [52, 508, 118, 529] (category: vehicle)\n- A1: [712, 524, 773, 549] (category: vehicle)\n- A5: [170, 463, 176, 479] (category: pedestrian)\n- A3: [550, 470, 556, 488] (category: pedestrian)\n\nAnswer options:\nA: G1->A4, G2->A5, G3->A3, G4->A1\nB: G1->A2, G2->A3, G3->A1, G4->A6\nC: G1->A3, G2->A1, G3->A6, G4->A4\nD: G1->A4, G2->A3, G3->A5, G4->A6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6b5ca664a0bec76fcb4e83634ea93c39a1cd48cb0e2e19cac5000ed55fa74d96.jpg", "reference_inputs/images/7585f41fe0b84cff321aa7f677c0cca23c67433ce4c183eb9baafe39b873f4d3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000023", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G6, G2, G1, G5, G3, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [146, 413, 180, 436] (category: pedestrian)\n- A2: [148, 172, 192, 189] (category: pedestrian)\n- A3: [570, 0, 612, 83] (category: vehicle)\n- A4: [874, 446, 918, 466] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G6: [542, 275, 570, 296] (category: vehicle)\n- G2: [906, 250, 928, 298] (category: pedestrian)\n- G1: [166, 256, 172, 299] (category: pedestrian)\n- G5: [278, 262, 285, 294] (category: pedestrian)\n- G3: [36, 250, 58, 298] (category: pedestrian)\n- G4: [446, 278, 466, 291] (category: vehicle)\n\nAnswer options:\nA: A1->G4, A2->G6, A3->G2, A4->G5\nB: A1->G1, A2->G5, A3->G4, A4->G6\nC: A1->G1, A2->G5, A3->G6, A4->G2\nD: A1->G3, A2->G4, A3->G5, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/80411caf2c694c923da91acf9ce65757d6d0b3cc4e9f32a494dc213f2e0988f9.jpg", "reference_inputs/images/9ee52d4b590978577fe3a21b750b8093f46ed3b4efce315bf8098e1feb0e6325.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000024", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000024", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A3, A1, A6, A2, A4, A5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [166, 256, 172, 299] (category: pedestrian)\n- G2: [906, 250, 928, 298] (category: pedestrian)\n- G3: [278, 262, 285, 294] (category: pedestrian)\n- G4: [36, 250, 58, 298] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A3: [570, 0, 612, 83] (category: vehicle)\n- A1: [146, 413, 180, 436] (category: pedestrian)\n- A6: [759, 385, 788, 402] (category: pedestrian)\n- A2: [874, 446, 918, 466] (category: pedestrian)\n- A4: [148, 172, 192, 189] (category: pedestrian)\n- A5: [42, 446, 91, 467] (category: pedestrian)\n\nAnswer options:\nA: G1->A6, G2->A4, G3->A2, G4->A1\nB: G1->A1, G2->A2, G3->A4, G4->A5\nC: G1->A4, G2->A3, G3->A6, G4->A1\nD: G1->A2, G2->A1, G3->A6, G4->A4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9ee52d4b590978577fe3a21b750b8093f46ed3b4efce315bf8098e1feb0e6325.jpg", "reference_inputs/images/80411caf2c694c923da91acf9ce65757d6d0b3cc4e9f32a494dc213f2e0988f9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000027", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G1, G5, G6, G3, G4, G2. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [454, 428, 470, 447] (category: vehicle)\n- A2: [494, 430, 516, 466] (category: vehicle)\n- A3: [650, 502, 659, 524] (category: pedestrian)\n- A4: [458, 468, 480, 490] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G1: [321, 268, 328, 293] (category: pedestrian)\n- G5: [443, 277, 462, 291] (category: vehicle)\n- G6: [690, 266, 703, 293] (category: pedestrian)\n- G3: [446, 281, 474, 294] (category: vehicle)\n- G4: [491, 258, 518, 292] (category: vehicle)\n- G2: [563, 272, 606, 300] (category: vehicle)\n\nAnswer options:\nA: A1->G2, A2->G3, A3->G6, A4->G1\nB: A1->G5, A2->G4, A3->G6, A4->G3\nC: A1->G3, A2->G2, A3->G4, A4->G6\nD: A1->G1, A2->G5, A3->G3, A4->G4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0338e6668ccc2980e70ebd4a9e32d2deccdc84389a69907b91810ac1184821af.jpg", "reference_inputs/images/34d027dbab11397f568f524e93a1358bb23b7502da274b0c04a1dba07548b837.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000028", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A5, A3, A6, A2, A1, A4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [563, 272, 606, 300] (category: vehicle)\n- G2: [690, 266, 703, 293] (category: pedestrian)\n- G3: [443, 277, 462, 291] (category: vehicle)\n- G4: [491, 258, 518, 292] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A5: [458, 468, 480, 490] (category: vehicle)\n- A3: [650, 502, 659, 524] (category: pedestrian)\n- A6: [549, 555, 576, 575] (category: vehicle)\n- A2: [386, 454, 393, 470] (category: pedestrian)\n- A1: [494, 430, 516, 466] (category: vehicle)\n- A4: [454, 428, 470, 447] (category: vehicle)\n\nAnswer options:\nA: G1->A4, G2->A1, G3->A2, G4->A6\nB: G1->A1, G2->A4, G3->A6, G4->A3\nC: G1->A6, G2->A3, G3->A4, G4->A1\nD: G1->A6, G2->A2, G3->A1, G4->A4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/34d027dbab11397f568f524e93a1358bb23b7502da274b0c04a1dba07548b837.jpg", "reference_inputs/images/0338e6668ccc2980e70ebd4a9e32d2deccdc84389a69907b91810ac1184821af.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000029", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G4, G3, G2, G5, G1, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [650, 519, 655, 542] (category: pedestrian)\n- A2: [423, 481, 449, 512] (category: vehicle)\n- A3: [635, 562, 641, 575] (category: pedestrian)\n- A4: [534, 490, 555, 514] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G4: [843, 269, 850, 292] (category: pedestrian)\n- G3: [914, 262, 924, 294] (category: pedestrian)\n- G2: [740, 276, 763, 295] (category: vehicle)\n- G5: [891, 258, 902, 297] (category: pedestrian)\n- G1: [659, 270, 678, 289] (category: vehicle)\n- G6: [595, 274, 626, 294] (category: vehicle)\n\nAnswer options:\nA: A1->G1, A2->G6, A3->G5, A4->G4\nB: A1->G5, A2->G4, A3->G1, A4->G2\nC: A1->G3, A2->G6, A3->G5, A4->G2\nD: A1->G4, A2->G2, A3->G5, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6d76248e7087fb5fcf6decf03cdcb4e685fdbb05d11f13983b35973a1a64e154.jpg", "reference_inputs/images/74f386330c7d9af7076f742a7d66db3df65f2fcc5027d9e6f0c2bb2d6c20ab41.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000030", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A6, A2, A1, A5, A3, A4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [914, 262, 924, 294] (category: pedestrian)\n- G2: [659, 270, 678, 289] (category: vehicle)\n- G3: [891, 258, 902, 297] (category: pedestrian)\n- G4: [595, 274, 626, 294] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A6: [423, 481, 449, 512] (category: vehicle)\n- A2: [469, 370, 484, 391] (category: vehicle)\n- A1: [650, 519, 655, 542] (category: pedestrian)\n- A5: [635, 562, 641, 575] (category: pedestrian)\n- A3: [503, 414, 532, 427] (category: vehicle)\n- A4: [534, 490, 555, 514] (category: vehicle)\n\nAnswer options:\nA: G1->A1, G2->A2, G3->A5, G4->A6\nB: G1->A6, G2->A3, G3->A2, G4->A4\nC: G1->A2, G2->A1, G3->A6, G4->A5\nD: G1->A2, G2->A4, G3->A5, G4->A1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/74f386330c7d9af7076f742a7d66db3df65f2fcc5027d9e6f0c2bb2d6c20ab41.jpg", "reference_inputs/images/6d76248e7087fb5fcf6decf03cdcb4e685fdbb05d11f13983b35973a1a64e154.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000031", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000031", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G5, G6, G3, G1, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [273, 557, 280, 575] (category: pedestrian)\n- A2: [531, 416, 544, 428] (category: vehicle)\n- A3: [460, 417, 475, 429] (category: vehicle)\n- A4: [508, 518, 530, 546] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [535, 280, 549, 290] (category: vehicle)\n- G5: [507, 278, 535, 297] (category: vehicle)\n- G6: [145, 254, 166, 301] (category: pedestrian)\n- G3: [182, 259, 198, 295] (category: pedestrian)\n- G1: [452, 282, 470, 290] (category: vehicle)\n- G4: [471, 280, 499, 293] (category: vehicle)\n\nAnswer options:\nA: A1->G3, A2->G2, A3->G1, A4->G5\nB: A1->G4, A2->G6, A3->G3, A4->G1\nC: A1->G2, A2->G3, A3->G1, A4->G4\nD: A1->G4, A2->G6, A3->G2, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1c4a984fa871acdb7a4fcedb40fb895e0a1f4c3d65635ac5eb19c51fb3d8efe8.jpg", "reference_inputs/images/dbdbafd1873f8374fc2da3b5dc8e39767781e722be8ed25fdcf2da41eac4a224.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000032", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A5, A2, A3, A6, A1, A4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [182, 259, 198, 295] (category: pedestrian)\n- G2: [535, 280, 549, 290] (category: vehicle)\n- G3: [452, 282, 470, 290] (category: vehicle)\n- G4: [507, 278, 535, 297] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A5: [531, 416, 544, 428] (category: vehicle)\n- A2: [273, 557, 280, 575] (category: pedestrian)\n- A3: [508, 458, 525, 478] (category: vehicle)\n- A6: [508, 518, 530, 546] (category: vehicle)\n- A1: [508, 397, 520, 405] (category: vehicle)\n- A4: [460, 417, 475, 429] (category: vehicle)\n\nAnswer options:\nA: G1->A4, G2->A1, G3->A3, G4->A5\nB: G1->A3, G2->A6, G3->A5, G4->A2\nC: G1->A2, G2->A5, G3->A4, G4->A6\nD: G1->A5, G2->A3, G3->A6, G4->A1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/dbdbafd1873f8374fc2da3b5dc8e39767781e722be8ed25fdcf2da41eac4a224.jpg", "reference_inputs/images/1c4a984fa871acdb7a4fcedb40fb895e0a1f4c3d65635ac5eb19c51fb3d8efe8.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000033", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000033", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G5, G3, G6, G4, G1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [434, 472, 457, 501] (category: vehicle)\n- A2: [686, 549, 691, 574] (category: pedestrian)\n- A3: [441, 515, 464, 546] (category: vehicle)\n- A4: [503, 423, 517, 438] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [671, 266, 679, 294] (category: pedestrian)\n- G5: [872, 234, 896, 307] (category: pedestrian)\n- G3: [906, 243, 924, 301] (category: pedestrian)\n- G6: [499, 275, 520, 292] (category: vehicle)\n- G4: [388, 269, 422, 289] (category: vehicle)\n- G1: [368, 274, 422, 303] (category: vehicle)\n\nAnswer options:\nA: A1->G1, A2->G3, A3->G5, A4->G6\nB: A1->G1, A2->G3, A3->G5, A4->G4\nC: A1->G1, A2->G2, A3->G6, A4->G5\nD: A1->G4, A2->G3, A3->G1, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/369d8c5a94fad57873a397e17382502d73f237070752a22f9e7d59767de6f1e1.jpg", "reference_inputs/images/52169b7a0b342f2d63699bcc716256cbbe071a240289254b98442a1483faace4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000034", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000034", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G3, G4, G5, G2, G1, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [122, 155, 167, 171] (category: pedestrian)\n- A2: [500, 406, 551, 518] (category: vehicle)\n- A3: [52, 151, 97, 170] (category: pedestrian)\n- A4: [325, 0, 378, 54] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G3: [693, 263, 702, 294] (category: pedestrian)\n- G4: [390, 279, 431, 297] (category: vehicle)\n- G5: [213, 262, 224, 296] (category: pedestrian)\n- G2: [166, 258, 178, 290] (category: pedestrian)\n- G1: [442, 279, 460, 291] (category: vehicle)\n- G6: [495, 247, 573, 315] (category: vehicle)\n\nAnswer options:\nA: A1->G6, A2->G3, A3->G4, A4->G1\nB: A1->G5, A2->G3, A3->G4, A4->G1\nC: A1->G5, A2->G1, A3->G6, A4->G3\nD: A1->G5, A2->G6, A3->G2, A4->G4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/686afe55c2fa06299ea2faa81211ce9eaa6045b06db4b14b5dbd4117d79dc52f.jpg", "reference_inputs/images/176edff7c377ecfb0a85bbb5fe6f909debb69e63dd867a4d9757b6b09591aec1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000035", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000035", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A6, A1, A3, A2, A5, A4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [166, 258, 178, 290] (category: pedestrian)\n- G2: [213, 262, 224, 296] (category: pedestrian)\n- G3: [495, 247, 573, 315] (category: vehicle)\n- G4: [390, 279, 431, 297] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A6: [122, 155, 167, 171] (category: pedestrian)\n- A1: [52, 151, 97, 170] (category: pedestrian)\n- A3: [505, 0, 552, 88] (category: vehicle)\n- A2: [500, 406, 551, 518] (category: vehicle)\n- A5: [799, 0, 808, 10] (category: pedestrian)\n- A4: [325, 0, 378, 54] (category: vehicle)\n\nAnswer options:\nA: G1->A4, G2->A2, G3->A6, G4->A5\nB: G1->A1, G2->A6, G3->A2, G4->A4\nC: G1->A5, G2->A3, G3->A1, G4->A6\nD: G1->A3, G2->A2, G3->A4, G4->A5\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/176edff7c377ecfb0a85bbb5fe6f909debb69e63dd867a4d9757b6b09591aec1.jpg", "reference_inputs/images/686afe55c2fa06299ea2faa81211ce9eaa6045b06db4b14b5dbd4117d79dc52f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000036", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000036", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A1, A5, A6, A3, A4, A2. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [562, 281, 578, 291] (category: vehicle)\n- G2: [522, 275, 534, 293] (category: vehicle)\n- G3: [389, 264, 397, 294] (category: pedestrian)\n- G4: [497, 274, 526, 297] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A1: [509, 402, 520, 412] (category: vehicle)\n- A5: [422, 522, 429, 546] (category: pedestrian)\n- A6: [501, 501, 522, 530] (category: vehicle)\n- A3: [512, 449, 529, 472] (category: vehicle)\n- A4: [587, 524, 619, 553] (category: vehicle)\n- A2: [554, 425, 567, 436] (category: vehicle)\n\nAnswer options:\nA: G1->A2, G2->A3, G3->A5, G4->A6\nB: G1->A1, G2->A2, G3->A5, G4->A6\nC: G1->A3, G2->A6, G3->A4, G4->A2\nD: G1->A6, G2->A4, G3->A5, G4->A3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/314d01bbfb4d254b346e7f6bc4fb62da813832c57ce8b979e13a5d0aae67e2bd.jpg", "reference_inputs/images/29a6541a1c03ccfac6d073b89565f4d67321d2e2acdf67f5fe0a474b8c507135.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000037", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000037", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G1, G5, G6, G3, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [346, 482, 351, 500] (category: pedestrian)\n- A2: [427, 489, 452, 522] (category: vehicle)\n- A3: [459, 406, 474, 417] (category: vehicle)\n- A4: [516, 418, 530, 427] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [188, 250, 329, 316] (category: vehicle)\n- G1: [2, 251, 40, 302] (category: pedestrian)\n- G5: [518, 281, 546, 294] (category: vehicle)\n- G6: [424, 282, 450, 293] (category: vehicle)\n- G3: [862, 238, 897, 306] (category: pedestrian)\n- G4: [700, 259, 728, 298] (category: pedestrian)\n\nAnswer options:\nA: A1->G5, A2->G6, A3->G1, A4->G3\nB: A1->G5, A2->G3, A3->G1, A4->G6\nC: A1->G1, A2->G2, A3->G6, A4->G5\nD: A1->G4, A2->G2, A3->G6, A4->G3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/01a33378cb0d1c10ec0222f6b5d05e9e374d1a4bae5fd14ef92f9858de4d88ff.jpg", "reference_inputs/images/948eb1874b239d46c2fb2d23b88f7be52db81266aea190904fab7655e01e84d4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000039", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000039", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A2, A3, A4, A1, A6, A5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [327, 282, 368, 297] (category: vehicle)\n- G2: [384, 284, 398, 291] (category: vehicle)\n- G3: [278, 278, 293, 295] (category: vehicle)\n- G4: [152, 270, 158, 292] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A2: [587, 524, 619, 553] (category: vehicle)\n- A3: [512, 449, 529, 472] (category: vehicle)\n- A4: [422, 522, 429, 546] (category: pedestrian)\n- A1: [554, 425, 567, 436] (category: vehicle)\n- A6: [501, 501, 522, 530] (category: vehicle)\n- A5: [509, 402, 520, 412] (category: vehicle)\n\nAnswer options:\nA: G1->A2, G2->A1, G3->A3, G4->A6\nB: G1->A5, G2->A3, G3->A4, G4->A6\nC: G1->A3, G2->A4, G3->A5, G4->A2\nD: G1->A2, G2->A5, G3->A6, G4->A4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/006e4b97e40fb9fb2f043974ca480bf6e4fbbcfac6f82f579c3ab98b45c6c1f4.jpg", "reference_inputs/images/29a6541a1c03ccfac6d073b89565f4d67321d2e2acdf67f5fe0a474b8c507135.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000040", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000040", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G3, G6, G1, G5, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [164, 222, 209, 237] (category: pedestrian)\n- A2: [147, 526, 180, 559] (category: pedestrian)\n- A3: [818, 76, 858, 97] (category: pedestrian)\n- A4: [823, 518, 856, 541] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [100, 250, 118, 302] (category: pedestrian)\n- G3: [266, 260, 274, 294] (category: pedestrian)\n- G6: [284, 266, 304, 294] (category: pedestrian)\n- G1: [694, 266, 700, 292] (category: pedestrian)\n- G5: [865, 248, 881, 299] (category: pedestrian)\n- G4: [635, 270, 641, 292] (category: pedestrian)\n\nAnswer options:\nA: A1->G6, A2->G2, A3->G1, A4->G5\nB: A1->G3, A2->G2, A3->G4, A4->G1\nC: A1->G6, A2->G1, A3->G4, A4->G5\nD: A1->G3, A2->G4, A3->G2, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b6e69bb6b20e30c6251d467415a026ad170366486c78c77a0625195e38cab5b5.jpg", "reference_inputs/images/214450fcb9ec8f693510b00eb2718a24d336b00136b94d6785fc63ff6db98864.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000041", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000041", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G6, G2, G5, G4, G3, G1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [754, 234, 781, 245] (category: pedestrian)\n- A2: [775, 52, 808, 77] (category: pedestrian)\n- A3: [363, 417, 412, 519] (category: vehicle)\n- A4: [315, 198, 374, 307] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G6: [217, 264, 356, 322] (category: vehicle)\n- G2: [709, 260, 718, 295] (category: pedestrian)\n- G5: [703, 265, 709, 292] (category: pedestrian)\n- G4: [326, 262, 382, 305] (category: vehicle)\n- G3: [748, 256, 761, 298] (category: pedestrian)\n- G1: [496, 273, 522, 294] (category: vehicle)\n\nAnswer options:\nA: A1->G4, A2->G5, A3->G1, A4->G6\nB: A1->G5, A2->G4, A3->G2, A4->G3\nC: A1->G1, A2->G5, A3->G2, A4->G4\nD: A1->G3, A2->G2, A3->G6, A4->G4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/19b97407fa1dc9b3bb12ab0247fd4fa6b09c797c0452b5a47ff0ca4a0408ea8c.jpg", "reference_inputs/images/e071a3d5ffba4f7bb6addc5eae0d5ad70da90edbd8c3be2635e149f0f653c9c0.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000044", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000044", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G3, G4, G6, G1, G2, G5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [158, 473, 163, 490] (category: pedestrian)\n- A2: [140, 542, 158, 563] (category: pedestrian)\n- A3: [352, 522, 381, 547] (category: vehicle)\n- A4: [244, 503, 271, 530] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G3: [763, 259, 776, 298] (category: pedestrian)\n- G4: [182, 259, 195, 298] (category: pedestrian)\n- G6: [521, 279, 558, 297] (category: vehicle)\n- G1: [400, 280, 434, 294] (category: vehicle)\n- G2: [616, 270, 631, 292] (category: pedestrian)\n- G5: [342, 270, 358, 292] (category: pedestrian)\n\nAnswer options:\nA: A1->G2, A2->G4, A3->G1, A4->G6\nB: A1->G3, A2->G2, A3->G1, A4->G5\nC: A1->G3, A2->G2, A3->G1, A4->G6\nD: A1->G6, A2->G5, A3->G1, A4->G2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0afbd9401679e8a5f9f739e851e2d6565ca0847966b9c870ff25e6bd5f0227c3.jpg", "reference_inputs/images/6a7f9e051b900b178f7d7c6eefc92cdd3bb4cf1baf907fdc77313333c8ba0243.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000045", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000045", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G1, G4, G6, G3, G2, G5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [454, 432, 471, 443] (category: vehicle)\n- A2: [533, 422, 549, 434] (category: vehicle)\n- A3: [510, 401, 521, 411] (category: vehicle)\n- A4: [506, 367, 520, 385] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G1: [509, 277, 532, 294] (category: vehicle)\n- G4: [332, 278, 397, 302] (category: vehicle)\n- G6: [502, 263, 526, 290] (category: vehicle)\n- G3: [565, 278, 612, 299] (category: vehicle)\n- G2: [454, 271, 484, 296] (category: vehicle)\n- G5: [442, 279, 454, 291] (category: vehicle)\n\nAnswer options:\nA: A1->G4, A2->G3, A3->G1, A4->G6\nB: A1->G4, A2->G1, A3->G2, A4->G6\nC: A1->G1, A2->G2, A3->G3, A4->G6\nD: A1->G5, A2->G1, A3->G6, A4->G2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/df59edf618f1768dd5184af7d6da5fda60eb9c250beb66b6d42ac32de9770ce0.jpg", "reference_inputs/images/cbfef6891d3e74555360b427a6a7211fb069abf7f7d51a894e28baa461754599.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000046", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000046", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A3, A4, A5, A1, A6, A2. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [332, 278, 397, 302] (category: vehicle)\n- G2: [565, 278, 612, 299] (category: vehicle)\n- G3: [509, 277, 532, 294] (category: vehicle)\n- G4: [624, 241, 754, 319] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A3: [487, 405, 499, 419] (category: vehicle)\n- A4: [434, 482, 454, 506] (category: vehicle)\n- A5: [454, 432, 471, 443] (category: vehicle)\n- A1: [538, 455, 557, 474] (category: vehicle)\n- A6: [510, 401, 521, 411] (category: vehicle)\n- A2: [533, 422, 549, 434] (category: vehicle)\n\nAnswer options:\nA: G1->A6, G2->A5, G3->A1, G4->A3\nB: G1->A6, G2->A4, G3->A1, G4->A2\nC: G1->A6, G2->A1, G3->A3, G4->A5\nD: G1->A5, G2->A2, G3->A6, G4->A1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cbfef6891d3e74555360b427a6a7211fb069abf7f7d51a894e28baa461754599.jpg", "reference_inputs/images/df59edf618f1768dd5184af7d6da5fda60eb9c250beb66b6d42ac32de9770ce0.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000047", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000047", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G1, G2, G6, G3, G5, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [398, 87, 454, 201] (category: vehicle)\n- A2: [504, 364, 554, 476] (category: vehicle)\n- A3: [812, 295, 846, 313] (category: pedestrian)\n- A4: [766, 0, 795, 29] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G1: [506, 269, 550, 304] (category: vehicle)\n- G2: [786, 256, 807, 299] (category: pedestrian)\n- G6: [329, 266, 335, 293] (category: pedestrian)\n- G3: [746, 262, 757, 296] (category: pedestrian)\n- G5: [650, 267, 657, 292] (category: pedestrian)\n- G4: [439, 279, 477, 297] (category: vehicle)\n\nAnswer options:\nA: A1->G6, A2->G1, A3->G2, A4->G4\nB: A1->G4, A2->G1, A3->G3, A4->G5\nC: A1->G2, A2->G4, A3->G5, A4->G3\nD: A1->G5, A2->G6, A3->G4, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9b07624335b4fa633b68a1113aa88dc561c5ac1dcdfee58fdfa373ca97d99437.jpg", "reference_inputs/images/67d3c4dfaf5334baa46fdb18857f233075e29e32dcbad3903bb91dcb5950e373.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_20m/000048", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000048", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A4, A6, A1, A3, A2, A5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [439, 279, 477, 297] (category: vehicle)\n- G2: [650, 267, 657, 292] (category: pedestrian)\n- G3: [506, 269, 550, 304] (category: vehicle)\n- G4: [746, 262, 757, 296] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A4: [766, 0, 795, 29] (category: pedestrian)\n- A6: [504, 364, 554, 476] (category: vehicle)\n- A1: [398, 87, 454, 201] (category: vehicle)\n- A3: [168, 341, 208, 355] (category: pedestrian)\n- A2: [122, 443, 161, 462] (category: pedestrian)\n- A5: [812, 295, 846, 313] (category: pedestrian)\n\nAnswer options:\nA: G1->A2, G2->A3, G3->A1, G4->A5\nB: G1->A4, G2->A5, G3->A1, G4->A3\nC: G1->A3, G2->A4, G3->A5, G4->A2\nD: G1->A1, G2->A4, G3->A6, G4->A5\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/67d3c4dfaf5334baa46fdb18857f233075e29e32dcbad3903bb91dcb5950e373.jpg", "reference_inputs/images/9b07624335b4fa633b68a1113aa88dc561c5ac1dcdfee58fdfa373ca97d99437.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000001", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G3, G1, G4, G5, G6, G2. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [476, 410, 490, 432] (category: vehicle)\n- A2: [526, 434, 540, 456] (category: vehicle)\n- A3: [459, 472, 486, 485] (category: vehicle)\n- A4: [506, 570, 527, 575] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G3: [536, 261, 560, 279] (category: vehicle)\n- G1: [460, 267, 479, 290] (category: vehicle)\n- G4: [76, 253, 277, 326] (category: vehicle)\n- G5: [492, 274, 568, 311] (category: vehicle)\n- G6: [872, 279, 893, 293] (category: vehicle)\n- G2: [414, 281, 466, 294] (category: vehicle)\n\nAnswer options:\nA: A1->G4, A2->G2, A3->G1, A4->G3\nB: A1->G1, A2->G3, A3->G2, A4->G5\nC: A1->G5, A2->G4, A3->G6, A4->G1\nD: A1->G1, A2->G5, A3->G6, A4->G2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f5158663b6493a36ff24cee09cee5d0205729b56889b7518818f61b53f9e3b4b.jpg", "reference_inputs/images/b683797c51a05883b7e76e4ff54291fc1eb45bd00558f4d84c85fd38534d3947.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000002", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G4, G2, G5, G3, G6, G1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [422, 152, 459, 227] (category: vehicle)\n- A2: [676, 166, 688, 178] (category: pedestrian)\n- A3: [244, 169, 261, 182] (category: pedestrian)\n- A4: [288, 195, 303, 208] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G4: [380, 267, 383, 294] (category: pedestrian)\n- G2: [412, 274, 417, 293] (category: pedestrian)\n- G5: [718, 271, 725, 297] (category: pedestrian)\n- G3: [494, 282, 529, 299] (category: vehicle)\n- G6: [408, 274, 411, 293] (category: pedestrian)\n- G1: [345, 270, 352, 295] (category: pedestrian)\n\nAnswer options:\nA: A1->G3, A2->G5, A3->G1, A4->G4\nB: A1->G4, A2->G1, A3->G3, A4->G5\nC: A1->G4, A2->G5, A3->G6, A4->G3\nD: A1->G1, A2->G4, A3->G3, A4->G2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/65ab20dcbba598480d730324f3c9ceffe765f5fccfc0bf381aeb02abec648f0f.jpg", "reference_inputs/images/9c6b3f29f4d9176113b4baac3a280ea580c2ac94190eae813932e337951513d2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000003", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G1, G5, G2, G3, G4, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [281, 205, 296, 215] (category: pedestrian)\n- A2: [446, 367, 479, 440] (category: vehicle)\n- A3: [678, 289, 690, 303] (category: pedestrian)\n- A4: [697, 205, 716, 215] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G1: [888, 266, 1023, 399] (category: vehicle)\n- G5: [254, 235, 423, 333] (category: vehicle)\n- G2: [782, 255, 797, 298] (category: pedestrian)\n- G3: [496, 246, 535, 295] (category: vehicle)\n- G4: [181, 255, 198, 298] (category: pedestrian)\n- G6: [829, 243, 846, 301] (category: pedestrian)\n\nAnswer options:\nA: A1->G4, A2->G5, A3->G6, A4->G2\nB: A1->G4, A2->G3, A3->G2, A4->G6\nC: A1->G5, A2->G1, A3->G3, A4->G6\nD: A1->G5, A2->G4, A3->G1, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c236d9347f55c0c3ec5ba178d174cdfed7be3be0fb69abec27fac1b60b6cc6cf.jpg", "reference_inputs/images/64ecb14e96931fa12d63d7d52e97e6551d9ab8794bec873526ad42f1cf64fd39.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000004", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A2, A4, A1, A6, A5, A3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [502, 261, 572, 314] (category: vehicle)\n- G2: [782, 255, 797, 298] (category: pedestrian)\n- G3: [254, 235, 423, 333] (category: vehicle)\n- G4: [829, 243, 846, 301] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A2: [691, 520, 704, 537] (category: pedestrian)\n- A4: [507, 283, 539, 356] (category: vehicle)\n- A1: [446, 367, 479, 440] (category: vehicle)\n- A6: [697, 205, 716, 215] (category: pedestrian)\n- A5: [678, 289, 690, 303] (category: pedestrian)\n- A3: [562, 484, 600, 559] (category: vehicle)\n\nAnswer options:\nA: G1->A4, G2->A3, G3->A6, G4->A5\nB: G1->A1, G2->A5, G3->A4, G4->A2\nC: G1->A4, G2->A6, G3->A1, G4->A5\nD: G1->A3, G2->A6, G3->A5, G4->A2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/64ecb14e96931fa12d63d7d52e97e6551d9ab8794bec873526ad42f1cf64fd39.jpg", "reference_inputs/images/c236d9347f55c0c3ec5ba178d174cdfed7be3be0fb69abec27fac1b60b6cc6cf.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000005", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G3, G1, G2, G6, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [558, 474, 566, 498] (category: pedestrian)\n- A2: [504, 43, 512, 60] (category: pedestrian)\n- A3: [0, 250, 125, 314] (category: vehicle)\n- A4: [545, 174, 622, 215] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [510, 238, 519, 265] (category: pedestrian)\n- G3: [853, 227, 859, 246] (category: pedestrian)\n- G1: [646, 227, 674, 255] (category: vehicle)\n- G2: [916, 218, 926, 246] (category: pedestrian)\n- G6: [572, 241, 575, 257] (category: pedestrian)\n- G4: [781, 238, 833, 258] (category: vehicle)\n\nAnswer options:\nA: A1->G5, A2->G2, A3->G1, A4->G4\nB: A1->G3, A2->G4, A3->G1, A4->G6\nC: A1->G2, A2->G6, A3->G5, A4->G4\nD: A1->G5, A2->G4, A3->G2, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9387fb07c1ea73e8176f603aa4bc4630b33ca3cd16a48c5c1865d623fa341a4d.jpg", "reference_inputs/images/a0ccf0f5e696281230600960e2603ce51a994ea07386f34ba1410ded71be1b4c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000006", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A1, A6, A3, A2, A4, A5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [916, 218, 926, 246] (category: pedestrian)\n- G2: [646, 227, 674, 255] (category: vehicle)\n- G3: [510, 238, 519, 265] (category: pedestrian)\n- G4: [781, 238, 833, 258] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A1: [504, 43, 512, 60] (category: pedestrian)\n- A6: [545, 174, 622, 215] (category: vehicle)\n- A3: [0, 250, 125, 314] (category: vehicle)\n- A2: [558, 474, 566, 498] (category: pedestrian)\n- A4: [247, 0, 263, 18] (category: pedestrian)\n- A5: [626, 26, 636, 47] (category: pedestrian)\n\nAnswer options:\nA: G1->A1, G2->A3, G3->A2, G4->A6\nB: G1->A2, G2->A5, G3->A6, G4->A1\nC: G1->A3, G2->A1, G3->A4, G4->A2\nD: G1->A3, G2->A6, G3->A1, G4->A4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a0ccf0f5e696281230600960e2603ce51a994ea07386f34ba1410ded71be1b4c.jpg", "reference_inputs/images/9387fb07c1ea73e8176f603aa4bc4630b33ca3cd16a48c5c1865d623fa341a4d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000007", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G3, G6, G1, G5, G2, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [286, 77, 307, 90] (category: pedestrian)\n- A2: [240, 231, 257, 245] (category: pedestrian)\n- A3: [694, 231, 707, 245] (category: pedestrian)\n- A4: [214, 203, 242, 212] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G3: [152, 261, 163, 296] (category: pedestrian)\n- G6: [753, 258, 765, 298] (category: pedestrian)\n- G1: [597, 275, 601, 289] (category: pedestrian)\n- G5: [162, 258, 173, 298] (category: pedestrian)\n- G2: [304, 267, 320, 294] (category: pedestrian)\n- G4: [254, 261, 264, 295] (category: pedestrian)\n\nAnswer options:\nA: A1->G1, A2->G4, A3->G3, A4->G5\nB: A1->G2, A2->G5, A3->G6, A4->G3\nC: A1->G5, A2->G1, A3->G6, A4->G2\nD: A1->G2, A2->G5, A3->G3, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5b432ec76c85bb5a57956a33adb0353060181098b08c679dfb6e3d264a7c7ff0.jpg", "reference_inputs/images/9a6e5f9f30ee237528c1c8bea12e3e2df67813b2f7eaf7df22f916e035be8caa.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000008", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G1, G6, G5, G3, G2, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [231, 238, 252, 246] (category: pedestrian)\n- A2: [239, 139, 257, 155] (category: pedestrian)\n- A3: [250, 279, 270, 287] (category: pedestrian)\n- A4: [234, 246, 254, 254] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G1: [719, 247, 740, 306] (category: pedestrian)\n- G6: [531, 250, 546, 306] (category: pedestrian)\n- G5: [32, 242, 54, 310] (category: pedestrian)\n- G3: [410, 253, 430, 306] (category: pedestrian)\n- G2: [516, 246, 534, 306] (category: pedestrian)\n- G4: [450, 248, 467, 308] (category: pedestrian)\n\nAnswer options:\nA: A1->G1, A2->G5, A3->G6, A4->G4\nB: A1->G6, A2->G1, A3->G4, A4->G2\nC: A1->G3, A2->G6, A3->G2, A4->G4\nD: A1->G5, A2->G4, A3->G6, A4->G2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3c008430886d13cc06b3af1ca92b44323c9fdc224aa91af5bf1a822dea3e92b5.jpg", "reference_inputs/images/9bd1be9135b490482662595471ecb13389cc21f0185e9afd81e9b1cf2152433c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000009", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A1, A6, A5, A4, A2, A3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [719, 247, 740, 306] (category: pedestrian)\n- G2: [450, 248, 467, 308] (category: pedestrian)\n- G3: [516, 246, 534, 306] (category: pedestrian)\n- G4: [531, 250, 546, 306] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A1: [690, 436, 704, 450] (category: pedestrian)\n- A6: [234, 246, 254, 254] (category: pedestrian)\n- A5: [214, 303, 244, 313] (category: pedestrian)\n- A4: [239, 139, 257, 155] (category: pedestrian)\n- A2: [250, 279, 270, 287] (category: pedestrian)\n- A3: [231, 238, 252, 246] (category: pedestrian)\n\nAnswer options:\nA: G1->A4, G2->A5, G3->A3, G4->A2\nB: G1->A4, G2->A6, G3->A1, G4->A3\nC: G1->A4, G2->A2, G3->A6, G4->A3\nD: G1->A5, G2->A6, G3->A3, G4->A4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9bd1be9135b490482662595471ecb13389cc21f0185e9afd81e9b1cf2152433c.jpg", "reference_inputs/images/3c008430886d13cc06b3af1ca92b44323c9fdc224aa91af5bf1a822dea3e92b5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000010", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G4, G3, G6, G2, G5, G1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [668, 494, 694, 509] (category: pedestrian)\n- A2: [715, 294, 735, 302] (category: pedestrian)\n- A3: [664, 341, 682, 350] (category: pedestrian)\n- A4: [447, 526, 478, 575] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G4: [685, 267, 691, 291] (category: pedestrian)\n- G3: [654, 266, 662, 293] (category: pedestrian)\n- G6: [414, 273, 464, 304] (category: vehicle)\n- G2: [608, 274, 612, 290] (category: pedestrian)\n- G5: [706, 262, 738, 295] (category: pedestrian)\n- G1: [503, 278, 518, 290] (category: vehicle)\n\nAnswer options:\nA: A1->G5, A2->G4, A3->G3, A4->G6\nB: A1->G1, A2->G2, A3->G6, A4->G3\nC: A1->G4, A2->G1, A3->G5, A4->G6\nD: A1->G5, A2->G4, A3->G6, A4->G3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a1d1537e3be8850fe6f3702a692d2159724b057ca4725db2e331af7a19f6bdbf.jpg", "reference_inputs/images/fc2f455a3a134b67bdb86f8967c4e888682308e77817c4a1c0ede8e7b7ffb7cb.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000011", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A5, A6, A3, A1, A2, A4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [706, 262, 738, 295] (category: pedestrian)\n- G2: [685, 267, 691, 291] (category: pedestrian)\n- G3: [414, 273, 464, 304] (category: vehicle)\n- G4: [654, 266, 662, 293] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A5: [674, 35, 686, 54] (category: pedestrian)\n- A6: [715, 294, 735, 302] (category: pedestrian)\n- A3: [664, 341, 682, 350] (category: pedestrian)\n- A1: [447, 526, 478, 575] (category: vehicle)\n- A2: [668, 494, 694, 509] (category: pedestrian)\n- A4: [414, 167, 444, 234] (category: vehicle)\n\nAnswer options:\nA: G1->A5, G2->A4, G3->A3, G4->A1\nB: G1->A3, G2->A1, G3->A4, G4->A6\nC: G1->A4, G2->A2, G3->A1, G4->A5\nD: G1->A2, G2->A6, G3->A1, G4->A3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fc2f455a3a134b67bdb86f8967c4e888682308e77817c4a1c0ede8e7b7ffb7cb.jpg", "reference_inputs/images/a1d1537e3be8850fe6f3702a692d2159724b057ca4725db2e331af7a19f6bdbf.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000012", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G1, G5, G6, G3, G2, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [302, 360, 324, 367] (category: pedestrian)\n- A2: [287, 222, 302, 236] (category: pedestrian)\n- A3: [58, 530, 86, 546] (category: pedestrian)\n- A4: [254, 479, 272, 495] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G1: [607, 242, 642, 310] (category: pedestrian)\n- G5: [275, 238, 290, 314] (category: pedestrian)\n- G6: [827, 245, 853, 303] (category: pedestrian)\n- G3: [0, 240, 10, 318] (category: pedestrian)\n- G2: [200, 266, 210, 301] (category: pedestrian)\n- G4: [68, 244, 92, 310] (category: pedestrian)\n\nAnswer options:\nA: A1->G4, A2->G1, A3->G6, A4->G3\nB: A1->G5, A2->G1, A3->G2, A4->G4\nC: A1->G3, A2->G4, A3->G2, A4->G6\nD: A1->G3, A2->G2, A3->G6, A4->G4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5503db5f5ae4225d2c6f3add36d2f86fa8ee0bd3d0e0cce90ad74ba26b5a2990.jpg", "reference_inputs/images/ee43fb8141aa4576e3ebdf22ef02f663d6e11cac7f7e49ef76189736638c9d2a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000013", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A3, A6, A5, A4, A2, A1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [827, 245, 853, 303] (category: pedestrian)\n- G2: [607, 242, 642, 310] (category: pedestrian)\n- G3: [200, 266, 210, 301] (category: pedestrian)\n- G4: [68, 244, 92, 310] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A3: [287, 222, 302, 236] (category: pedestrian)\n- A6: [254, 479, 272, 495] (category: pedestrian)\n- A5: [748, 409, 766, 426] (category: pedestrian)\n- A4: [222, 90, 247, 104] (category: pedestrian)\n- A2: [58, 530, 86, 546] (category: pedestrian)\n- A1: [330, 448, 343, 463] (category: pedestrian)\n\nAnswer options:\nA: G1->A2, G2->A3, G3->A5, G4->A4\nB: G1->A4, G2->A3, G3->A2, G4->A6\nC: G1->A1, G2->A2, G3->A5, G4->A3\nD: G1->A1, G2->A5, G3->A4, G4->A6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ee43fb8141aa4576e3ebdf22ef02f663d6e11cac7f7e49ef76189736638c9d2a.jpg", "reference_inputs/images/5503db5f5ae4225d2c6f3add36d2f86fa8ee0bd3d0e0cce90ad74ba26b5a2990.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000014", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G4, G1, G5, G6, G3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [286, 197, 305, 205] (category: pedestrian)\n- A2: [736, 230, 757, 239] (category: pedestrian)\n- A3: [188, 251, 210, 262] (category: pedestrian)\n- A4: [386, 266, 418, 333] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [304, 264, 312, 294] (category: pedestrian)\n- G4: [744, 262, 754, 294] (category: pedestrian)\n- G1: [359, 275, 406, 300] (category: vehicle)\n- G5: [634, 269, 639, 291] (category: pedestrian)\n- G6: [178, 262, 188, 294] (category: pedestrian)\n- G3: [797, 254, 813, 299] (category: pedestrian)\n\nAnswer options:\nA: A1->G4, A2->G2, A3->G1, A4->G3\nB: A1->G1, A2->G6, A3->G5, A4->G2\nC: A1->G2, A2->G4, A3->G6, A4->G1\nD: A1->G4, A2->G6, A3->G1, A4->G2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/01bec1e6063f2abdaa5a2a82f0cc79873e8527ba96cf4d200ccc3916f0c1bc18.jpg", "reference_inputs/images/502fe2b7bf5d1fe5c44e96178c57cdc8f6cda7fad2cf013c6853ede35a442838.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000015", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A5, A3, A1, A4, A2, A6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [359, 275, 406, 300] (category: vehicle)\n- G2: [304, 264, 312, 294] (category: pedestrian)\n- G3: [178, 262, 188, 294] (category: pedestrian)\n- G4: [744, 262, 754, 294] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A5: [693, 401, 714, 410] (category: pedestrian)\n- A3: [506, 0, 536, 51] (category: vehicle)\n- A1: [736, 230, 757, 239] (category: pedestrian)\n- A4: [386, 266, 418, 333] (category: vehicle)\n- A2: [188, 251, 210, 262] (category: pedestrian)\n- A6: [286, 197, 305, 205] (category: pedestrian)\n\nAnswer options:\nA: G1->A4, G2->A6, G3->A2, G4->A1\nB: G1->A1, G2->A4, G3->A5, G4->A2\nC: G1->A2, G2->A6, G3->A1, G4->A5\nD: G1->A2, G2->A5, G3->A3, G4->A4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/502fe2b7bf5d1fe5c44e96178c57cdc8f6cda7fad2cf013c6853ede35a442838.jpg", "reference_inputs/images/01bec1e6063f2abdaa5a2a82f0cc79873e8527ba96cf4d200ccc3916f0c1bc18.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000018", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G3, G1, G4, G6, G2, G5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [280, 187, 295, 198] (category: pedestrian)\n- A2: [417, 356, 455, 430] (category: vehicle)\n- A3: [670, 374, 683, 385] (category: pedestrian)\n- A4: [283, 398, 299, 411] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G3: [792, 272, 799, 298] (category: pedestrian)\n- G1: [468, 274, 473, 294] (category: pedestrian)\n- G4: [454, 268, 461, 295] (category: pedestrian)\n- G6: [569, 283, 606, 300] (category: vehicle)\n- G2: [478, 274, 482, 293] (category: pedestrian)\n- G5: [461, 274, 463, 293] (category: pedestrian)\n\nAnswer options:\nA: A1->G1, A2->G6, A3->G3, A4->G4\nB: A1->G3, A2->G2, A3->G1, A4->G4\nC: A1->G1, A2->G6, A3->G3, A4->G2\nD: A1->G3, A2->G6, A3->G2, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b4d0eda699256237ed11a4a95c83ee5937f0838e00ca027a698437dc64ed5d94.jpg", "reference_inputs/images/47f9e862659644a60fa5865a905e3616ae64a6438419abb7aae6d257478f5f7c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000019", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A6, A2, A3, A4, A1, A5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [468, 274, 473, 294] (category: pedestrian)\n- G2: [792, 272, 799, 298] (category: pedestrian)\n- G3: [569, 283, 606, 300] (category: vehicle)\n- G4: [454, 268, 461, 295] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A6: [417, 356, 455, 430] (category: vehicle)\n- A2: [280, 187, 295, 198] (category: pedestrian)\n- A3: [670, 374, 683, 385] (category: pedestrian)\n- A4: [283, 398, 299, 411] (category: pedestrian)\n- A1: [258, 125, 278, 138] (category: pedestrian)\n- A5: [239, 374, 256, 384] (category: pedestrian)\n\nAnswer options:\nA: G1->A5, G2->A6, G3->A1, G4->A3\nB: G1->A3, G2->A6, G3->A5, G4->A2\nC: G1->A6, G2->A3, G3->A2, G4->A1\nD: G1->A2, G2->A3, G3->A6, G4->A4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/47f9e862659644a60fa5865a905e3616ae64a6438419abb7aae6d257478f5f7c.jpg", "reference_inputs/images/b4d0eda699256237ed11a4a95c83ee5937f0838e00ca027a698437dc64ed5d94.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000020", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G3, G4, G2, G1, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [790, 392, 801, 407] (category: vehicle)\n- A2: [723, 475, 738, 490] (category: vehicle)\n- A3: [802, 566, 837, 575] (category: vehicle)\n- A4: [801, 454, 812, 463] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [522, 274, 534, 288] (category: vehicle)\n- G3: [507, 277, 526, 291] (category: vehicle)\n- G4: [436, 282, 453, 292] (category: vehicle)\n- G2: [484, 276, 495, 287] (category: vehicle)\n- G1: [550, 224, 630, 302] (category: vehicle)\n- G6: [0, 253, 105, 342] (category: vehicle)\n\nAnswer options:\nA: A1->G6, A2->G1, A3->G4, A4->G2\nB: A1->G5, A2->G4, A3->G1, A4->G6\nC: A1->G1, A2->G3, A3->G5, A4->G6\nD: A1->G4, A2->G2, A3->G3, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/12ba627932e572d5a61af9cd22f432b21d00b9f8a92045064ca84ac488070f7c.jpg", "reference_inputs/images/068981882776608705076f4c4c3e998b66611be3d5507c1608daae3a21fad37c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000021", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A1, A6, A4, A3, A5, A2. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [522, 274, 534, 288] (category: vehicle)\n- G2: [484, 276, 495, 287] (category: vehicle)\n- G3: [436, 282, 453, 292] (category: vehicle)\n- G4: [450, 275, 478, 296] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A1: [723, 475, 738, 490] (category: vehicle)\n- A6: [732, 439, 744, 454] (category: vehicle)\n- A4: [750, 374, 762, 387] (category: vehicle)\n- A3: [741, 525, 758, 547] (category: vehicle)\n- A5: [776, 477, 790, 492] (category: vehicle)\n- A2: [790, 392, 801, 407] (category: vehicle)\n\nAnswer options:\nA: G1->A3, G2->A5, G3->A2, G4->A6\nB: G1->A2, G2->A4, G3->A1, G4->A3\nC: G1->A1, G2->A3, G3->A6, G4->A2\nD: G1->A2, G2->A6, G3->A3, G4->A5\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/068981882776608705076f4c4c3e998b66611be3d5507c1608daae3a21fad37c.jpg", "reference_inputs/images/12ba627932e572d5a61af9cd22f432b21d00b9f8a92045064ca84ac488070f7c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000022", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G3, G6, G2, G1, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [235, 70, 250, 99] (category: pedestrian)\n- A2: [674, 340, 686, 355] (category: pedestrian)\n- A3: [383, 94, 419, 168] (category: vehicle)\n- A4: [548, 499, 581, 575] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [667, 266, 674, 294] (category: pedestrian)\n- G3: [423, 274, 450, 294] (category: vehicle)\n- G6: [626, 270, 630, 291] (category: pedestrian)\n- G2: [333, 270, 350, 292] (category: pedestrian)\n- G1: [564, 268, 621, 305] (category: vehicle)\n- G4: [318, 268, 322, 292] (category: pedestrian)\n\nAnswer options:\nA: A1->G6, A2->G3, A3->G1, A4->G2\nB: A1->G2, A2->G5, A3->G3, A4->G1\nC: A1->G1, A2->G3, A3->G4, A4->G5\nD: A1->G2, A2->G6, A3->G4, A4->G5\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2942222141645b774a4b3bd6e5b7422af344083fdde6423f2b414362c1d9f968.jpg", "reference_inputs/images/75bb9f3ac1921d7fed1cce6bc9bbfa9956c3988c4fcf366d0c757f6c48b81399.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000023", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G2, G1, G4, G3, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [680, 108, 696, 121] (category: pedestrian)\n- A2: [674, 261, 687, 268] (category: pedestrian)\n- A3: [654, 279, 666, 288] (category: pedestrian)\n- A4: [280, 434, 300, 449] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [220, 256, 226, 297] (category: pedestrian)\n- G2: [659, 267, 667, 293] (category: pedestrian)\n- G1: [635, 270, 643, 291] (category: pedestrian)\n- G4: [449, 198, 534, 308] (category: vehicle)\n- G3: [334, 271, 338, 291] (category: pedestrian)\n- G6: [647, 265, 654, 294] (category: pedestrian)\n\nAnswer options:\nA: A1->G1, A2->G2, A3->G6, A4->G5\nB: A1->G3, A2->G4, A3->G5, A4->G2\nC: A1->G1, A2->G6, A3->G3, A4->G5\nD: A1->G3, A2->G5, A3->G4, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b54de1266305b9b20f5b628edb6cc3be8bf303f5eaaa4ba4235d674b186ebab7.jpg", "reference_inputs/images/4183a0b6de334d0e851eef8ba64a5005f51b443781882f4bdd69e0315019eef6.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000024", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000024", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A5, A3, A1, A6, A2, A4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [220, 256, 226, 297] (category: pedestrian)\n- G2: [449, 198, 534, 308] (category: vehicle)\n- G3: [647, 265, 654, 294] (category: pedestrian)\n- G4: [635, 270, 643, 291] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A5: [234, 54, 254, 71] (category: pedestrian)\n- A3: [474, 458, 522, 575] (category: vehicle)\n- A1: [674, 261, 687, 268] (category: pedestrian)\n- A6: [680, 108, 696, 121] (category: pedestrian)\n- A2: [654, 279, 666, 288] (category: pedestrian)\n- A4: [280, 434, 300, 449] (category: pedestrian)\n\nAnswer options:\nA: G1->A4, G2->A5, G3->A3, G4->A2\nB: G1->A1, G2->A4, G3->A5, G4->A6\nC: G1->A2, G2->A3, G3->A1, G4->A4\nD: G1->A4, G2->A3, G3->A2, G4->A6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4183a0b6de334d0e851eef8ba64a5005f51b443781882f4bdd69e0315019eef6.jpg", "reference_inputs/images/b54de1266305b9b20f5b628edb6cc3be8bf303f5eaaa4ba4235d674b186ebab7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000025", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G1, G6, G2, G4, G5, G3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [530, 466, 543, 477] (category: vehicle)\n- A2: [510, 442, 520, 451] (category: vehicle)\n- A3: [462, 476, 477, 488] (category: vehicle)\n- A4: [534, 503, 549, 521] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G1: [508, 275, 535, 296] (category: vehicle)\n- G6: [577, 277, 637, 302] (category: vehicle)\n- G2: [502, 262, 527, 291] (category: vehicle)\n- G4: [686, 202, 1000, 350] (category: vehicle)\n- G5: [443, 269, 480, 298] (category: vehicle)\n- G3: [274, 275, 367, 307] (category: vehicle)\n\nAnswer options:\nA: A1->G1, A2->G2, A3->G5, A4->G6\nB: A1->G5, A2->G4, A3->G1, A4->G3\nC: A1->G3, A2->G6, A3->G2, A4->G5\nD: A1->G6, A2->G1, A3->G3, A4->G4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4e33cdd8ea49b40c1dc79aa8a47c3247f0846a5bfc6bf23c11356d4b495d047c.jpg", "reference_inputs/images/e8f86a144ff07fab1b6ed4595caeffbf95d4782f342c62d61c804499bb0065bf.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000026", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A4, A2, A5, A3, A6, A1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [686, 202, 1000, 350] (category: vehicle)\n- G2: [274, 275, 367, 307] (category: vehicle)\n- G3: [577, 277, 637, 302] (category: vehicle)\n- G4: [508, 275, 535, 296] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A4: [490, 446, 501, 461] (category: vehicle)\n- A2: [449, 530, 465, 553] (category: vehicle)\n- A5: [530, 466, 543, 477] (category: vehicle)\n- A3: [462, 476, 477, 488] (category: vehicle)\n- A6: [510, 442, 520, 451] (category: vehicle)\n- A1: [534, 503, 549, 521] (category: vehicle)\n\nAnswer options:\nA: G1->A3, G2->A4, G3->A1, G4->A6\nB: G1->A3, G2->A6, G3->A2, G4->A5\nC: G1->A1, G2->A3, G3->A5, G4->A6\nD: G1->A3, G2->A4, G3->A5, G4->A2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e8f86a144ff07fab1b6ed4595caeffbf95d4782f342c62d61c804499bb0065bf.jpg", "reference_inputs/images/4e33cdd8ea49b40c1dc79aa8a47c3247f0846a5bfc6bf23c11356d4b495d047c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000027", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G6, G5, G2, G1, G3, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [1, 180, 33, 191] (category: pedestrian)\n- A2: [748, 237, 770, 248] (category: pedestrian)\n- A3: [448, 466, 478, 534] (category: vehicle)\n- A4: [274, 149, 297, 159] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G6: [26, 263, 38, 294] (category: pedestrian)\n- G5: [342, 270, 347, 282] (category: pedestrian)\n- G2: [334, 265, 435, 320] (category: vehicle)\n- G1: [295, 267, 306, 294] (category: pedestrian)\n- G3: [626, 272, 631, 290] (category: pedestrian)\n- G4: [774, 261, 782, 295] (category: pedestrian)\n\nAnswer options:\nA: A1->G5, A2->G2, A3->G4, A4->G6\nB: A1->G3, A2->G1, A3->G5, A4->G2\nC: A1->G4, A2->G1, A3->G3, A4->G2\nD: A1->G6, A2->G4, A3->G2, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6e4af1e6834a23f481da3fb5481f4f060494ac2b58b1035dff9f3ad5ad3f2661.jpg", "reference_inputs/images/1f4e5bfc1dc07dd2fc6ab971db43666d4bee8f303a3194ecb836bf23599c8aa9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000028", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A1, A2, A4, A6, A3, A5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [334, 265, 435, 320] (category: vehicle)\n- G2: [295, 267, 306, 294] (category: pedestrian)\n- G3: [26, 263, 38, 294] (category: pedestrian)\n- G4: [774, 261, 782, 295] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A1: [1, 180, 33, 191] (category: pedestrian)\n- A2: [274, 149, 297, 159] (category: pedestrian)\n- A4: [273, 0, 296, 15] (category: pedestrian)\n- A6: [448, 466, 478, 534] (category: vehicle)\n- A3: [1010, 267, 1023, 272] (category: pedestrian)\n- A5: [748, 237, 770, 248] (category: pedestrian)\n\nAnswer options:\nA: G1->A1, G2->A2, G3->A5, G4->A3\nB: G1->A6, G2->A2, G3->A5, G4->A3\nC: G1->A6, G2->A2, G3->A1, G4->A5\nD: G1->A5, G2->A6, G3->A1, G4->A3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1f4e5bfc1dc07dd2fc6ab971db43666d4bee8f303a3194ecb836bf23599c8aa9.jpg", "reference_inputs/images/6e4af1e6834a23f481da3fb5481f4f060494ac2b58b1035dff9f3ad5ad3f2661.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000029", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G4, G3, G2, G5, G1, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [822, 108, 848, 122] (category: pedestrian)\n- A2: [910, 325, 982, 358] (category: vehicle)\n- A3: [673, 50, 691, 64] (category: pedestrian)\n- A4: [750, 0, 766, 15] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G4: [432, 273, 438, 294] (category: pedestrian)\n- G3: [354, 251, 371, 302] (category: pedestrian)\n- G2: [70, 238, 99, 305] (category: pedestrian)\n- G5: [27, 225, 74, 314] (category: pedestrian)\n- G1: [622, 279, 692, 304] (category: vehicle)\n- G6: [438, 274, 442, 293] (category: pedestrian)\n\nAnswer options:\nA: A1->G1, A2->G6, A3->G5, A4->G4\nB: A1->G5, A2->G4, A3->G1, A4->G2\nC: A1->G3, A2->G1, A3->G5, A4->G2\nD: A1->G4, A2->G2, A3->G5, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d23def54bc38c6919f30f1cf05d51e32af4daf8d2413d5881837e1e5da81ad0b.jpg", "reference_inputs/images/eb97a3246d862464d2b51b720f73d209b8d3a46bb24e54d9e5a2b762b5dfa240.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000030", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A6, A2, A1, A5, A3, A4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [622, 279, 692, 304] (category: vehicle)\n- G2: [27, 225, 74, 314] (category: pedestrian)\n- G3: [70, 238, 99, 305] (category: pedestrian)\n- G4: [354, 251, 371, 302] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A6: [822, 108, 848, 122] (category: pedestrian)\n- A2: [673, 50, 691, 64] (category: pedestrian)\n- A1: [910, 325, 982, 358] (category: vehicle)\n- A5: [750, 0, 766, 15] (category: pedestrian)\n- A3: [329, 28, 342, 47] (category: pedestrian)\n- A4: [252, 61, 270, 78] (category: pedestrian)\n\nAnswer options:\nA: G1->A1, G2->A2, G3->A5, G4->A6\nB: G1->A6, G2->A3, G3->A2, G4->A4\nC: G1->A2, G2->A1, G3->A6, G4->A5\nD: G1->A2, G2->A4, G3->A5, G4->A1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/eb97a3246d862464d2b51b720f73d209b8d3a46bb24e54d9e5a2b762b5dfa240.jpg", "reference_inputs/images/d23def54bc38c6919f30f1cf05d51e32af4daf8d2413d5881837e1e5da81ad0b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000031", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000031", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G5, G6, G3, G1, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [222, 306, 239, 316] (category: pedestrian)\n- A2: [506, 159, 538, 227] (category: vehicle)\n- A3: [554, 498, 608, 575] (category: vehicle)\n- A4: [288, 430, 302, 446] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [506, 276, 540, 299] (category: vehicle)\n- G5: [2, 242, 33, 307] (category: pedestrian)\n- G6: [466, 279, 498, 294] (category: vehicle)\n- G3: [98, 253, 125, 298] (category: pedestrian)\n- G1: [299, 222, 414, 305] (category: vehicle)\n- G4: [447, 282, 466, 291] (category: vehicle)\n\nAnswer options:\nA: A1->G3, A2->G2, A3->G1, A4->G5\nB: A1->G4, A2->G6, A3->G3, A4->G1\nC: A1->G2, A2->G3, A3->G1, A4->G4\nD: A1->G4, A2->G6, A3->G2, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fa77697dfc4085fced112433ab375a16fb84c526f445ccec94d4d1edbce39c59.jpg", "reference_inputs/images/5a50298408b5a44380b325f06a63c9e6f92a85c81613f17465ad2b8e5b4836a3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000032", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A2, A1, A6, A4, A5, A3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [506, 276, 540, 299] (category: vehicle)\n- G2: [299, 222, 414, 305] (category: vehicle)\n- G3: [2, 242, 33, 307] (category: pedestrian)\n- G4: [98, 253, 125, 298] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A2: [506, 159, 538, 227] (category: vehicle)\n- A1: [288, 430, 302, 446] (category: pedestrian)\n- A6: [222, 306, 239, 316] (category: pedestrian)\n- A4: [303, 566, 326, 575] (category: pedestrian)\n- A5: [302, 142, 318, 158] (category: pedestrian)\n- A3: [554, 498, 608, 575] (category: vehicle)\n\nAnswer options:\nA: G1->A2, G2->A3, G3->A1, G4->A6\nB: G1->A4, G2->A6, G3->A3, G4->A2\nC: G1->A3, G2->A4, G3->A2, G4->A1\nD: G1->A2, G2->A3, G3->A6, G4->A1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5a50298408b5a44380b325f06a63c9e6f92a85c81613f17465ad2b8e5b4836a3.jpg", "reference_inputs/images/fa77697dfc4085fced112433ab375a16fb84c526f445ccec94d4d1edbce39c59.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000033", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000033", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G4, G6, G2, G1, G3, G5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [273, 251, 289, 266] (category: pedestrian)\n- A2: [214, 92, 242, 104] (category: pedestrian)\n- A3: [234, 31, 254, 50] (category: pedestrian)\n- A4: [249, 66, 269, 80] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G4: [630, 247, 649, 306] (category: pedestrian)\n- G6: [703, 250, 716, 305] (category: pedestrian)\n- G2: [892, 246, 914, 304] (category: pedestrian)\n- G1: [232, 240, 250, 309] (category: pedestrian)\n- G3: [690, 245, 706, 306] (category: pedestrian)\n- G5: [572, 252, 589, 304] (category: pedestrian)\n\nAnswer options:\nA: A1->G6, A2->G3, A3->G4, A4->G1\nB: A1->G1, A2->G5, A3->G6, A4->G2\nC: A1->G6, A2->G4, A3->G2, A4->G5\nD: A1->G1, A2->G5, A3->G3, A4->G4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e4457e1a53e081868651ab4f5bdb7954e2e6197ba1a7b0805272ee9e12908fcd.jpg", "reference_inputs/images/fb65f7995f334114946a52d3172118daba99133ed34921956d2365e3e709e2a6.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000034", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000034", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A1, A4, A6, A5, A2, A3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [572, 252, 589, 304] (category: pedestrian)\n- G2: [232, 240, 250, 309] (category: pedestrian)\n- G3: [630, 247, 649, 306] (category: pedestrian)\n- G4: [690, 245, 706, 306] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A1: [691, 227, 705, 238] (category: pedestrian)\n- A4: [234, 31, 254, 50] (category: pedestrian)\n- A6: [214, 92, 242, 104] (category: pedestrian)\n- A5: [249, 66, 269, 80] (category: pedestrian)\n- A2: [651, 294, 663, 303] (category: pedestrian)\n- A3: [273, 251, 289, 266] (category: pedestrian)\n\nAnswer options:\nA: G1->A4, G2->A5, G3->A6, G4->A3\nB: G1->A1, G2->A3, G3->A5, G4->A2\nC: G1->A3, G2->A5, G3->A4, G4->A1\nD: G1->A6, G2->A3, G3->A5, G4->A4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fb65f7995f334114946a52d3172118daba99133ed34921956d2365e3e709e2a6.jpg", "reference_inputs/images/e4457e1a53e081868651ab4f5bdb7954e2e6197ba1a7b0805272ee9e12908fcd.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000035", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000035", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G1, G5, G3, G6, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [384, 58, 414, 134] (category: vehicle)\n- A2: [285, 63, 300, 77] (category: pedestrian)\n- A3: [298, 210, 313, 220] (category: pedestrian)\n- A4: [244, 234, 261, 241] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [930, 272, 938, 292] (category: pedestrian)\n- G1: [815, 274, 867, 294] (category: vehicle)\n- G5: [218, 233, 246, 314] (category: pedestrian)\n- G3: [370, 234, 734, 354] (category: vehicle)\n- G6: [239, 248, 254, 308] (category: pedestrian)\n- G4: [574, 241, 594, 268] (category: pedestrian)\n\nAnswer options:\nA: A1->G1, A2->G4, A3->G5, A4->G6\nB: A1->G3, A2->G4, A3->G1, A4->G2\nC: A1->G6, A2->G1, A3->G3, A4->G2\nD: A1->G3, A2->G4, A3->G5, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/223cbee8041d6f1a55407a61b86a321c281765434511cd82b74d02a28bad5c15.jpg", "reference_inputs/images/5c64e995683b1de21c283c671349dac82351d1b3a64375df3ccdb6bffaedb802.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000036", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000036", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A6, A5, A4, A3, A2, A1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [574, 241, 594, 268] (category: pedestrian)\n- G2: [218, 233, 246, 314] (category: pedestrian)\n- G3: [239, 248, 254, 308] (category: pedestrian)\n- G4: [370, 234, 734, 354] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A6: [244, 234, 261, 241] (category: pedestrian)\n- A5: [502, 0, 538, 50] (category: vehicle)\n- A4: [298, 210, 313, 220] (category: pedestrian)\n- A3: [562, 245, 596, 317] (category: vehicle)\n- A2: [285, 63, 300, 77] (category: pedestrian)\n- A1: [384, 58, 414, 134] (category: vehicle)\n\nAnswer options:\nA: G1->A2, G2->A4, G3->A6, G4->A1\nB: G1->A2, G2->A4, G3->A3, G4->A1\nC: G1->A3, G2->A5, G3->A4, G4->A6\nD: G1->A1, G2->A6, G3->A3, G4->A5\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5c64e995683b1de21c283c671349dac82351d1b3a64375df3ccdb6bffaedb802.jpg", "reference_inputs/images/223cbee8041d6f1a55407a61b86a321c281765434511cd82b74d02a28bad5c15.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000037", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000037", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G2, G3, G6, G4, G1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [0, 461, 66, 500] (category: vehicle)\n- A2: [38, 148, 69, 166] (category: pedestrian)\n- A3: [239, 158, 262, 170] (category: pedestrian)\n- A4: [42, 540, 72, 564] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [0, 258, 13, 305] (category: pedestrian)\n- G2: [646, 242, 668, 306] (category: pedestrian)\n- G3: [411, 258, 455, 298] (category: vehicle)\n- G6: [222, 283, 290, 303] (category: vehicle)\n- G4: [146, 266, 166, 301] (category: pedestrian)\n- G1: [598, 263, 615, 298] (category: pedestrian)\n\nAnswer options:\nA: A1->G2, A2->G5, A3->G4, A4->G3\nB: A1->G6, A2->G1, A3->G2, A4->G4\nC: A1->G5, A2->G6, A3->G4, A4->G2\nD: A1->G4, A2->G6, A3->G1, A4->G5\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0fef7bf43103cdecd4e2c7ba577149ffb74ee952c482f428f478d115ae197001.jpg", "reference_inputs/images/a6bbb87b287f5401a21be4253d6a9ba6d1b1fb7dd5a4446eec287e86133e81c1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000038", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000038", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A4, A2, A1, A6, A5, A3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [222, 283, 290, 303] (category: vehicle)\n- G2: [598, 263, 615, 298] (category: pedestrian)\n- G3: [646, 242, 668, 306] (category: pedestrian)\n- G4: [146, 266, 166, 301] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A4: [490, 458, 562, 497] (category: vehicle)\n- A2: [239, 158, 262, 170] (category: pedestrian)\n- A1: [0, 461, 66, 500] (category: vehicle)\n- A6: [42, 540, 72, 564] (category: pedestrian)\n- A5: [850, 178, 874, 194] (category: pedestrian)\n- A3: [38, 148, 69, 166] (category: pedestrian)\n\nAnswer options:\nA: G1->A6, G2->A5, G3->A2, G4->A1\nB: G1->A5, G2->A3, G3->A1, G4->A4\nC: G1->A2, G2->A3, G3->A6, G4->A1\nD: G1->A1, G2->A3, G3->A2, G4->A6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a6bbb87b287f5401a21be4253d6a9ba6d1b1fb7dd5a4446eec287e86133e81c1.jpg", "reference_inputs/images/0fef7bf43103cdecd4e2c7ba577149ffb74ee952c482f428f478d115ae197001.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000039", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000039", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G1, G5, G6, G2, G3, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [490, 310, 525, 382] (category: vehicle)\n- A2: [686, 553, 702, 574] (category: pedestrian)\n- A3: [719, 456, 734, 468] (category: pedestrian)\n- A4: [721, 167, 736, 177] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G1: [635, 270, 641, 292] (category: pedestrian)\n- G5: [865, 248, 881, 299] (category: pedestrian)\n- G6: [694, 266, 700, 292] (category: pedestrian)\n- G2: [962, 226, 994, 310] (category: pedestrian)\n- G3: [318, 271, 322, 290] (category: pedestrian)\n- G4: [478, 265, 529, 301] (category: vehicle)\n\nAnswer options:\nA: A1->G6, A2->G3, A3->G4, A4->G5\nB: A1->G3, A2->G4, A3->G1, A4->G2\nC: A1->G4, A2->G1, A3->G6, A4->G3\nD: A1->G4, A2->G2, A3->G5, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cec587f7f71d59f91723be411e06f634b2c38d94eef8cefcfe273dea4be0b7c4.jpg", "reference_inputs/images/214450fcb9ec8f693510b00eb2718a24d336b00136b94d6785fc63ff6db98864.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000040", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000040", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A1, A2, A6, A4, A5, A3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [962, 226, 994, 310] (category: pedestrian)\n- G2: [635, 270, 641, 292] (category: pedestrian)\n- G3: [478, 265, 529, 301] (category: vehicle)\n- G4: [865, 248, 881, 299] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A1: [719, 456, 734, 468] (category: pedestrian)\n- A2: [275, 468, 290, 482] (category: pedestrian)\n- A6: [674, 78, 686, 93] (category: pedestrian)\n- A4: [490, 310, 525, 382] (category: vehicle)\n- A5: [686, 553, 702, 574] (category: pedestrian)\n- A3: [721, 167, 736, 177] (category: pedestrian)\n\nAnswer options:\nA: G1->A1, G2->A6, G3->A3, G4->A5\nB: G1->A2, G2->A3, G3->A6, G4->A4\nC: G1->A4, G2->A2, G3->A3, G4->A6\nD: G1->A5, G2->A6, G3->A4, G4->A1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/214450fcb9ec8f693510b00eb2718a24d336b00136b94d6785fc63ff6db98864.jpg", "reference_inputs/images/cec587f7f71d59f91723be411e06f634b2c38d94eef8cefcfe273dea4be0b7c4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000041", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000041", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G4, G2, G1, G3, G6, G5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [370, 324, 410, 398] (category: vehicle)\n- A2: [666, 248, 678, 255] (category: pedestrian)\n- A3: [522, 426, 560, 501] (category: vehicle)\n- A4: [649, 350, 662, 360] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G4: [789, 268, 819, 281] (category: vehicle)\n- G2: [932, 269, 986, 284] (category: vehicle)\n- G1: [217, 254, 230, 272] (category: pedestrian)\n- G3: [834, 266, 855, 278] (category: vehicle)\n- G6: [704, 254, 710, 278] (category: pedestrian)\n- G5: [722, 258, 730, 276] (category: pedestrian)\n\nAnswer options:\nA: A1->G3, A2->G6, A3->G1, A4->G5\nB: A1->G4, A2->G1, A3->G5, A4->G2\nC: A1->G2, A2->G6, A3->G4, A4->G5\nD: A1->G4, A2->G2, A3->G5, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5dbeddd1c3c7318355a2bef03ca4c4171d13aa2262ccc2af44a86fbae0d714a3.jpg", "reference_inputs/images/db28015dd827c7a87eca54c8bed03ee4194eed92a3b8f6f40966f1f3a8f4b38c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000043", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000043", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G1, G4, G2, G3, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [598, 485, 610, 501] (category: vehicle)\n- A2: [557, 446, 567, 460] (category: vehicle)\n- A3: [576, 372, 586, 385] (category: vehicle)\n- A4: [562, 550, 579, 575] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [500, 264, 544, 301] (category: vehicle)\n- G1: [411, 274, 441, 294] (category: vehicle)\n- G4: [476, 273, 490, 288] (category: vehicle)\n- G2: [0, 210, 334, 399] (category: vehicle)\n- G3: [478, 281, 494, 291] (category: vehicle)\n- G6: [554, 278, 579, 294] (category: vehicle)\n\nAnswer options:\nA: A1->G6, A2->G4, A3->G5, A4->G1\nB: A1->G5, A2->G1, A3->G4, A4->G2\nC: A1->G3, A2->G5, A3->G1, A4->G2\nD: A1->G2, A2->G3, A3->G6, A4->G4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/982b4a63b2edbf1c4acd3dca191ce0fee7ef5a177a699d51862a19ce21f26430.jpg", "reference_inputs/images/876e52177c54d2f1a987ee7d6b60aef790c909d21871bccc19c29ad50f3249e4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000044", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000044", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A1, A3, A6, A2, A5, A4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [500, 264, 544, 301] (category: vehicle)\n- G2: [411, 274, 441, 294] (category: vehicle)\n- G3: [476, 273, 490, 288] (category: vehicle)\n- G4: [0, 210, 334, 399] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A1: [576, 372, 586, 385] (category: vehicle)\n- A3: [608, 389, 619, 403] (category: vehicle)\n- A6: [557, 446, 567, 460] (category: vehicle)\n- A2: [562, 550, 579, 575] (category: vehicle)\n- A5: [546, 489, 562, 505] (category: vehicle)\n- A4: [598, 485, 610, 501] (category: vehicle)\n\nAnswer options:\nA: G1->A3, G2->A6, G3->A2, G4->A4\nB: G1->A5, G2->A3, G3->A6, G4->A2\nC: G1->A4, G2->A6, G3->A1, G4->A2\nD: G1->A1, G2->A2, G3->A5, G4->A6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/876e52177c54d2f1a987ee7d6b60aef790c909d21871bccc19c29ad50f3249e4.jpg", "reference_inputs/images/982b4a63b2edbf1c4acd3dca191ce0fee7ef5a177a699d51862a19ce21f26430.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000045", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000045", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G6, G1, G2, G5, G4, G3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [714, 397, 734, 406] (category: pedestrian)\n- A2: [674, 291, 690, 299] (category: pedestrian)\n- A3: [664, 130, 683, 142] (category: pedestrian)\n- A4: [447, 324, 478, 391] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G6: [986, 234, 1014, 305] (category: pedestrian)\n- G1: [679, 262, 688, 294] (category: pedestrian)\n- G2: [376, 269, 458, 311] (category: vehicle)\n- G5: [762, 256, 795, 298] (category: pedestrian)\n- G4: [502, 278, 518, 290] (category: vehicle)\n- G3: [712, 264, 721, 293] (category: pedestrian)\n\nAnswer options:\nA: A1->G3, A2->G2, A3->G5, A4->G4\nB: A1->G2, A2->G3, A3->G1, A4->G5\nC: A1->G3, A2->G2, A3->G6, A4->G1\nD: A1->G6, A2->G5, A3->G1, A4->G2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/acba93faab2b2b94bd8f121833dcdc0686ddae44c5819cbb8c7ea4fc08b05082.jpg", "reference_inputs/images/51c4243b98d1a3e1b646c00d0570aefc32d478d1128648c52c2f84bc12b8f8ff.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000046", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000046", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A5, A1, A3, A2, A4, A6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [986, 234, 1014, 305] (category: pedestrian)\n- G2: [376, 269, 458, 311] (category: vehicle)\n- G3: [679, 262, 688, 294] (category: pedestrian)\n- G4: [762, 256, 795, 298] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A5: [715, 82, 735, 94] (category: pedestrian)\n- A1: [249, 389, 266, 408] (category: pedestrian)\n- A3: [674, 291, 690, 299] (category: pedestrian)\n- A2: [447, 324, 478, 391] (category: vehicle)\n- A4: [714, 397, 734, 406] (category: pedestrian)\n- A6: [664, 130, 683, 142] (category: pedestrian)\n\nAnswer options:\nA: G1->A3, G2->A4, G3->A5, G4->A6\nB: G1->A2, G2->A1, G3->A5, G4->A3\nC: G1->A4, G2->A2, G3->A6, G4->A3\nD: G1->A4, G2->A2, G3->A6, G4->A1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/51c4243b98d1a3e1b646c00d0570aefc32d478d1128648c52c2f84bc12b8f8ff.jpg", "reference_inputs/images/acba93faab2b2b94bd8f121833dcdc0686ddae44c5819cbb8c7ea4fc08b05082.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000047", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000047", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G3, G4, G5, G1, G6, G2. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [260, 281, 278, 290] (category: pedestrian)\n- A2: [286, 346, 304, 354] (category: pedestrian)\n- A3: [283, 215, 298, 224] (category: pedestrian)\n- A4: [374, 61, 409, 134] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G3: [221, 248, 243, 308] (category: pedestrian)\n- G4: [353, 246, 370, 307] (category: pedestrian)\n- G5: [446, 245, 462, 306] (category: pedestrian)\n- G1: [860, 250, 1023, 318] (category: vehicle)\n- G6: [579, 245, 594, 304] (category: pedestrian)\n- G2: [310, 240, 333, 312] (category: pedestrian)\n\nAnswer options:\nA: A1->G6, A2->G5, A3->G1, A4->G3\nB: A1->G6, A2->G4, A3->G1, A4->G2\nC: A1->G6, A2->G1, A3->G3, A4->G5\nD: A1->G5, A2->G2, A3->G6, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e15b5ab4486d18e49579e63edec4137ea576208c7b3ed60f09684c21f8bbd4b2.jpg", "reference_inputs/images/83de54604647f8550f7370570918454c1f86c25ce7ef1fa73fb95060dbcecc4f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000048", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000048", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A5, A1, A3, A2, A4, A6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [446, 245, 462, 306] (category: pedestrian)\n- G2: [860, 250, 1023, 318] (category: vehicle)\n- G3: [579, 245, 594, 304] (category: pedestrian)\n- G4: [310, 240, 333, 312] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A5: [260, 281, 278, 290] (category: pedestrian)\n- A1: [235, 322, 255, 332] (category: pedestrian)\n- A3: [214, 382, 242, 392] (category: pedestrian)\n- A2: [286, 346, 304, 354] (category: pedestrian)\n- A4: [374, 61, 409, 134] (category: vehicle)\n- A6: [283, 215, 298, 224] (category: pedestrian)\n\nAnswer options:\nA: G1->A5, G2->A4, G3->A6, G4->A2\nB: G1->A3, G2->A4, G3->A1, G4->A5\nC: G1->A1, G2->A4, G3->A3, G4->A2\nD: G1->A4, G2->A5, G3->A1, G4->A2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/83de54604647f8550f7370570918454c1f86c25ce7ef1fa73fb95060dbcecc4f.jpg", "reference_inputs/images/e15b5ab4486d18e49579e63edec4137ea576208c7b3ed60f09684c21f8bbd4b2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000051", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000051", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G6, G3, G1, G4, G5, G2. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [594, 529, 609, 547] (category: vehicle)\n- A2: [550, 386, 558, 402] (category: vehicle)\n- A3: [600, 464, 611, 475] (category: vehicle)\n- A4: [539, 456, 550, 472] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G6: [293, 269, 310, 292] (category: pedestrian)\n- G3: [750, 258, 764, 298] (category: pedestrian)\n- G1: [530, 276, 562, 298] (category: vehicle)\n- G4: [459, 276, 482, 291] (category: vehicle)\n- G5: [562, 278, 579, 292] (category: vehicle)\n- G2: [504, 273, 518, 288] (category: vehicle)\n\nAnswer options:\nA: A1->G6, A2->G4, A3->G2, A4->G1\nB: A1->G4, A2->G6, A3->G1, A4->G2\nC: A1->G5, A2->G1, A3->G3, A4->G2\nD: A1->G1, A2->G2, A3->G5, A4->G4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f836ca059921239eb42024ba8423dd2007a5475b13086e3b198e74a73afe9fb5.jpg", "reference_inputs/images/6735f98cc46d273f973863ea4bee7b2468bdd114361efff6508e8164bd3036ee.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000052", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000052", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G1, G5, G3, G4, G2, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [462, 470, 474, 487] (category: vehicle)\n- A2: [446, 529, 466, 551] (category: vehicle)\n- A3: [487, 378, 495, 393] (category: vehicle)\n- A4: [508, 518, 523, 539] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G1: [500, 268, 538, 298] (category: vehicle)\n- G5: [482, 281, 495, 290] (category: vehicle)\n- G3: [549, 279, 571, 293] (category: vehicle)\n- G4: [422, 275, 449, 293] (category: vehicle)\n- G2: [407, 282, 414, 298] (category: vehicle)\n- G6: [479, 274, 491, 288] (category: vehicle)\n\nAnswer options:\nA: A1->G5, A2->G1, A3->G6, A4->G3\nB: A1->G6, A2->G3, A3->G1, A4->G5\nC: A1->G5, A2->G4, A3->G6, A4->G1\nD: A1->G4, A2->G2, A3->G6, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/dd4850896eac74d7cb7eb6106ed750d6f3f7a800b678e68b687900ca4ba4251e.jpg", "reference_inputs/images/9ad8a50cccc69d4d19587748ca75a29e617773e0bff87ecdaec8ddeb22d583f1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000053", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000053", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G4, G2, G1, G5, G3, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [699, 550, 738, 575] (category: vehicle)\n- A2: [746, 198, 770, 209] (category: pedestrian)\n- A3: [766, 298, 785, 308] (category: pedestrian)\n- A4: [545, 130, 597, 198] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G4: [301, 278, 334, 296] (category: vehicle)\n- G2: [530, 274, 534, 290] (category: pedestrian)\n- G1: [434, 273, 450, 288] (category: vehicle)\n- G5: [270, 270, 274, 290] (category: pedestrian)\n- G3: [220, 270, 226, 291] (category: pedestrian)\n- G6: [391, 280, 406, 290] (category: vehicle)\n\nAnswer options:\nA: A1->G5, A2->G1, A3->G4, A4->G2\nB: A1->G6, A2->G3, A3->G5, A4->G4\nC: A1->G5, A2->G3, A3->G1, A4->G2\nD: A1->G5, A2->G3, A3->G6, A4->G4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e4d2d3e636b4d3b3ff2b02938c80dc6fca66ffd777ec8927b57b34db548cdb62.jpg", "reference_inputs/images/474cb395aa9651abbdae88daf92c2b5ade56e19b9392b0a17f662fedcfd84053.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000054", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000054", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A2, A4, A1, A6, A3, A5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [391, 280, 406, 290] (category: vehicle)\n- G2: [301, 278, 334, 296] (category: vehicle)\n- G3: [220, 270, 226, 291] (category: pedestrian)\n- G4: [270, 270, 274, 290] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A2: [746, 198, 770, 209] (category: pedestrian)\n- A4: [882, 549, 907, 572] (category: pedestrian)\n- A1: [545, 130, 597, 198] (category: vehicle)\n- A6: [417, 517, 435, 538] (category: pedestrian)\n- A3: [766, 298, 785, 308] (category: pedestrian)\n- A5: [699, 550, 738, 575] (category: vehicle)\n\nAnswer options:\nA: G1->A5, G2->A1, G3->A2, G4->A3\nB: G1->A1, G2->A6, G3->A3, G4->A4\nC: G1->A1, G2->A3, G3->A5, G4->A6\nD: G1->A5, G2->A6, G3->A4, G4->A2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/474cb395aa9651abbdae88daf92c2b5ade56e19b9392b0a17f662fedcfd84053.jpg", "reference_inputs/images/e4d2d3e636b4d3b3ff2b02938c80dc6fca66ffd777ec8927b57b34db548cdb62.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000056", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000056", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A2, A5, A1, A4, A6, A3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [27, 239, 138, 299] (category: vehicle)\n- G2: [242, 276, 274, 297] (category: vehicle)\n- G3: [291, 284, 310, 292] (category: vehicle)\n- G4: [318, 278, 349, 294] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A2: [510, 494, 523, 510] (category: vehicle)\n- A5: [482, 505, 498, 528] (category: vehicle)\n- A1: [537, 539, 557, 559] (category: vehicle)\n- A4: [466, 450, 477, 462] (category: vehicle)\n- A6: [439, 559, 462, 575] (category: vehicle)\n- A3: [507, 434, 522, 457] (category: vehicle)\n\nAnswer options:\nA: G1->A3, G2->A4, G3->A1, G4->A5\nB: G1->A6, G2->A1, G3->A4, G4->A5\nC: G1->A4, G2->A5, G3->A3, G4->A6\nD: G1->A4, G2->A2, G3->A6, G4->A1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1be0d45fc5eb88503ea3915f0b8fe008a69f91db0a9f5307220b605630c8732b.jpg", "reference_inputs/images/9a4b8a0ef75644fc6326ef5f14e66451e3488ddd11d148c7e81dea717a438b9b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000058", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000058", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A4, A3, A2, A6, A1, A5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [276, 265, 288, 299] (category: pedestrian)\n- G2: [554, 231, 628, 303] (category: vehicle)\n- G3: [866, 249, 886, 302] (category: pedestrian)\n- G4: [42, 242, 74, 314] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A4: [292, 450, 308, 465] (category: pedestrian)\n- A3: [19, 474, 50, 491] (category: pedestrian)\n- A2: [0, 172, 130, 221] (category: vehicle)\n- A6: [186, 34, 209, 52] (category: pedestrian)\n- A1: [511, 428, 530, 439] (category: pedestrian)\n- A5: [302, 513, 317, 527] (category: pedestrian)\n\nAnswer options:\nA: G1->A1, G2->A4, G3->A6, G4->A3\nB: G1->A3, G2->A2, G3->A6, G4->A4\nC: G1->A1, G2->A2, G3->A6, G4->A4\nD: G1->A6, G2->A2, G3->A1, G4->A3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ca8487ff33e0062cfeb68ed777e96a88e0c9431a893d7a94f47e9934bcac648e.jpg", "reference_inputs/images/78c4d62d677825deee57612396c5c7a9551bb4faba10dc9c14391fb1ab84f3b3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000060", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000060", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A4, A2, A3, A5, A6, A1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [938, 235, 988, 310] (category: pedestrian)\n- G2: [178, 258, 485, 379] (category: vehicle)\n- G3: [741, 234, 754, 273] (category: pedestrian)\n- G4: [610, 241, 632, 307] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A4: [748, 409, 766, 426] (category: pedestrian)\n- A2: [302, 360, 324, 367] (category: pedestrian)\n- A3: [254, 479, 272, 495] (category: pedestrian)\n- A5: [330, 448, 343, 463] (category: pedestrian)\n- A6: [287, 222, 302, 236] (category: pedestrian)\n- A1: [386, 540, 446, 575] (category: vehicle)\n\nAnswer options:\nA: G1->A1, G2->A3, G3->A6, G4->A2\nB: G1->A2, G2->A1, G3->A5, G4->A3\nC: G1->A2, G2->A6, G3->A1, G4->A5\nD: G1->A6, G2->A1, G3->A2, G4->A5\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6de9c3585cf47e11898d5f77222fa22a23cf12d738d4e6bd2252109ea1039268.jpg", "reference_inputs/images/5503db5f5ae4225d2c6f3add36d2f86fa8ee0bd3d0e0cce90ad74ba26b5a2990.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000061", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000061", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G1, G6, G4, G2, G5, G3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [510, 498, 522, 514] (category: vehicle)\n- A2: [458, 493, 474, 510] (category: vehicle)\n- A3: [530, 458, 541, 470] (category: vehicle)\n- A4: [487, 372, 497, 385] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G1: [580, 265, 616, 300] (category: vehicle)\n- G6: [491, 260, 600, 330] (category: vehicle)\n- G4: [115, 269, 306, 323] (category: vehicle)\n- G2: [400, 275, 432, 295] (category: vehicle)\n- G5: [450, 282, 482, 290] (category: vehicle)\n- G3: [473, 272, 487, 288] (category: vehicle)\n\nAnswer options:\nA: A1->G4, A2->G3, A3->G5, A4->G1\nB: A1->G6, A2->G4, A3->G1, A4->G3\nC: A1->G6, A2->G1, A3->G5, A4->G4\nD: A1->G2, A2->G4, A3->G1, A4->G5\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d244fd3db364087e2ec3d7b95f0825a29b0801934b47170eb9f4cba1fe921459.jpg", "reference_inputs/images/b719686ac3ce75d89888dbdd46c328c0141d8802f7ef17467032f5f9c3aea3c0.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000062", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000062", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A2, A4, A1, A6, A5, A3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [473, 272, 487, 288] (category: vehicle)\n- G2: [491, 260, 600, 330] (category: vehicle)\n- G3: [580, 265, 616, 300] (category: vehicle)\n- G4: [115, 269, 306, 323] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A2: [510, 498, 522, 514] (category: vehicle)\n- A4: [508, 559, 525, 575] (category: vehicle)\n- A1: [487, 372, 497, 385] (category: vehicle)\n- A6: [520, 382, 530, 396] (category: vehicle)\n- A5: [458, 493, 474, 510] (category: vehicle)\n- A3: [530, 458, 541, 470] (category: vehicle)\n\nAnswer options:\nA: G1->A1, G2->A2, G3->A3, G4->A5\nB: G1->A5, G2->A1, G3->A2, G4->A4\nC: G1->A5, G2->A4, G3->A1, G4->A6\nD: G1->A5, G2->A2, G3->A4, G4->A1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b719686ac3ce75d89888dbdd46c328c0141d8802f7ef17467032f5f9c3aea3c0.jpg", "reference_inputs/images/d244fd3db364087e2ec3d7b95f0825a29b0801934b47170eb9f4cba1fe921459.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000063", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000063", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G1, G2, G4, G3, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [530, 493, 543, 509] (category: vehicle)\n- A2: [501, 429, 516, 451] (category: vehicle)\n- A3: [480, 453, 493, 464] (category: vehicle)\n- A4: [474, 428, 484, 440] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [446, 281, 474, 294] (category: vehicle)\n- G1: [491, 258, 518, 292] (category: vehicle)\n- G2: [358, 271, 368, 290] (category: pedestrian)\n- G4: [443, 277, 462, 291] (category: vehicle)\n- G3: [690, 266, 703, 293] (category: pedestrian)\n- G6: [563, 272, 606, 300] (category: vehicle)\n\nAnswer options:\nA: A1->G6, A2->G1, A3->G5, A4->G4\nB: A1->G5, A2->G4, A3->G1, A4->G6\nC: A1->G3, A2->G6, A3->G1, A4->G5\nD: A1->G4, A2->G2, A3->G5, A4->G3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cfa18aef7c5f1fd61ebebff7dafc64d1918ac5e9d9bb3839c48b3dd33dba961b.jpg", "reference_inputs/images/34d027dbab11397f568f524e93a1358bb23b7502da274b0c04a1dba07548b837.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000064", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000064", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A1, A2, A3, A4, A6, A5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [388, 267, 394, 292] (category: pedestrian)\n- G2: [773, 267, 780, 291] (category: pedestrian)\n- G3: [642, 278, 664, 293] (category: vehicle)\n- G4: [527, 275, 553, 293] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A1: [236, 528, 256, 548] (category: pedestrian)\n- A2: [698, 422, 711, 438] (category: pedestrian)\n- A3: [286, 553, 305, 570] (category: pedestrian)\n- A4: [550, 279, 578, 352] (category: vehicle)\n- A6: [261, 487, 279, 505] (category: pedestrian)\n- A5: [374, 273, 409, 344] (category: vehicle)\n\nAnswer options:\nA: G1->A6, G2->A1, G3->A3, G4->A4\nB: G1->A1, G2->A2, G3->A4, G4->A5\nC: G1->A4, G2->A1, G3->A5, G4->A2\nD: G1->A1, G2->A2, G3->A5, G4->A3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b3574f508469643224817c0968375f2dbc9b8fb5695e469f277f5de5349a47a5.jpg", "reference_inputs/images/4ab1ce6d791179a724a0abda241d57be67280c3d399e30581a826bced7e1b7b5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000065", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000065", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A2, A3, A4, A6, A5, A1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [497, 276, 554, 305] (category: vehicle)\n- G2: [424, 282, 470, 293] (category: vehicle)\n- G3: [534, 263, 556, 290] (category: vehicle)\n- G4: [234, 263, 342, 311] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A2: [472, 429, 492, 437] (category: vehicle)\n- A3: [522, 402, 534, 420] (category: vehicle)\n- A4: [508, 481, 522, 493] (category: vehicle)\n- A6: [458, 490, 470, 508] (category: vehicle)\n- A5: [482, 387, 494, 404] (category: vehicle)\n- A1: [536, 526, 553, 547] (category: vehicle)\n\nAnswer options:\nA: G1->A4, G2->A5, G3->A3, G4->A1\nB: G1->A4, G2->A2, G3->A3, G4->A6\nC: G1->A6, G2->A4, G3->A1, G4->A5\nD: G1->A6, G2->A5, G3->A1, G4->A3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8d896736401234e10331aeea3ac1fabfe73c9ad39a25d19bf9916657bd895404.jpg", "reference_inputs/images/e243169b9506053af12d4cec254d3ac2558083486bf9158c6ce0ca07d9a10508.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000067", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000067", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G3, G5, G1, G2, G4, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [535, 524, 550, 545] (category: vehicle)\n- A2: [507, 482, 522, 495] (category: vehicle)\n- A3: [530, 466, 542, 477] (category: vehicle)\n- A4: [488, 466, 498, 478] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G3: [716, 278, 762, 293] (category: vehicle)\n- G5: [492, 275, 557, 307] (category: vehicle)\n- G1: [179, 256, 262, 293] (category: vehicle)\n- G2: [428, 275, 474, 301] (category: vehicle)\n- G4: [566, 278, 618, 301] (category: vehicle)\n- G6: [770, 197, 1023, 386] (category: vehicle)\n\nAnswer options:\nA: A1->G1, A2->G5, A3->G4, A4->G2\nB: A1->G6, A2->G4, A3->G2, A4->G3\nC: A1->G6, A2->G5, A3->G4, A4->G2\nD: A1->G4, A2->G2, A3->G6, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f6f1c52deeda02696e616062d3700b17ffeac1e2d963d9ec5a0d768c6fccd7b7.jpg", "reference_inputs/images/4361054b280cd9d73394edc6e371afc818b989aed9f9327cd8d6243c79133116.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000068", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000068", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A6, A3, A4, A2, A5, A1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [770, 197, 1023, 386] (category: vehicle)\n- G2: [566, 278, 618, 301] (category: vehicle)\n- G3: [492, 275, 557, 307] (category: vehicle)\n- G4: [428, 275, 474, 301] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A6: [530, 466, 542, 477] (category: vehicle)\n- A3: [450, 524, 469, 546] (category: vehicle)\n- A4: [535, 524, 550, 545] (category: vehicle)\n- A2: [382, 561, 386, 575] (category: pedestrian)\n- A5: [507, 482, 522, 495] (category: vehicle)\n- A1: [488, 466, 498, 478] (category: vehicle)\n\nAnswer options:\nA: G1->A4, G2->A6, G3->A5, G4->A1\nB: G1->A3, G2->A5, G3->A2, G4->A6\nC: G1->A5, G2->A6, G3->A4, G4->A1\nD: G1->A1, G2->A4, G3->A5, G4->A3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4361054b280cd9d73394edc6e371afc818b989aed9f9327cd8d6243c79133116.jpg", "reference_inputs/images/f6f1c52deeda02696e616062d3700b17ffeac1e2d963d9ec5a0d768c6fccd7b7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000069", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000069", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G2, G4, G1, G3, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [651, 81, 663, 97] (category: pedestrian)\n- A2: [692, 12, 706, 32] (category: pedestrian)\n- A3: [673, 106, 684, 124] (category: pedestrian)\n- A4: [670, 306, 682, 314] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [212, 223, 237, 317] (category: pedestrian)\n- G2: [790, 257, 806, 303] (category: pedestrian)\n- G4: [824, 278, 830, 294] (category: pedestrian)\n- G1: [826, 227, 858, 320] (category: pedestrian)\n- G3: [86, 208, 129, 320] (category: pedestrian)\n- G6: [2, 222, 35, 311] (category: pedestrian)\n\nAnswer options:\nA: A1->G3, A2->G6, A3->G5, A4->G1\nB: A1->G5, A2->G6, A3->G1, A4->G2\nC: A1->G3, A2->G6, A3->G2, A4->G5\nD: A1->G6, A2->G5, A3->G3, A4->G2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/54fe527ada2d1060e9abe51e500c229043388ec7321da0fcfb22ea61353b10b0.jpg", "reference_inputs/images/794cfbcbe7aebf0fd07de502de177b39301e6da824fb87e6bb1e4aa716ca8342.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000070", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000070", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A1, A3, A6, A4, A5, A2. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [212, 223, 237, 317] (category: pedestrian)\n- G2: [86, 208, 129, 320] (category: pedestrian)\n- G3: [850, 280, 894, 303] (category: vehicle)\n- G4: [2, 222, 35, 311] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A1: [651, 81, 663, 97] (category: pedestrian)\n- A3: [670, 306, 682, 314] (category: pedestrian)\n- A6: [692, 12, 706, 32] (category: pedestrian)\n- A4: [272, 39, 288, 61] (category: pedestrian)\n- A5: [879, 488, 953, 524] (category: vehicle)\n- A2: [673, 106, 684, 124] (category: pedestrian)\n\nAnswer options:\nA: G1->A1, G2->A5, G3->A4, G4->A2\nB: G1->A4, G2->A6, G3->A1, G4->A5\nC: G1->A2, G2->A4, G3->A6, G4->A3\nD: G1->A2, G2->A1, G3->A5, G4->A6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/794cfbcbe7aebf0fd07de502de177b39301e6da824fb87e6bb1e4aa716ca8342.jpg", "reference_inputs/images/54fe527ada2d1060e9abe51e500c229043388ec7321da0fcfb22ea61353b10b0.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000071", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000071", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G1, G5, G6, G2, G3, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [286, 470, 315, 482] (category: pedestrian)\n- A2: [673, 292, 686, 298] (category: pedestrian)\n- A3: [186, 140, 212, 154] (category: pedestrian)\n- A4: [279, 492, 297, 508] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G1: [771, 258, 781, 295] (category: pedestrian)\n- G5: [120, 246, 137, 301] (category: pedestrian)\n- G6: [229, 274, 233, 290] (category: pedestrian)\n- G2: [162, 255, 193, 300] (category: pedestrian)\n- G3: [678, 265, 686, 294] (category: pedestrian)\n- G4: [265, 267, 271, 292] (category: pedestrian)\n\nAnswer options:\nA: A1->G4, A2->G6, A3->G5, A4->G2\nB: A1->G2, A2->G3, A3->G4, A4->G5\nC: A1->G4, A2->G3, A3->G6, A4->G1\nD: A1->G5, A2->G1, A3->G6, A4->G2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1a0ea3e0c43e1f8333c7b5ac26d86b0d4127ca6debff07ec7d82fe47ce562749.jpg", "reference_inputs/images/c6777d37805bc1e28ba259a064593108c7c27c287b70aceb3538e5e222e83c53.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000072", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000072", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A4, A6, A3, A1, A5, A2. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [265, 267, 271, 292] (category: pedestrian)\n- G2: [771, 258, 781, 295] (category: pedestrian)\n- G3: [162, 255, 193, 300] (category: pedestrian)\n- G4: [120, 246, 137, 301] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A4: [719, 379, 735, 388] (category: pedestrian)\n- A6: [186, 140, 212, 154] (category: pedestrian)\n- A3: [286, 470, 315, 482] (category: pedestrian)\n- A1: [490, 517, 525, 575] (category: vehicle)\n- A5: [673, 292, 686, 298] (category: pedestrian)\n- A2: [279, 492, 297, 508] (category: pedestrian)\n\nAnswer options:\nA: G1->A5, G2->A3, G3->A1, G4->A6\nB: G1->A1, G2->A4, G3->A5, G4->A6\nC: G1->A4, G2->A5, G3->A3, G4->A6\nD: G1->A6, G2->A4, G3->A3, G4->A2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c6777d37805bc1e28ba259a064593108c7c27c287b70aceb3538e5e222e83c53.jpg", "reference_inputs/images/1a0ea3e0c43e1f8333c7b5ac26d86b0d4127ca6debff07ec7d82fe47ce562749.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_30m/000073", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000073", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G3, G5, G4, G1, G2, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [756, 191, 775, 200] (category: pedestrian)\n- A2: [678, 146, 690, 159] (category: pedestrian)\n- A3: [674, 392, 686, 403] (category: pedestrian)\n- A4: [755, 370, 775, 378] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G3: [813, 272, 823, 298] (category: pedestrian)\n- G5: [412, 262, 424, 298] (category: pedestrian)\n- G4: [197, 218, 230, 314] (category: pedestrian)\n- G1: [741, 246, 761, 306] (category: pedestrian)\n- G2: [385, 243, 401, 304] (category: pedestrian)\n- G6: [918, 232, 971, 322] (category: pedestrian)\n\nAnswer options:\nA: A1->G6, A2->G2, A3->G1, A4->G4\nB: A1->G4, A2->G6, A3->G1, A4->G2\nC: A1->G3, A2->G1, A3->G2, A4->G6\nD: A1->G2, A2->G4, A3->G6, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/425ae2a6e520d6325a6ef5ca44e168b8600759bb84f76528dbe67ab98b445533.jpg", "reference_inputs/images/3b84b5d82519d128b30a19c0cdb9e0532e599f2540e6a0fb617747af66d36169.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_50m/000001", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G1, G6, G2, G4, G3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [470, 190, 489, 234] (category: vehicle)\n- A2: [536, 289, 562, 367] (category: vehicle)\n- A3: [434, 12, 455, 58] (category: vehicle)\n- A4: [434, 492, 453, 532] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [537, 250, 579, 294] (category: vehicle)\n- G1: [311, 274, 377, 304] (category: vehicle)\n- G6: [508, 278, 522, 290] (category: vehicle)\n- G2: [522, 275, 531, 287] (category: vehicle)\n- G4: [454, 282, 470, 290] (category: vehicle)\n- G3: [470, 278, 489, 292] (category: vehicle)\n\nAnswer options:\nA: A1->G3, A2->G5, A3->G4, A4->G1\nB: A1->G4, A2->G3, A3->G6, A4->G1\nC: A1->G4, A2->G5, A3->G1, A4->G3\nD: A1->G4, A2->G3, A3->G1, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1336f727d14e0a43920ba0e565b7a3a66a5ef82e0946e2742228982971657d7b.jpg", "reference_inputs/images/326e309493b74656ab397a5c7fe4a7d782f7c238e5d6a88d3395323169ed9146.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_50m/000002", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A3, A4, A6, A5, A2, A1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [537, 250, 579, 294] (category: vehicle)\n- G2: [470, 278, 489, 292] (category: vehicle)\n- G3: [311, 274, 377, 304] (category: vehicle)\n- G4: [454, 282, 470, 290] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A3: [434, 12, 455, 58] (category: vehicle)\n- A4: [434, 492, 453, 532] (category: vehicle)\n- A6: [470, 190, 489, 234] (category: vehicle)\n- A5: [507, 0, 527, 34] (category: vehicle)\n- A2: [536, 289, 562, 367] (category: vehicle)\n- A1: [378, 549, 384, 562] (category: pedestrian)\n\nAnswer options:\nA: G1->A4, G2->A5, G3->A1, G4->A3\nB: G1->A2, G2->A6, G3->A4, G4->A3\nC: G1->A4, G2->A3, G3->A1, G4->A2\nD: G1->A5, G2->A2, G3->A3, G4->A4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/326e309493b74656ab397a5c7fe4a7d782f7c238e5d6a88d3395323169ed9146.jpg", "reference_inputs/images/1336f727d14e0a43920ba0e565b7a3a66a5ef82e0946e2742228982971657d7b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_50m/000003", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G3, G1, G6, G2, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [574, 458, 596, 503] (category: vehicle)\n- A2: [511, 294, 532, 337] (category: vehicle)\n- A3: [508, 30, 526, 72] (category: vehicle)\n- A4: [570, 162, 589, 203] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [574, 280, 582, 306] (category: pedestrian)\n- G3: [442, 290, 489, 310] (category: vehicle)\n- G1: [310, 292, 335, 308] (category: vehicle)\n- G6: [266, 288, 278, 297] (category: vehicle)\n- G2: [341, 290, 366, 305] (category: vehicle)\n- G4: [261, 295, 295, 304] (category: vehicle)\n\nAnswer options:\nA: A1->G4, A2->G2, A3->G3, A4->G1\nB: A1->G6, A2->G3, A3->G2, A4->G1\nC: A1->G6, A2->G3, A3->G5, A4->G4\nD: A1->G5, A2->G3, A3->G1, A4->G2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6335503211ddf6b2368041e62c88cd8c440ca0c151d187f98a89d26e416d6c09.jpg", "reference_inputs/images/606dd5fea012af2756b373ee21b6a9f4a7ff6eca37a4eb356fb07f1cb3476f3e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_50m/000005", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A2, A3, A5, A6, A4, A1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [502, 261, 572, 314] (category: vehicle)\n- G2: [181, 255, 198, 298] (category: pedestrian)\n- G3: [782, 255, 797, 298] (category: pedestrian)\n- G4: [888, 266, 1023, 399] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A2: [873, 518, 886, 526] (category: pedestrian)\n- A3: [1002, 360, 1023, 379] (category: vehicle)\n- A5: [957, 502, 973, 513] (category: pedestrian)\n- A6: [574, 0, 596, 8] (category: vehicle)\n- A4: [958, 258, 974, 264] (category: pedestrian)\n- A1: [655, 532, 662, 542] (category: pedestrian)\n\nAnswer options:\nA: G1->A4, G2->A3, G3->A6, G4->A1\nB: G1->A3, G2->A2, G3->A4, G4->A6\nC: G1->A3, G2->A5, G3->A4, G4->A6\nD: G1->A5, G2->A3, G3->A6, G4->A4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/64ecb14e96931fa12d63d7d52e97e6551d9ab8794bec873526ad42f1cf64fd39.jpg", "reference_inputs/images/d1ee783bb91590ea5128726b99ed4acd64eb536a4dfa2e81f2a4271e019ffe16.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_50m/000006", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G3, G5, G4, G2, G1, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [496, 197, 522, 276] (category: vehicle)\n- A2: [571, 120, 590, 161] (category: vehicle)\n- A3: [530, 442, 551, 486] (category: vehicle)\n- A4: [574, 454, 597, 498] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G3: [430, 276, 450, 286] (category: vehicle)\n- G5: [329, 260, 381, 286] (category: vehicle)\n- G4: [467, 281, 482, 290] (category: vehicle)\n- G2: [481, 247, 523, 302] (category: vehicle)\n- G1: [242, 214, 270, 282] (category: pedestrian)\n- G6: [0, 185, 28, 245] (category: pedestrian)\n\nAnswer options:\nA: A1->G5, A2->G4, A3->G1, A4->G2\nB: A1->G2, A2->G5, A3->G4, A4->G3\nC: A1->G3, A2->G6, A3->G2, A4->G5\nD: A1->G1, A2->G2, A3->G4, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f1b1fabf0f694b63c2ad925139ff55b74e5134f92f21fc72456f28796a8d42c4.jpg", "reference_inputs/images/30e6bd80fec7eeb677a4b663c71af085dbdb8ae9f65814a556f247652e14e01b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_50m/000009", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G4, G2, G6, G1, G3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [507, 365, 527, 407] (category: vehicle)\n- A2: [474, 24, 493, 66] (category: vehicle)\n- A3: [434, 202, 454, 245] (category: vehicle)\n- A4: [472, 559, 487, 575] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [554, 278, 579, 294] (category: vehicle)\n- G4: [411, 274, 441, 294] (category: vehicle)\n- G2: [0, 210, 334, 399] (category: vehicle)\n- G6: [476, 273, 490, 288] (category: vehicle)\n- G1: [500, 264, 544, 301] (category: vehicle)\n- G3: [478, 281, 494, 291] (category: vehicle)\n\nAnswer options:\nA: A1->G1, A2->G2, A3->G3, A4->G4\nB: A1->G3, A2->G1, A3->G5, A4->G4\nC: A1->G2, A2->G3, A3->G5, A4->G6\nD: A1->G1, A2->G3, A3->G4, A4->G2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5279243a189dac5df5abe787abc68a73bb0ecaf5b22deee4090ef8daf12ab644.jpg", "reference_inputs/images/876e52177c54d2f1a987ee7d6b60aef790c909d21871bccc19c29ad50f3249e4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_50m/000010", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A2, A6, A5, A4, A1, A3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [500, 264, 544, 301] (category: vehicle)\n- G2: [411, 274, 441, 294] (category: vehicle)\n- G3: [478, 281, 494, 291] (category: vehicle)\n- G4: [0, 210, 334, 399] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A2: [546, 208, 563, 248] (category: vehicle)\n- A6: [434, 385, 456, 429] (category: vehicle)\n- A5: [434, 202, 454, 245] (category: vehicle)\n- A4: [507, 365, 527, 407] (category: vehicle)\n- A1: [472, 559, 487, 575] (category: vehicle)\n- A3: [474, 24, 493, 66] (category: vehicle)\n\nAnswer options:\nA: G1->A4, G2->A5, G3->A3, G4->A1\nB: G1->A4, G2->A2, G3->A1, G4->A6\nC: G1->A2, G2->A4, G3->A3, G4->A1\nD: G1->A4, G2->A1, G3->A6, G4->A3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/876e52177c54d2f1a987ee7d6b60aef790c909d21871bccc19c29ad50f3249e4.jpg", "reference_inputs/images/5279243a189dac5df5abe787abc68a73bb0ecaf5b22deee4090ef8daf12ab644.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_50m/000011", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G1, G5, G6, G3, G4, G2. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [433, 209, 450, 253] (category: vehicle)\n- A2: [473, 453, 493, 496] (category: vehicle)\n- A3: [499, 162, 527, 239] (category: vehicle)\n- A4: [511, 402, 526, 446] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G1: [300, 273, 305, 293] (category: pedestrian)\n- G5: [337, 241, 417, 302] (category: vehicle)\n- G6: [472, 280, 484, 294] (category: vehicle)\n- G3: [498, 278, 528, 300] (category: vehicle)\n- G4: [614, 272, 621, 296] (category: pedestrian)\n- G2: [425, 281, 446, 294] (category: vehicle)\n\nAnswer options:\nA: A1->G3, A2->G4, A3->G2, A4->G6\nB: A1->G3, A2->G6, A3->G5, A4->G2\nC: A1->G2, A2->G3, A3->G4, A4->G5\nD: A1->G5, A2->G2, A3->G1, A4->G3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d78eac01e2212aeff0e4dc193dad35c2de808cf46bddb2f94cf15c2e06f99339.jpg", "reference_inputs/images/772ae829aab097e813edc986df6337aa53a4dfb46188b630cff56caf74228b46.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/City06_ground_aerial_v2_50m/000012", "task_name": "multi_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A3, A5, A4, A2, A6, A1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [337, 241, 417, 302] (category: vehicle)\n- G2: [472, 280, 484, 294] (category: vehicle)\n- G3: [498, 278, 528, 300] (category: vehicle)\n- G4: [425, 281, 446, 294] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A3: [511, 402, 526, 446] (category: vehicle)\n- A5: [542, 522, 564, 567] (category: vehicle)\n- A4: [499, 162, 527, 239] (category: vehicle)\n- A2: [437, 566, 456, 575] (category: vehicle)\n- A6: [433, 209, 450, 253] (category: vehicle)\n- A1: [473, 453, 493, 496] (category: vehicle)\n\nAnswer options:\nA: G1->A4, G2->A1, G3->A6, G4->A3\nB: G1->A1, G2->A5, G3->A6, G4->A2\nC: G1->A2, G2->A4, G3->A1, G4->A3\nD: G1->A2, G2->A1, G3->A4, G4->A3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/772ae829aab097e813edc986df6337aa53a4dfb46188b630cff56caf74228b46.jpg", "reference_inputs/images/d78eac01e2212aeff0e4dc193dad35c2de808cf46bddb2f94cf15c2e06f99339.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000001", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G4, G1, G3, G5, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [504, 440, 518, 456] (category: vehicle)\n- A2: [450, 430, 466, 444] (category: vehicle)\n- A3: [565, 472, 570, 490] (category: pedestrian)\n- A4: [707, 472, 713, 490] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [451, 289, 479, 301] (category: vehicle)\n- G4: [912, 288, 924, 324] (category: pedestrian)\n- G1: [410, 298, 439, 311] (category: vehicle)\n- G3: [574, 287, 578, 306] (category: pedestrian)\n- G5: [498, 295, 523, 314] (category: vehicle)\n- G6: [623, 288, 632, 324] (category: pedestrian)\n\nAnswer options:\nA: A1->G1, A2->G4, A3->G2, A4->G5\nB: A1->G5, A2->G6, A3->G3, A4->G4\nC: A1->G5, A2->G1, A3->G6, A4->G4\nD: A1->G4, A2->G3, A3->G5, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/532a7bb18fb47c2c7a6afad29eee2e22fe5703751da3632aa370b49a67654041.jpg", "reference_inputs/images/bb1b3b3955a3f2a6491e5176b2d74a97d2bd6453e48df238d57ef60e4f23c296.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000003", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G1, G4, G2, G3, G6, G5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [392, 538, 443, 569] (category: vehicle)\n- A2: [528, 438, 562, 454] (category: vehicle)\n- A3: [531, 538, 588, 561] (category: vehicle)\n- A4: [424, 469, 462, 488] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G1: [437, 291, 473, 305] (category: vehicle)\n- G4: [530, 290, 563, 298] (category: vehicle)\n- G2: [398, 291, 450, 310] (category: vehicle)\n- G3: [525, 294, 554, 306] (category: vehicle)\n- G6: [532, 298, 570, 314] (category: vehicle)\n- G5: [387, 298, 441, 314] (category: vehicle)\n\nAnswer options:\nA: A1->G3, A2->G6, A3->G5, A4->G4\nB: A1->G5, A2->G4, A3->G6, A4->G1\nC: A1->G5, A2->G3, A3->G6, A4->G4\nD: A1->G4, A2->G5, A3->G1, A4->G3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/17b62a74f74c1ea060246e106301a4bc1105d82173d30d693bf2f9d1dd4a700a.jpg", "reference_inputs/images/c050fcbf6a2fbb7ae73cec88f53be285420ba47b9cda22afa8774910d8ad59db.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000004", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A5, A3, A1, A4, A6, A2. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [398, 291, 450, 310] (category: vehicle)\n- G2: [658, 289, 666, 314] (category: pedestrian)\n- G3: [387, 298, 441, 314] (category: vehicle)\n- G4: [530, 290, 563, 298] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A5: [392, 538, 443, 569] (category: vehicle)\n- A3: [424, 469, 462, 488] (category: vehicle)\n- A1: [528, 438, 562, 454] (category: vehicle)\n- A4: [401, 511, 451, 541] (category: vehicle)\n- A6: [524, 479, 569, 513] (category: vehicle)\n- A2: [651, 565, 658, 575] (category: pedestrian)\n\nAnswer options:\nA: G1->A4, G2->A2, G3->A5, G4->A1\nB: G1->A3, G2->A5, G3->A1, G4->A2\nC: G1->A4, G2->A2, G3->A1, G4->A5\nD: G1->A3, G2->A5, G3->A6, G4->A2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c050fcbf6a2fbb7ae73cec88f53be285420ba47b9cda22afa8774910d8ad59db.jpg", "reference_inputs/images/17b62a74f74c1ea060246e106301a4bc1105d82173d30d693bf2f9d1dd4a700a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000006", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A5, A3, A2, A6, A1, A4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [442, 290, 470, 303] (category: vehicle)\n- G2: [655, 287, 666, 320] (category: pedestrian)\n- G3: [125, 290, 154, 338] (category: pedestrian)\n- G4: [524, 296, 550, 310] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A5: [635, 424, 665, 437] (category: vehicle)\n- A3: [506, 547, 512, 573] (category: pedestrian)\n- A2: [623, 445, 654, 458] (category: vehicle)\n- A6: [719, 412, 750, 426] (category: vehicle)\n- A1: [720, 458, 765, 486] (category: vehicle)\n- A4: [822, 542, 831, 568] (category: pedestrian)\n\nAnswer options:\nA: G1->A1, G2->A5, G3->A3, G4->A2\nB: G1->A3, G2->A5, G3->A1, G4->A2\nC: G1->A3, G2->A6, G3->A5, G4->A1\nD: G1->A5, G2->A4, G3->A3, G4->A1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/daafd56b7f9b0584e99977588f276d6d9f561cd5540ec6e91fc46c16fc19f642.jpg", "reference_inputs/images/d5bfe9a4746a03c92cbeee7518807d18504fafcbc4319a64d05fe881f42e9aab.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000007", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G4, G6, G1, G2, G3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [398, 519, 450, 550] (category: vehicle)\n- A2: [354, 522, 359, 542] (category: pedestrian)\n- A3: [529, 502, 573, 521] (category: vehicle)\n- A4: [534, 548, 593, 573] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [354, 289, 360, 308] (category: pedestrian)\n- G4: [438, 291, 474, 305] (category: vehicle)\n- G6: [525, 294, 554, 306] (category: vehicle)\n- G1: [543, 296, 583, 317] (category: vehicle)\n- G2: [532, 298, 570, 313] (category: vehicle)\n- G3: [399, 291, 450, 310] (category: vehicle)\n\nAnswer options:\nA: A1->G5, A2->G6, A3->G1, A4->G4\nB: A1->G4, A2->G2, A3->G1, A4->G5\nC: A1->G6, A2->G5, A3->G2, A4->G4\nD: A1->G3, A2->G5, A3->G6, A4->G2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8d623e668ea8bf1fb46683001b988f91917220a05ed970bf04e08492689ac461.jpg", "reference_inputs/images/f598fb5a843da633d25b3a73385b1f3542e3af3c5ae2199248979d89ad1ff582.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000009", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G4, G1, G2, G3, G5, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [438, 552, 444, 575] (category: pedestrian)\n- A2: [591, 406, 605, 418] (category: vehicle)\n- A3: [588, 525, 594, 546] (category: pedestrian)\n- A4: [581, 511, 586, 534] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G4: [271, 290, 293, 374] (category: pedestrian)\n- G1: [690, 298, 711, 312] (category: vehicle)\n- G2: [626, 292, 648, 304] (category: vehicle)\n- G3: [677, 290, 691, 342] (category: pedestrian)\n- G5: [714, 294, 729, 350] (category: pedestrian)\n- G6: [588, 290, 605, 299] (category: vehicle)\n\nAnswer options:\nA: A1->G4, A2->G2, A3->G5, A4->G3\nB: A1->G2, A2->G1, A3->G6, A4->G5\nC: A1->G3, A2->G6, A3->G1, A4->G4\nD: A1->G4, A2->G5, A3->G6, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5decbba8daa7c5df5a7fff7c056b20766744bd23a17ecce6a98368f20a0ea9b9.jpg", "reference_inputs/images/b85a6b29f399b960fda0517b94c5f99319b190ff4a324961395b1f408498e61b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000010", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A5, A6, A4, A2, A3, A1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [677, 290, 691, 342] (category: pedestrian)\n- G2: [626, 292, 648, 304] (category: vehicle)\n- G3: [714, 294, 729, 350] (category: pedestrian)\n- G4: [690, 298, 711, 312] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A5: [588, 525, 594, 546] (category: pedestrian)\n- A6: [581, 511, 586, 534] (category: pedestrian)\n- A4: [438, 552, 444, 575] (category: pedestrian)\n- A2: [640, 415, 657, 427] (category: vehicle)\n- A3: [611, 438, 631, 453] (category: vehicle)\n- A1: [591, 406, 605, 418] (category: vehicle)\n\nAnswer options:\nA: G1->A6, G2->A1, G3->A5, G4->A3\nB: G1->A3, G2->A1, G3->A2, G4->A6\nC: G1->A4, G2->A3, G3->A5, G4->A2\nD: G1->A1, G2->A2, G3->A6, G4->A4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b85a6b29f399b960fda0517b94c5f99319b190ff4a324961395b1f408498e61b.jpg", "reference_inputs/images/5decbba8daa7c5df5a7fff7c056b20766744bd23a17ecce6a98368f20a0ea9b9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000011", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G4, G1, G6, G5, G3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [582, 426, 614, 437] (category: vehicle)\n- A2: [842, 422, 877, 433] (category: vehicle)\n- A3: [826, 399, 856, 410] (category: vehicle)\n- A4: [583, 442, 614, 482] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [280, 291, 305, 300] (category: vehicle)\n- G4: [597, 226, 1023, 418] (category: vehicle)\n- G1: [2, 330, 157, 390] (category: vehicle)\n- G6: [430, 294, 444, 302] (category: vehicle)\n- G5: [385, 295, 410, 310] (category: vehicle)\n- G3: [0, 312, 23, 332] (category: vehicle)\n\nAnswer options:\nA: A1->G4, A2->G3, A3->G6, A4->G5\nB: A1->G1, A2->G5, A3->G2, A4->G4\nC: A1->G6, A2->G2, A3->G1, A4->G3\nD: A1->G1, A2->G6, A3->G5, A4->G4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8502f61f181f5d74610587b6ad6b6b5a8d2308a26aaf20c07336a3dfbc0b0ca5.jpg", "reference_inputs/images/efa5658689482f7351dfb2d3e76efc3166d2ac78bdd9cedd302f88a92851f48f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000012", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A2, A3, A6, A5, A4, A1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [280, 291, 305, 300] (category: vehicle)\n- G2: [2, 330, 157, 390] (category: vehicle)\n- G3: [597, 226, 1023, 418] (category: vehicle)\n- G4: [385, 295, 410, 310] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A2: [842, 422, 877, 433] (category: vehicle)\n- A3: [696, 422, 730, 431] (category: vehicle)\n- A6: [628, 403, 652, 410] (category: vehicle)\n- A5: [583, 442, 614, 482] (category: vehicle)\n- A4: [826, 399, 856, 410] (category: vehicle)\n- A1: [582, 426, 614, 437] (category: vehicle)\n\nAnswer options:\nA: G1->A3, G2->A6, G3->A5, G4->A1\nB: G1->A4, G2->A1, G3->A2, G4->A3\nC: G1->A4, G2->A1, G3->A5, G4->A2\nD: G1->A3, G2->A6, G3->A5, G4->A4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/efa5658689482f7351dfb2d3e76efc3166d2ac78bdd9cedd302f88a92851f48f.jpg", "reference_inputs/images/8502f61f181f5d74610587b6ad6b6b5a8d2308a26aaf20c07336a3dfbc0b0ca5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000014", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G6, G4, G1, G2, G3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [204, 525, 278, 575] (category: vehicle)\n- A2: [185, 222, 299, 286] (category: vehicle)\n- A3: [518, 311, 594, 404] (category: vehicle)\n- A4: [129, 133, 168, 148] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [341, 302, 348, 327] (category: pedestrian)\n- G6: [563, 315, 604, 330] (category: vehicle)\n- G4: [558, 319, 602, 340] (category: vehicle)\n- G1: [396, 310, 463, 337] (category: vehicle)\n- G2: [570, 330, 626, 353] (category: vehicle)\n- G3: [587, 333, 666, 364] (category: vehicle)\n\nAnswer options:\nA: A1->G2, A2->G3, A3->G6, A4->G5\nB: A1->G3, A2->G1, A3->G4, A4->G5\nC: A1->G4, A2->G2, A3->G5, A4->G1\nD: A1->G3, A2->G5, A3->G4, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c1fe12231e3c6505f4561f7b59d589c5fc02488b59c32ec3725a2de26e313288.jpg", "reference_inputs/images/12612aa50c4da7e27c053652a3c4e25219e5ca62e5385dfabd28088fd37f4ec1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000015", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A6, A3, A4, A2, A1, A5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [587, 333, 666, 364] (category: vehicle)\n- G2: [558, 319, 602, 340] (category: vehicle)\n- G3: [570, 330, 626, 353] (category: vehicle)\n- G4: [396, 310, 463, 337] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A6: [185, 222, 299, 286] (category: vehicle)\n- A3: [518, 311, 594, 404] (category: vehicle)\n- A4: [282, 459, 373, 556] (category: vehicle)\n- A2: [204, 525, 278, 575] (category: vehicle)\n- A1: [366, 414, 454, 512] (category: vehicle)\n- A5: [442, 364, 518, 455] (category: vehicle)\n\nAnswer options:\nA: G1->A2, G2->A3, G3->A4, G4->A6\nB: G1->A4, G2->A3, G3->A5, G4->A2\nC: G1->A3, G2->A2, G3->A1, G4->A5\nD: G1->A5, G2->A6, G3->A3, G4->A4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/12612aa50c4da7e27c053652a3c4e25219e5ca62e5385dfabd28088fd37f4ec1.jpg", "reference_inputs/images/c1fe12231e3c6505f4561f7b59d589c5fc02488b59c32ec3725a2de26e313288.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000017", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G1, G6, G2, G4, G3, G5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [595, 383, 618, 389] (category: vehicle)\n- A2: [462, 400, 474, 414] (category: vehicle)\n- A3: [702, 370, 723, 378] (category: vehicle)\n- A4: [501, 390, 518, 412] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G1: [977, 294, 1023, 310] (category: vehicle)\n- G6: [921, 290, 946, 297] (category: vehicle)\n- G2: [529, 298, 568, 313] (category: vehicle)\n- G4: [313, 296, 379, 336] (category: vehicle)\n- G3: [717, 298, 778, 311] (category: vehicle)\n- G5: [471, 259, 531, 334] (category: vehicle)\n\nAnswer options:\nA: A1->G4, A2->G5, A3->G6, A4->G3\nB: A1->G3, A2->G4, A3->G6, A4->G5\nC: A1->G3, A2->G4, A3->G2, A4->G6\nD: A1->G2, A2->G1, A3->G4, A4->G3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fc020544619452bceeb4c01d861680999eff0a57983a98dd5efea5e57ebef25d.jpg", "reference_inputs/images/6c1faeec1b212385a6473aba22bc10edfcccad8ee4a4d132b5fa6116fc167924.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000018", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A1, A5, A2, A4, A6, A3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [717, 298, 778, 311] (category: vehicle)\n- G2: [313, 296, 379, 336] (category: vehicle)\n- G3: [977, 294, 1023, 310] (category: vehicle)\n- G4: [921, 290, 946, 297] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A1: [595, 383, 618, 389] (category: vehicle)\n- A5: [698, 382, 721, 390] (category: vehicle)\n- A2: [822, 383, 846, 390] (category: vehicle)\n- A4: [462, 400, 474, 414] (category: vehicle)\n- A6: [702, 370, 723, 378] (category: vehicle)\n- A3: [511, 385, 533, 392] (category: vehicle)\n\nAnswer options:\nA: G1->A3, G2->A6, G3->A4, G4->A5\nB: G1->A5, G2->A6, G3->A1, G4->A4\nC: G1->A1, G2->A4, G3->A5, G4->A6\nD: G1->A2, G2->A1, G3->A4, G4->A5\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6c1faeec1b212385a6473aba22bc10edfcccad8ee4a4d132b5fa6116fc167924.jpg", "reference_inputs/images/fc020544619452bceeb4c01d861680999eff0a57983a98dd5efea5e57ebef25d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000019", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G3, G6, G4, G5, G1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [685, 494, 689, 514] (category: pedestrian)\n- A2: [522, 481, 541, 504] (category: vehicle)\n- A3: [454, 511, 459, 534] (category: pedestrian)\n- A4: [577, 432, 592, 444] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [590, 294, 609, 304] (category: vehicle)\n- G3: [526, 295, 550, 314] (category: vehicle)\n- G6: [742, 286, 749, 312] (category: pedestrian)\n- G4: [433, 287, 440, 316] (category: pedestrian)\n- G5: [462, 287, 468, 307] (category: pedestrian)\n- G1: [556, 279, 571, 296] (category: vehicle)\n\nAnswer options:\nA: A1->G3, A2->G6, A3->G5, A4->G4\nB: A1->G1, A2->G4, A3->G2, A4->G3\nC: A1->G4, A2->G6, A3->G5, A4->G2\nD: A1->G6, A2->G3, A3->G4, A4->G2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fb34b53b679d1237735335f190243505fad3e06e30ee7b8151f738312973ddd5.jpg", "reference_inputs/images/fb83154e0a62656e6b5c69979f54dd7b956389860903523b5d68cc091995ba8d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000020", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A4, A1, A5, A6, A3, A2. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [742, 286, 749, 312] (category: pedestrian)\n- G2: [590, 294, 609, 304] (category: vehicle)\n- G3: [526, 295, 550, 314] (category: vehicle)\n- G4: [433, 287, 440, 316] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A4: [577, 432, 592, 444] (category: vehicle)\n- A1: [519, 418, 530, 430] (category: vehicle)\n- A5: [453, 477, 458, 495] (category: pedestrian)\n- A6: [685, 494, 689, 514] (category: pedestrian)\n- A3: [454, 511, 459, 534] (category: pedestrian)\n- A2: [522, 481, 541, 504] (category: vehicle)\n\nAnswer options:\nA: G1->A4, G2->A2, G3->A1, G4->A3\nB: G1->A2, G2->A5, G3->A4, G4->A6\nC: G1->A6, G2->A3, G3->A5, G4->A1\nD: G1->A6, G2->A4, G3->A2, G4->A3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fb83154e0a62656e6b5c69979f54dd7b956389860903523b5d68cc091995ba8d.jpg", "reference_inputs/images/fb34b53b679d1237735335f190243505fad3e06e30ee7b8151f738312973ddd5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000021", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G6, G1, G2, G3, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [899, 211, 938, 226] (category: pedestrian)\n- A2: [631, 329, 646, 338] (category: pedestrian)\n- A3: [575, 0, 589, 22] (category: pedestrian)\n- A4: [781, 534, 806, 560] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [312, 291, 320, 324] (category: pedestrian)\n- G6: [340, 294, 349, 322] (category: pedestrian)\n- G1: [330, 292, 334, 313] (category: pedestrian)\n- G2: [539, 292, 545, 310] (category: pedestrian)\n- G3: [493, 290, 498, 311] (category: pedestrian)\n- G4: [172, 294, 180, 318] (category: pedestrian)\n\nAnswer options:\nA: A1->G6, A2->G1, A3->G3, A4->G4\nB: A1->G1, A2->G4, A3->G3, A4->G5\nC: A1->G4, A2->G5, A3->G6, A4->G2\nD: A1->G1, A2->G2, A3->G4, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/96282862df9fed1e54bdafbd1cefb080486775ae3b35c0d1cf0a80741f19ab67.jpg", "reference_inputs/images/6c19efe1630f6eca1709b8bf989300cf0ebc534fdb36b20b60aeb358ef605f59.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000022", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A2, A5, A6, A3, A4, A1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [172, 294, 180, 318] (category: pedestrian)\n- G2: [340, 294, 349, 322] (category: pedestrian)\n- G3: [419, 291, 426, 320] (category: pedestrian)\n- G4: [330, 292, 334, 313] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A2: [781, 534, 806, 560] (category: pedestrian)\n- A5: [899, 211, 938, 226] (category: pedestrian)\n- A6: [631, 329, 646, 338] (category: pedestrian)\n- A3: [575, 0, 589, 22] (category: pedestrian)\n- A4: [658, 364, 678, 379] (category: pedestrian)\n- A1: [871, 98, 908, 118] (category: pedestrian)\n\nAnswer options:\nA: G1->A2, G2->A5, G3->A1, G4->A6\nB: G1->A6, G2->A3, G3->A5, G4->A2\nC: G1->A4, G2->A5, G3->A6, G4->A2\nD: G1->A1, G2->A2, G3->A4, G4->A3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6c19efe1630f6eca1709b8bf989300cf0ebc534fdb36b20b60aeb358ef605f59.jpg", "reference_inputs/images/96282862df9fed1e54bdafbd1cefb080486775ae3b35c0d1cf0a80741f19ab67.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000023", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G2, G6, G1, G4, G3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [626, 551, 643, 575] (category: pedestrian)\n- A2: [797, 118, 853, 217] (category: vehicle)\n- A3: [654, 409, 669, 424] (category: pedestrian)\n- A4: [970, 370, 1023, 468] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [794, 286, 800, 304] (category: pedestrian)\n- G2: [888, 305, 985, 333] (category: vehicle)\n- G6: [632, 295, 642, 326] (category: pedestrian)\n- G1: [638, 288, 662, 338] (category: pedestrian)\n- G4: [679, 299, 722, 318] (category: vehicle)\n- G3: [820, 290, 827, 306] (category: pedestrian)\n\nAnswer options:\nA: A1->G1, A2->G4, A3->G6, A4->G2\nB: A1->G2, A2->G5, A3->G4, A4->G1\nC: A1->G6, A2->G1, A3->G4, A4->G2\nD: A1->G5, A2->G4, A3->G3, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b8200a86bbc7cdf8a4e6f13a73c77702bd6d6aef0e6249858945696f82d3c305.jpg", "reference_inputs/images/8e90901750b16872fdaa935ea78ac05d94b56c4f0c1df5921fade7be28e20261.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000025", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G2, G3, G4, G1, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [675, 533, 737, 575] (category: vehicle)\n- A2: [602, 439, 619, 454] (category: vehicle)\n- A3: [685, 470, 710, 486] (category: vehicle)\n- A4: [561, 484, 566, 502] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [619, 294, 640, 305] (category: vehicle)\n- G2: [598, 289, 626, 358] (category: pedestrian)\n- G3: [573, 286, 582, 310] (category: pedestrian)\n- G4: [723, 296, 737, 309] (category: vehicle)\n- G1: [763, 357, 826, 575] (category: pedestrian)\n- G6: [738, 261, 882, 337] (category: vehicle)\n\nAnswer options:\nA: A1->G5, A2->G6, A3->G1, A4->G4\nB: A1->G4, A2->G3, A3->G2, A4->G6\nC: A1->G6, A2->G5, A3->G4, A4->G3\nD: A1->G6, A2->G3, A3->G1, A4->G5\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/01986be9f391684280a129ee31aa9ba31926cad35759a3b95020608ed7dfb573.jpg", "reference_inputs/images/15efa56c21d2bc6e235bb1a3ccf4483e7a48fdd00089930beb82826e589d1667.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000028", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G3, G4, G5, G2, G1, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [626, 489, 654, 514] (category: vehicle)\n- A2: [624, 402, 639, 413] (category: vehicle)\n- A3: [595, 418, 610, 430] (category: vehicle)\n- A4: [491, 393, 506, 400] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G3: [723, 295, 760, 341] (category: pedestrian)\n- G4: [475, 294, 504, 303] (category: vehicle)\n- G5: [886, 307, 1023, 376] (category: vehicle)\n- G2: [666, 297, 701, 313] (category: vehicle)\n- G1: [562, 281, 574, 294] (category: vehicle)\n- G6: [702, 290, 736, 307] (category: vehicle)\n\nAnswer options:\nA: A1->G6, A2->G3, A3->G4, A4->G1\nB: A1->G5, A2->G3, A3->G4, A4->G1\nC: A1->G5, A2->G1, A3->G6, A4->G3\nD: A1->G5, A2->G6, A3->G2, A4->G4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a05c6e1fc084f2c02448041fed5d2848f797979541b7617a3dcd203089276c02.jpg", "reference_inputs/images/a64a4eeffc4e041cfee0750c7ca09cfaaa10ab0dd50029a956576eef779531a2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000029", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A4, A1, A2, A3, A5, A6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [886, 307, 1023, 376] (category: vehicle)\n- G2: [475, 294, 504, 303] (category: vehicle)\n- G3: [702, 290, 736, 307] (category: vehicle)\n- G4: [666, 297, 701, 313] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A4: [676, 453, 702, 467] (category: vehicle)\n- A1: [774, 470, 786, 486] (category: pedestrian)\n- A2: [624, 402, 639, 413] (category: vehicle)\n- A3: [491, 393, 506, 400] (category: vehicle)\n- A5: [595, 418, 610, 430] (category: vehicle)\n- A6: [626, 489, 654, 514] (category: vehicle)\n\nAnswer options:\nA: G1->A4, G2->A3, G3->A5, G4->A2\nB: G1->A6, G2->A3, G3->A2, G4->A5\nC: G1->A5, G2->A4, G3->A2, G4->A6\nD: G1->A3, G2->A2, G3->A4, G4->A5\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a64a4eeffc4e041cfee0750c7ca09cfaaa10ab0dd50029a956576eef779531a2.jpg", "reference_inputs/images/a05c6e1fc084f2c02448041fed5d2848f797979541b7617a3dcd203089276c02.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000030", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G3, G1, G6, G5, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [629, 465, 654, 485] (category: vehicle)\n- A2: [658, 435, 679, 450] (category: vehicle)\n- A3: [575, 398, 589, 407] (category: vehicle)\n- A4: [602, 423, 618, 436] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [758, 306, 839, 330] (category: vehicle)\n- G3: [774, 295, 828, 314] (category: vehicle)\n- G1: [607, 289, 623, 312] (category: pedestrian)\n- G6: [664, 294, 698, 311] (category: vehicle)\n- G5: [614, 294, 630, 303] (category: vehicle)\n- G4: [653, 293, 674, 300] (category: vehicle)\n\nAnswer options:\nA: A1->G2, A2->G3, A3->G5, A4->G6\nB: A1->G4, A2->G1, A3->G5, A4->G3\nC: A1->G5, A2->G6, A3->G1, A4->G3\nD: A1->G4, A2->G6, A3->G2, A4->G3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c391ab20637d039794ce863c1b9db7c92b6d6325ce8db865c9cab7217923f7de.jpg", "reference_inputs/images/fa38e654e7d24f439480ca85078b43de575e08da51ef24adf0cb83b923b47e43.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000031", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000031", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A4, A2, A1, A5, A3, A6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [774, 295, 828, 314] (category: vehicle)\n- G2: [614, 294, 630, 303] (category: vehicle)\n- G3: [664, 294, 698, 311] (category: vehicle)\n- G4: [758, 306, 839, 330] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A4: [799, 486, 807, 505] (category: pedestrian)\n- A2: [629, 465, 654, 485] (category: vehicle)\n- A1: [575, 398, 589, 407] (category: vehicle)\n- A5: [602, 423, 618, 436] (category: vehicle)\n- A3: [658, 435, 679, 450] (category: vehicle)\n- A6: [758, 459, 761, 474] (category: pedestrian)\n\nAnswer options:\nA: G1->A4, G2->A3, G3->A6, G4->A1\nB: G1->A4, G2->A3, G3->A5, G4->A1\nC: G1->A4, G2->A5, G3->A6, G4->A1\nD: G1->A3, G2->A1, G3->A5, G4->A2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fa38e654e7d24f439480ca85078b43de575e08da51ef24adf0cb83b923b47e43.jpg", "reference_inputs/images/c391ab20637d039794ce863c1b9db7c92b6d6325ce8db865c9cab7217923f7de.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000033", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000033", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G3, G4, G6, G2, G1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [567, 475, 571, 493] (category: pedestrian)\n- A2: [434, 470, 454, 490] (category: vehicle)\n- A3: [454, 424, 467, 437] (category: vehicle)\n- A4: [506, 437, 519, 453] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [644, 289, 654, 330] (category: pedestrian)\n- G3: [503, 289, 522, 302] (category: vehicle)\n- G4: [502, 298, 527, 315] (category: vehicle)\n- G6: [413, 294, 438, 311] (category: vehicle)\n- G2: [585, 286, 590, 310] (category: pedestrian)\n- G1: [442, 291, 458, 303] (category: vehicle)\n\nAnswer options:\nA: A1->G6, A2->G1, A3->G3, A4->G4\nB: A1->G6, A2->G3, A3->G5, A4->G4\nC: A1->G2, A2->G6, A3->G1, A4->G3\nD: A1->G6, A2->G4, A3->G1, A4->G3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/58535b94fbff0387d3d936599a30966b8283f1f79f82b2557b1e6cb15cb992b7.jpg", "reference_inputs/images/d364b0e3541250a939545692adbf2b7f5dd4eda4eed7a82b53526764b5620522.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000034", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000034", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A1, A2, A3, A5, A4, A6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [585, 286, 590, 310] (category: pedestrian)\n- G2: [442, 291, 458, 303] (category: vehicle)\n- G3: [413, 294, 438, 311] (category: vehicle)\n- G4: [502, 298, 527, 315] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A1: [504, 498, 523, 522] (category: vehicle)\n- A2: [434, 470, 454, 490] (category: vehicle)\n- A3: [454, 424, 467, 437] (category: vehicle)\n- A5: [506, 437, 519, 453] (category: vehicle)\n- A4: [567, 475, 571, 493] (category: pedestrian)\n- A6: [509, 397, 519, 408] (category: vehicle)\n\nAnswer options:\nA: G1->A2, G2->A6, G3->A1, G4->A4\nB: G1->A3, G2->A2, G3->A4, G4->A5\nC: G1->A5, G2->A3, G3->A4, G4->A6\nD: G1->A4, G2->A3, G3->A2, G4->A1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d364b0e3541250a939545692adbf2b7f5dd4eda4eed7a82b53526764b5620522.jpg", "reference_inputs/images/58535b94fbff0387d3d936599a30966b8283f1f79f82b2557b1e6cb15cb992b7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000035", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000035", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G1, G6, G3, G2, G5, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [605, 474, 626, 493] (category: vehicle)\n- A2: [723, 558, 730, 575] (category: pedestrian)\n- A3: [450, 498, 455, 517] (category: pedestrian)\n- A4: [362, 516, 366, 538] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G1: [598, 288, 605, 320] (category: pedestrian)\n- G6: [306, 289, 311, 310] (category: pedestrian)\n- G3: [408, 297, 436, 310] (category: vehicle)\n- G2: [254, 287, 262, 311] (category: pedestrian)\n- G5: [179, 286, 187, 313] (category: pedestrian)\n- G4: [505, 291, 528, 308] (category: vehicle)\n\nAnswer options:\nA: A1->G2, A2->G6, A3->G4, A4->G5\nB: A1->G1, A2->G5, A3->G6, A4->G3\nC: A1->G4, A2->G1, A3->G6, A4->G5\nD: A1->G1, A2->G3, A3->G4, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/07ca824b99b314bd94d57eda78a95cebe85ce30bc8e7d8698d800d83e704a7c3.jpg", "reference_inputs/images/1075720d899aa88c5aa14cb7eeb20344e150199624b9bde5c85be648db3d26cb.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000036", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000036", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A6, A5, A3, A4, A2, A1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [505, 291, 528, 308] (category: vehicle)\n- G2: [179, 286, 187, 313] (category: pedestrian)\n- G3: [598, 288, 605, 320] (category: pedestrian)\n- G4: [306, 289, 311, 310] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A6: [522, 475, 543, 497] (category: vehicle)\n- A5: [362, 516, 366, 538] (category: pedestrian)\n- A3: [605, 474, 626, 493] (category: vehicle)\n- A4: [723, 558, 730, 575] (category: pedestrian)\n- A2: [450, 498, 455, 517] (category: pedestrian)\n- A1: [413, 504, 418, 524] (category: pedestrian)\n\nAnswer options:\nA: G1->A3, G2->A5, G3->A4, G4->A2\nB: G1->A2, G2->A3, G3->A5, G4->A1\nC: G1->A4, G2->A5, G3->A1, G4->A2\nD: G1->A6, G2->A2, G3->A3, G4->A4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1075720d899aa88c5aa14cb7eeb20344e150199624b9bde5c85be648db3d26cb.jpg", "reference_inputs/images/07ca824b99b314bd94d57eda78a95cebe85ce30bc8e7d8698d800d83e704a7c3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000037", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000037", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G4, G1, G6, G5, G2, G3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [838, 86, 880, 110] (category: pedestrian)\n- A2: [707, 266, 737, 279] (category: pedestrian)\n- A3: [507, 197, 518, 213] (category: pedestrian)\n- A4: [577, 276, 590, 288] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G4: [550, 292, 559, 313] (category: pedestrian)\n- G1: [614, 271, 639, 292] (category: pedestrian)\n- G6: [674, 286, 682, 307] (category: pedestrian)\n- G5: [761, 288, 774, 318] (category: pedestrian)\n- G2: [625, 290, 632, 304] (category: pedestrian)\n- G3: [510, 287, 516, 310] (category: pedestrian)\n\nAnswer options:\nA: A1->G6, A2->G2, A3->G3, A4->G4\nB: A1->G4, A2->G2, A3->G1, A4->G3\nC: A1->G3, A2->G5, A3->G4, A4->G1\nD: A1->G1, A2->G5, A3->G2, A4->G4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/38b14bfd106225a2e22f062250eb7d51bddb58c2ae67d4a7ab3f274540ea8a91.jpg", "reference_inputs/images/57ae96560ad6020664e5a95327c1992dcf7337d3980ecf843a71447b788d5bfd.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000039", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000039", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G1, G3, G5, G4, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [401, 382, 414, 395] (category: pedestrian)\n- A2: [887, 546, 935, 574] (category: pedestrian)\n- A3: [538, 361, 582, 459] (category: vehicle)\n- A4: [903, 22, 948, 47] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [583, 289, 592, 308] (category: pedestrian)\n- G1: [375, 286, 380, 301] (category: pedestrian)\n- G3: [484, 308, 526, 340] (category: vehicle)\n- G5: [636, 289, 655, 330] (category: pedestrian)\n- G4: [251, 288, 262, 319] (category: pedestrian)\n- G6: [334, 302, 382, 326] (category: vehicle)\n\nAnswer options:\nA: A1->G4, A2->G5, A3->G6, A4->G2\nB: A1->G5, A2->G2, A3->G6, A4->G3\nC: A1->G5, A2->G2, A3->G4, A4->G1\nD: A1->G3, A2->G5, A3->G2, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/71627e38503f8fa008a5c48ecbf7cb63719a3f00c0377de6c1e90fbd4fd13671.jpg", "reference_inputs/images/18af91d51cab25cd269148c24de3e6c462001bcbe4cdf2ad120569133edaefec.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000040", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000040", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A5, A6, A4, A2, A1, A3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [583, 289, 592, 308] (category: pedestrian)\n- G2: [251, 288, 262, 319] (category: pedestrian)\n- G3: [636, 289, 655, 330] (category: pedestrian)\n- G4: [334, 302, 382, 326] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A5: [360, 305, 382, 325] (category: pedestrian)\n- A6: [903, 22, 948, 47] (category: pedestrian)\n- A4: [401, 382, 414, 395] (category: pedestrian)\n- A2: [887, 546, 935, 574] (category: pedestrian)\n- A1: [538, 361, 582, 459] (category: vehicle)\n- A3: [730, 517, 782, 575] (category: vehicle)\n\nAnswer options:\nA: G1->A3, G2->A2, G3->A5, G4->A6\nB: G1->A3, G2->A1, G3->A2, G4->A6\nC: G1->A6, G2->A4, G3->A2, G4->A1\nD: G1->A3, G2->A2, G3->A6, G4->A4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/18af91d51cab25cd269148c24de3e6c462001bcbe4cdf2ad120569133edaefec.jpg", "reference_inputs/images/71627e38503f8fa008a5c48ecbf7cb63719a3f00c0377de6c1e90fbd4fd13671.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000041", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000041", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G4, G6, G2, G1, G3, G5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [621, 518, 659, 546] (category: vehicle)\n- A2: [371, 435, 393, 446] (category: vehicle)\n- A3: [624, 535, 671, 575] (category: vehicle)\n- A4: [510, 510, 534, 534] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G4: [805, 308, 887, 325] (category: vehicle)\n- G6: [461, 334, 557, 391] (category: vehicle)\n- G2: [811, 298, 872, 312] (category: vehicle)\n- G1: [760, 302, 815, 317] (category: vehicle)\n- G3: [893, 300, 958, 317] (category: vehicle)\n- G5: [980, 298, 1023, 315] (category: vehicle)\n\nAnswer options:\nA: A1->G6, A2->G1, A3->G3, A4->G5\nB: A1->G6, A2->G5, A3->G4, A4->G1\nC: A1->G1, A2->G6, A3->G2, A4->G4\nD: A1->G2, A2->G5, A3->G1, A4->G4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d76465ef7fbd5baaf0fe865c9d44135d8dba217611268bbefb6586e6993147b2.jpg", "reference_inputs/images/c52d3ed15c8ad4699b4bc938831c5b06416414b65cdf6d09847efe920d0a1925.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000042", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000042", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A6, A3, A1, A4, A2, A5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [811, 298, 872, 312] (category: vehicle)\n- G2: [805, 308, 887, 325] (category: vehicle)\n- G3: [760, 302, 815, 317] (category: vehicle)\n- G4: [461, 334, 557, 391] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A6: [371, 435, 393, 446] (category: vehicle)\n- A3: [510, 510, 534, 534] (category: vehicle)\n- A1: [621, 518, 659, 546] (category: vehicle)\n- A4: [624, 535, 671, 575] (category: vehicle)\n- A2: [439, 561, 487, 575] (category: vehicle)\n- A5: [523, 447, 558, 462] (category: vehicle)\n\nAnswer options:\nA: G1->A4, G2->A3, G3->A1, G4->A6\nB: G1->A5, G2->A3, G3->A4, G4->A2\nC: G1->A5, G2->A2, G3->A1, G4->A3\nD: G1->A3, G2->A4, G3->A2, G4->A5\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c52d3ed15c8ad4699b4bc938831c5b06416414b65cdf6d09847efe920d0a1925.jpg", "reference_inputs/images/d76465ef7fbd5baaf0fe865c9d44135d8dba217611268bbefb6586e6993147b2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000043", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000043", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G3, G1, G6, G5, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [570, 431, 584, 451] (category: pedestrian)\n- A2: [654, 116, 670, 134] (category: pedestrian)\n- A3: [649, 530, 682, 550] (category: pedestrian)\n- A4: [809, 134, 870, 243] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [561, 287, 572, 327] (category: pedestrian)\n- G3: [807, 286, 814, 307] (category: pedestrian)\n- G1: [530, 290, 534, 307] (category: pedestrian)\n- G6: [689, 301, 739, 318] (category: vehicle)\n- G5: [656, 293, 690, 335] (category: pedestrian)\n- G4: [592, 292, 598, 312] (category: pedestrian)\n\nAnswer options:\nA: A1->G1, A2->G2, A3->G4, A4->G3\nB: A1->G4, A2->G2, A3->G5, A4->G1\nC: A1->G3, A2->G2, A3->G4, A4->G5\nD: A1->G2, A2->G4, A3->G5, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/05b6f20e4d9f1346db76bff920938e567e5ec7a0b5c20003a050d61e4e631ef7.jpg", "reference_inputs/images/153c42acc503b7f49f9717f56313e3f98e87cf0cd70405ca91151993eafe3310.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000044", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000044", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G4, G5, G3, G6, G1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [434, 283, 456, 294] (category: pedestrian)\n- A2: [700, 43, 723, 70] (category: pedestrian)\n- A3: [557, 519, 590, 540] (category: pedestrian)\n- A4: [598, 283, 621, 294] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [263, 294, 271, 312] (category: pedestrian)\n- G4: [423, 290, 428, 311] (category: pedestrian)\n- G5: [353, 290, 363, 308] (category: pedestrian)\n- G3: [334, 290, 337, 307] (category: pedestrian)\n- G6: [337, 290, 342, 313] (category: pedestrian)\n- G1: [546, 292, 553, 318] (category: pedestrian)\n\nAnswer options:\nA: A1->G1, A2->G4, A3->G2, A4->G5\nB: A1->G6, A2->G3, A3->G4, A4->G1\nC: A1->G4, A2->G2, A3->G5, A4->G6\nD: A1->G6, A2->G5, A3->G3, A4->G4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ad6fe422531a93b9485098c6c3e4bc860f8aeb03907c9f838ade835d139050c2.jpg", "reference_inputs/images/d2fdc4683ecf48ad9dd5cf6d1739017f93cacc73846c598a25716dd0416a30a8.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000046", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000046", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A3, A6, A5, A4, A1, A2. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [297, 289, 316, 357] (category: pedestrian)\n- G2: [394, 290, 410, 338] (category: pedestrian)\n- G3: [670, 291, 696, 297] (category: vehicle)\n- G4: [551, 313, 633, 375] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A3: [688, 363, 706, 370] (category: vehicle)\n- A6: [308, 530, 314, 553] (category: pedestrian)\n- A5: [588, 374, 608, 379] (category: vehicle)\n- A4: [682, 374, 702, 380] (category: vehicle)\n- A1: [391, 533, 421, 569] (category: vehicle)\n- A2: [347, 505, 354, 525] (category: pedestrian)\n\nAnswer options:\nA: G1->A6, G2->A1, G3->A4, G4->A3\nB: G1->A3, G2->A5, G3->A2, G4->A4\nC: G1->A6, G2->A1, G3->A4, G4->A5\nD: G1->A6, G2->A2, G3->A5, G4->A1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ef1c9e15657d46d314b74cd595a8bf5a6aa2686d5fedec13ec8318fc99ba7f6c.jpg", "reference_inputs/images/5bc11b7f2e38ac137619a4800f8fe2d0896a00695683ac0192ea35c976912819.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000047", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000047", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A1, A5, A6, A2, A3, A4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [420, 297, 444, 342] (category: pedestrian)\n- G2: [279, 321, 436, 402] (category: vehicle)\n- G3: [608, 310, 835, 409] (category: vehicle)\n- G4: [998, 294, 1018, 350] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A1: [965, 14, 1003, 59] (category: pedestrian)\n- A5: [739, 0, 842, 76] (category: vehicle)\n- A6: [373, 0, 436, 35] (category: vehicle)\n- A2: [382, 126, 482, 225] (category: vehicle)\n- A3: [917, 504, 962, 535] (category: pedestrian)\n- A4: [731, 166, 840, 269] (category: vehicle)\n\nAnswer options:\nA: G1->A6, G2->A5, G3->A2, G4->A4\nB: G1->A2, G2->A1, G3->A4, G4->A5\nC: G1->A3, G2->A6, G3->A1, G4->A4\nD: G1->A1, G2->A5, G3->A4, G4->A3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9d7dd89769e249a4eddd5467312146be0428ac4007260c0ea7d8f0137e5d7c62.jpg", "reference_inputs/images/1aedfd6862af453cdd58f1227e627eccd860b652ccd30b1767d50432e1f9439a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000048", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000048", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G6, G5, G3, G4, G1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [91, 530, 130, 559] (category: pedestrian)\n- A2: [281, 490, 334, 575] (category: vehicle)\n- A3: [138, 421, 173, 437] (category: pedestrian)\n- A4: [482, 43, 526, 143] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [430, 294, 447, 304] (category: vehicle)\n- G6: [214, 289, 225, 323] (category: pedestrian)\n- G5: [365, 297, 404, 319] (category: vehicle)\n- G3: [500, 294, 518, 306] (category: vehicle)\n- G4: [319, 287, 326, 310] (category: pedestrian)\n- G1: [267, 287, 275, 314] (category: pedestrian)\n\nAnswer options:\nA: A1->G2, A2->G5, A3->G3, A4->G6\nB: A1->G6, A2->G5, A3->G2, A4->G4\nC: A1->G1, A2->G5, A3->G4, A4->G3\nD: A1->G1, A2->G2, A3->G4, A4->G3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5479ce473b2ca0acb6639443c9e9e1ea4a026e7a50028f282cb78759b8584efd.jpg", "reference_inputs/images/fdf2d889c2dffa46fe69ae36b3d498a5696e384b59b17da14b65c985f03e64a3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000049", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000049", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G1, G5, G3, G4, G2, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [585, 400, 596, 410] (category: vehicle)\n- A2: [694, 473, 724, 490] (category: vehicle)\n- A3: [594, 421, 611, 433] (category: vehicle)\n- A4: [457, 517, 502, 538] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G1: [414, 304, 496, 325] (category: vehicle)\n- G5: [578, 279, 596, 297] (category: vehicle)\n- G3: [600, 290, 615, 300] (category: vehicle)\n- G4: [765, 299, 817, 314] (category: vehicle)\n- G2: [615, 294, 638, 303] (category: vehicle)\n- G6: [978, 302, 1023, 338] (category: vehicle)\n\nAnswer options:\nA: A1->G3, A2->G4, A3->G2, A4->G1\nB: A1->G2, A2->G6, A3->G1, A4->G5\nC: A1->G3, A2->G5, A3->G1, A4->G6\nD: A1->G5, A2->G1, A3->G3, A4->G4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/09e4d63e55bd5f22dd2cec7666efaac4ee4d83856b562dc1c9726bfdc0600797.jpg", "reference_inputs/images/8c822065e6a4a3137d25a6abc18bb3c2c267e21f7d088cad1a3a17dbee7b5c2a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000050", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000050", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A4, A1, A2, A3, A6, A5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [578, 279, 596, 297] (category: vehicle)\n- G2: [600, 290, 615, 300] (category: vehicle)\n- G3: [765, 299, 817, 314] (category: vehicle)\n- G4: [615, 294, 638, 303] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A4: [585, 400, 596, 410] (category: vehicle)\n- A1: [457, 517, 502, 538] (category: vehicle)\n- A2: [594, 421, 611, 433] (category: vehicle)\n- A3: [779, 567, 816, 575] (category: vehicle)\n- A6: [694, 473, 724, 490] (category: vehicle)\n- A5: [568, 370, 585, 390] (category: vehicle)\n\nAnswer options:\nA: G1->A5, G2->A6, G3->A3, G4->A2\nB: G1->A5, G2->A4, G3->A6, G4->A2\nC: G1->A1, G2->A3, G3->A2, G4->A5\nD: G1->A5, G2->A2, G3->A1, G4->A4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8c822065e6a4a3137d25a6abc18bb3c2c267e21f7d088cad1a3a17dbee7b5c2a.jpg", "reference_inputs/images/09e4d63e55bd5f22dd2cec7666efaac4ee4d83856b562dc1c9726bfdc0600797.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000051", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000051", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G5, G1, G3, G6, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [620, 218, 637, 234] (category: pedestrian)\n- A2: [100, 486, 213, 564] (category: vehicle)\n- A3: [652, 297, 760, 382] (category: vehicle)\n- A4: [769, 289, 875, 368] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [540, 291, 598, 310] (category: vehicle)\n- G5: [213, 302, 360, 349] (category: vehicle)\n- G1: [400, 294, 449, 311] (category: vehicle)\n- G3: [392, 300, 430, 315] (category: vehicle)\n- G6: [116, 323, 309, 383] (category: vehicle)\n- G4: [305, 287, 310, 304] (category: pedestrian)\n\nAnswer options:\nA: A1->G5, A2->G2, A3->G6, A4->G1\nB: A1->G3, A2->G6, A3->G5, A4->G4\nC: A1->G6, A2->G5, A3->G1, A4->G4\nD: A1->G4, A2->G6, A3->G3, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0671ae8ee169440e6627d92e84b2ba53644dd471a3dc9097fb5bbae57d101527.jpg", "reference_inputs/images/cd09321292f0788f78323c26abbf46a241c86ce743f184ba8aece6a49f533751.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000052", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000052", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A2, A4, A3, A1, A6, A5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [305, 287, 310, 304] (category: pedestrian)\n- G2: [400, 294, 449, 311] (category: vehicle)\n- G3: [392, 300, 430, 315] (category: vehicle)\n- G4: [116, 323, 309, 383] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A2: [620, 218, 637, 234] (category: pedestrian)\n- A4: [100, 486, 213, 564] (category: vehicle)\n- A3: [470, 373, 577, 454] (category: vehicle)\n- A1: [769, 289, 875, 368] (category: vehicle)\n- A6: [652, 297, 760, 382] (category: vehicle)\n- A5: [196, 450, 314, 539] (category: vehicle)\n\nAnswer options:\nA: G1->A2, G2->A1, G3->A6, G4->A4\nB: G1->A4, G2->A3, G3->A6, G4->A1\nC: G1->A6, G2->A5, G3->A1, G4->A4\nD: G1->A5, G2->A2, G3->A3, G4->A1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cd09321292f0788f78323c26abbf46a241c86ce743f184ba8aece6a49f533751.jpg", "reference_inputs/images/0671ae8ee169440e6627d92e84b2ba53644dd471a3dc9097fb5bbae57d101527.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000053", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000053", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G5, G3, G4, G2, G6, G1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [573, 410, 586, 422] (category: vehicle)\n- A2: [634, 547, 667, 575] (category: vehicle)\n- A3: [534, 506, 556, 534] (category: vehicle)\n- A4: [700, 516, 710, 538] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G5: [578, 289, 594, 300] (category: vehicle)\n- G3: [519, 290, 564, 314] (category: vehicle)\n- G4: [605, 290, 631, 307] (category: vehicle)\n- G2: [966, 290, 986, 342] (category: pedestrian)\n- G6: [661, 301, 705, 322] (category: vehicle)\n- G1: [735, 287, 745, 313] (category: pedestrian)\n\nAnswer options:\nA: A1->G6, A2->G5, A3->G2, A4->G1\nB: A1->G5, A2->G6, A3->G3, A4->G1\nC: A1->G3, A2->G6, A3->G1, A4->G5\nD: A1->G2, A2->G1, A3->G6, A4->G3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7ff9a063ea7e900809d71200db9a76b02c6c06d8d46e1304fb4b04a2199158b9.jpg", "reference_inputs/images/ffeabae965ca87a9a76fa6420a55de21d573a30f19c479c2e2f9f7fc91d2e546.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000056", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000056", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G3, G1, G6, G5, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [138, 486, 174, 510] (category: pedestrian)\n- A2: [653, 66, 677, 90] (category: pedestrian)\n- A3: [649, 333, 667, 346] (category: pedestrian)\n- A4: [695, 122, 721, 140] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [305, 287, 312, 310] (category: pedestrian)\n- G3: [608, 287, 618, 310] (category: pedestrian)\n- G1: [602, 288, 608, 317] (category: pedestrian)\n- G6: [583, 286, 587, 309] (category: pedestrian)\n- G5: [810, 288, 814, 298] (category: pedestrian)\n- G4: [222, 288, 231, 325] (category: pedestrian)\n\nAnswer options:\nA: A1->G4, A2->G6, A3->G1, A4->G3\nB: A1->G4, A2->G3, A3->G2, A4->G6\nC: A1->G1, A2->G5, A3->G6, A4->G3\nD: A1->G1, A2->G5, A3->G4, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b93ab76ddb5a25b217c5c906a07d5efb3a35760afac1f9e5f0eaea7e0b6e0003.jpg", "reference_inputs/images/30142c91b4a377fd8bb01708b286c070258fd5c4a84e7771a9bec6c4b1def4c8.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000057", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000057", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A2, A1, A6, A3, A4, A5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [602, 288, 608, 317] (category: pedestrian)\n- G2: [222, 288, 231, 325] (category: pedestrian)\n- G3: [608, 287, 618, 310] (category: pedestrian)\n- G4: [583, 286, 587, 309] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A2: [138, 486, 174, 510] (category: pedestrian)\n- A1: [658, 562, 674, 575] (category: pedestrian)\n- A6: [89, 122, 131, 140] (category: pedestrian)\n- A3: [695, 122, 721, 140] (category: pedestrian)\n- A4: [653, 66, 677, 90] (category: pedestrian)\n- A5: [649, 333, 667, 346] (category: pedestrian)\n\nAnswer options:\nA: G1->A5, G2->A2, G3->A3, G4->A4\nB: G1->A4, G2->A2, G3->A6, G4->A5\nC: G1->A4, G2->A5, G3->A3, G4->A1\nD: G1->A2, G2->A3, G3->A1, G4->A5\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/30142c91b4a377fd8bb01708b286c070258fd5c4a84e7771a9bec6c4b1def4c8.jpg", "reference_inputs/images/b93ab76ddb5a25b217c5c906a07d5efb3a35760afac1f9e5f0eaea7e0b6e0003.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000058", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000058", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G1, G3, G6, G5, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [959, 277, 998, 302] (category: pedestrian)\n- A2: [730, 417, 839, 520] (category: vehicle)\n- A3: [382, 374, 482, 476] (category: vehicle)\n- A4: [722, 19, 830, 126] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [512, 294, 561, 310] (category: vehicle)\n- G1: [838, 290, 854, 321] (category: pedestrian)\n- G3: [679, 298, 737, 325] (category: vehicle)\n- G6: [735, 296, 813, 342] (category: vehicle)\n- G5: [631, 291, 680, 316] (category: vehicle)\n- G4: [388, 296, 488, 336] (category: vehicle)\n\nAnswer options:\nA: A1->G1, A2->G6, A3->G4, A4->G5\nB: A1->G1, A2->G2, A3->G4, A4->G6\nC: A1->G1, A2->G5, A3->G4, A4->G6\nD: A1->G5, A2->G3, A3->G2, A4->G4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/593aba1a5f50088dcad24abfed915ae977e9337ecee2638c883220b5be6434b8.jpg", "reference_inputs/images/c78a0378bceef9047b848116311ccaa5e549a203c8ea138f175b0306b5e4c7b7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000059", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000059", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A4, A2, A1, A6, A5, A3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [388, 296, 488, 336] (category: vehicle)\n- G2: [735, 296, 813, 342] (category: vehicle)\n- G3: [838, 290, 854, 321] (category: pedestrian)\n- G4: [631, 291, 680, 316] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A4: [730, 417, 839, 520] (category: vehicle)\n- A2: [373, 193, 465, 284] (category: vehicle)\n- A1: [382, 374, 482, 476] (category: vehicle)\n- A6: [722, 19, 830, 126] (category: vehicle)\n- A5: [959, 277, 998, 302] (category: pedestrian)\n- A3: [738, 226, 844, 324] (category: vehicle)\n\nAnswer options:\nA: G1->A1, G2->A4, G3->A5, G4->A6\nB: G1->A2, G2->A5, G3->A3, G4->A1\nC: G1->A1, G2->A6, G3->A2, G4->A5\nD: G1->A3, G2->A5, G3->A1, G4->A4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c78a0378bceef9047b848116311ccaa5e549a203c8ea138f175b0306b5e4c7b7.jpg", "reference_inputs/images/593aba1a5f50088dcad24abfed915ae977e9337ecee2638c883220b5be6434b8.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000060", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000060", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G4, G1, G6, G3, G5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [644, 430, 676, 442] (category: vehicle)\n- A2: [763, 553, 815, 575] (category: vehicle)\n- A3: [730, 490, 767, 510] (category: vehicle)\n- A4: [632, 475, 654, 494] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [979, 300, 1023, 316] (category: vehicle)\n- G4: [886, 298, 953, 312] (category: vehicle)\n- G1: [610, 306, 662, 320] (category: vehicle)\n- G6: [547, 322, 615, 358] (category: vehicle)\n- G3: [833, 302, 889, 315] (category: vehicle)\n- G5: [869, 307, 954, 324] (category: vehicle)\n\nAnswer options:\nA: A1->G1, A2->G2, A3->G4, A4->G5\nB: A1->G5, A2->G2, A3->G4, A4->G6\nC: A1->G1, A2->G2, A3->G3, A4->G5\nD: A1->G5, A2->G4, A3->G1, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d4e07913a5a14259f4b28417bcd90f7478a358c22747ce6cf3e6cfa47b710b39.jpg", "reference_inputs/images/6aeff7f4a3fe7015e7da04d5c9d75664c3247c67da7bc0db28bb73528a27a828.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000061", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000061", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A1, A6, A5, A3, A2, A4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [610, 306, 662, 320] (category: vehicle)\n- G2: [869, 307, 954, 324] (category: vehicle)\n- G3: [979, 300, 1023, 316] (category: vehicle)\n- G4: [833, 302, 889, 315] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A1: [560, 531, 607, 553] (category: vehicle)\n- A6: [763, 553, 815, 575] (category: vehicle)\n- A5: [730, 490, 767, 510] (category: vehicle)\n- A3: [567, 504, 613, 527] (category: vehicle)\n- A2: [644, 430, 676, 442] (category: vehicle)\n- A4: [632, 475, 654, 494] (category: vehicle)\n\nAnswer options:\nA: G1->A1, G2->A2, G3->A4, G4->A6\nB: G1->A2, G2->A4, G3->A6, G4->A5\nC: G1->A1, G2->A3, G3->A2, G4->A6\nD: G1->A4, G2->A5, G3->A3, G4->A6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6aeff7f4a3fe7015e7da04d5c9d75664c3247c67da7bc0db28bb73528a27a828.jpg", "reference_inputs/images/d4e07913a5a14259f4b28417bcd90f7478a358c22747ce6cf3e6cfa47b710b39.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000062", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000062", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G3, G5, G6, G4, G1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [132, 455, 166, 475] (category: pedestrian)\n- A2: [561, 126, 626, 238] (category: vehicle)\n- A3: [295, 390, 356, 500] (category: vehicle)\n- A4: [744, 453, 771, 479] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [493, 291, 498, 310] (category: pedestrian)\n- G3: [558, 296, 579, 311] (category: vehicle)\n- G5: [459, 294, 478, 332] (category: pedestrian)\n- G6: [736, 288, 742, 306] (category: pedestrian)\n- G4: [677, 301, 718, 315] (category: vehicle)\n- G1: [620, 297, 642, 306] (category: vehicle)\n\nAnswer options:\nA: A1->G5, A2->G2, A3->G4, A4->G1\nB: A1->G5, A2->G4, A3->G1, A4->G3\nC: A1->G4, A2->G2, A3->G5, A4->G1\nD: A1->G2, A2->G1, A3->G3, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5c0e4b7fd28d747f1c4c10d28e50b31f60ce5a6fbd644a44f8b265ae00c19b94.jpg", "reference_inputs/images/9a8ae0b3d983add7ddedd2d91a1213471f9aed9b1ff95a593c93278990174473.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000063", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000063", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G4, G6, G1, G3, G5. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [621, 401, 634, 411] (category: vehicle)\n- A2: [450, 489, 454, 506] (category: pedestrian)\n- A3: [574, 378, 590, 401] (category: vehicle)\n- A4: [569, 480, 579, 496] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [599, 289, 629, 317] (category: pedestrian)\n- G4: [544, 286, 555, 314] (category: pedestrian)\n- G6: [590, 278, 617, 300] (category: vehicle)\n- G1: [654, 291, 675, 298] (category: vehicle)\n- G3: [622, 292, 645, 304] (category: vehicle)\n- G5: [404, 290, 412, 321] (category: pedestrian)\n\nAnswer options:\nA: A1->G3, A2->G4, A3->G1, A4->G6\nB: A1->G1, A2->G5, A3->G6, A4->G2\nC: A1->G3, A2->G1, A3->G4, A4->G5\nD: A1->G2, A2->G1, A3->G3, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/39f792258ee5da7ad110e4a0028f5a09df7eaabf2742743885a78cb2856015d7.jpg", "reference_inputs/images/702dd7ba3d6815564455245f3199e57483d587890f24e75a7c679226bfe87823.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000064", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000064", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A2, A6, A5, A1, A3, A4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [654, 291, 675, 298] (category: vehicle)\n- G2: [404, 290, 412, 321] (category: pedestrian)\n- G3: [599, 289, 629, 317] (category: pedestrian)\n- G4: [590, 278, 617, 300] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A2: [450, 489, 454, 506] (category: pedestrian)\n- A6: [621, 401, 634, 411] (category: vehicle)\n- A5: [532, 464, 540, 481] (category: pedestrian)\n- A1: [569, 480, 579, 496] (category: pedestrian)\n- A3: [574, 378, 590, 401] (category: vehicle)\n- A4: [594, 414, 607, 426] (category: vehicle)\n\nAnswer options:\nA: G1->A2, G2->A3, G3->A6, G4->A1\nB: G1->A6, G2->A2, G3->A1, G4->A3\nC: G1->A3, G2->A4, G3->A1, G4->A2\nD: G1->A2, G2->A5, G3->A4, G4->A3\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/702dd7ba3d6815564455245f3199e57483d587890f24e75a7c679226bfe87823.jpg", "reference_inputs/images/39f792258ee5da7ad110e4a0028f5a09df7eaabf2742743885a78cb2856015d7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000065", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000065", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G1, G4, G3, G2, G5, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [578, 470, 590, 494] (category: pedestrian)\n- A2: [115, 372, 223, 429] (category: vehicle)\n- A3: [517, 375, 537, 388] (category: pedestrian)\n- A4: [928, 494, 1023, 552] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G1: [518, 294, 534, 328] (category: pedestrian)\n- G4: [576, 294, 585, 336] (category: pedestrian)\n- G3: [173, 307, 274, 331] (category: vehicle)\n- G2: [558, 286, 565, 314] (category: pedestrian)\n- G5: [514, 278, 526, 297] (category: vehicle)\n- G6: [939, 312, 1023, 344] (category: vehicle)\n\nAnswer options:\nA: A1->G4, A2->G1, A3->G3, A4->G2\nB: A1->G4, A2->G3, A3->G1, A4->G6\nC: A1->G4, A2->G2, A3->G5, A4->G6\nD: A1->G5, A2->G2, A3->G1, A4->G4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/40f61ea9edb503112601c62b3f52b58bfbfc3755589baafaa08d713e3b776977.jpg", "reference_inputs/images/12ff10e2b317b9f7c0aff4ccda2d4142e0e5b16c61dab7dfee15d3580d0725a9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000067", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000067", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A5, A1, A4, A6, A2, A3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [540, 297, 574, 304] (category: vehicle)\n- G2: [696, 302, 730, 315] (category: vehicle)\n- G3: [795, 298, 835, 307] (category: vehicle)\n- G4: [694, 294, 730, 305] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A5: [608, 438, 625, 452] (category: vehicle)\n- A1: [618, 393, 632, 400] (category: vehicle)\n- A4: [707, 486, 742, 506] (category: vehicle)\n- A6: [488, 406, 506, 414] (category: vehicle)\n- A2: [553, 346, 564, 359] (category: vehicle)\n- A3: [639, 418, 658, 430] (category: vehicle)\n\nAnswer options:\nA: G1->A6, G2->A5, G3->A3, G4->A4\nB: G1->A5, G2->A2, G3->A3, G4->A6\nC: G1->A1, G2->A6, G3->A5, G4->A3\nD: G1->A6, G2->A3, G3->A1, G4->A5\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d3734fe6806cf1bea75aedd8fed2e857ca548f7986cb59e09507251fc79338c3.jpg", "reference_inputs/images/4eb701386ccf91f87d8153ba6786d69689da1ca369180d1c62781fc6f8b641d2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000068", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000068", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G3, G5, G4, G1, G6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [919, 498, 962, 524] (category: pedestrian)\n- A2: [870, 362, 906, 374] (category: pedestrian)\n- A3: [898, 470, 935, 488] (category: pedestrian)\n- A4: [575, 263, 588, 273] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [663, 294, 674, 315] (category: pedestrian)\n- G3: [642, 296, 653, 336] (category: pedestrian)\n- G5: [565, 314, 654, 575] (category: pedestrian)\n- G4: [810, 294, 828, 350] (category: pedestrian)\n- G1: [642, 337, 647, 359] (category: pedestrian)\n- G6: [786, 298, 806, 352] (category: pedestrian)\n\nAnswer options:\nA: A1->G4, A2->G2, A3->G6, A4->G5\nB: A1->G1, A2->G4, A3->G5, A4->G3\nC: A1->G4, A2->G5, A3->G2, A4->G3\nD: A1->G5, A2->G2, A3->G3, A4->G6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4f1d592425f43ddd44ef33551b00627e3bfa2b20112e736d3202b2b1e08e70c9.jpg", "reference_inputs/images/a04ee1aa01447531e8a5ff396360422102cee8a5d040a61a205db0f05731279d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000069", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000069", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A2, A5, A1, A4, A6, A3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [565, 314, 654, 575] (category: pedestrian)\n- G2: [810, 294, 828, 350] (category: pedestrian)\n- G3: [786, 298, 806, 352] (category: pedestrian)\n- G4: [663, 294, 674, 315] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A2: [898, 470, 935, 488] (category: pedestrian)\n- A5: [561, 158, 581, 173] (category: pedestrian)\n- A1: [919, 498, 962, 524] (category: pedestrian)\n- A4: [1002, 384, 1023, 394] (category: pedestrian)\n- A6: [575, 263, 588, 273] (category: pedestrian)\n- A3: [870, 362, 906, 374] (category: pedestrian)\n\nAnswer options:\nA: G1->A2, G2->A5, G3->A6, G4->A3\nB: G1->A6, G2->A1, G3->A2, G4->A3\nC: G1->A5, G2->A1, G3->A2, G4->A3\nD: G1->A1, G2->A5, G3->A2, G4->A6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a04ee1aa01447531e8a5ff396360422102cee8a5d040a61a205db0f05731279d.jpg", "reference_inputs/images/4f1d592425f43ddd44ef33551b00627e3bfa2b20112e736d3202b2b1e08e70c9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000070", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000070", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G2, G3, G1, G6, G5, G4. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [360, 305, 382, 325] (category: pedestrian)\n- A2: [401, 382, 414, 395] (category: pedestrian)\n- A3: [903, 22, 948, 47] (category: pedestrian)\n- A4: [887, 546, 935, 574] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G2: [673, 290, 703, 345] (category: pedestrian)\n- G3: [593, 290, 604, 311] (category: pedestrian)\n- G1: [192, 288, 206, 327] (category: pedestrian)\n- G6: [582, 286, 587, 302] (category: pedestrian)\n- G5: [202, 288, 210, 323] (category: pedestrian)\n- G4: [332, 287, 336, 306] (category: pedestrian)\n\nAnswer options:\nA: A1->G5, A2->G1, A3->G3, A4->G2\nB: A1->G3, A2->G6, A3->G1, A4->G2\nC: A1->G2, A2->G5, A3->G6, A4->G3\nD: A1->G5, A2->G3, A3->G1, A4->G2\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/71627e38503f8fa008a5c48ecbf7cb63719a3f00c0377de6c1e90fbd4fd13671.jpg", "reference_inputs/images/d3558a936a65613cbbec820d11b51cc2ac958af87155f8146abf0217b9691467.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000071", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000071", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A5, A2, A4, A1, A3, A6. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [202, 288, 210, 323] (category: pedestrian)\n- G2: [192, 288, 206, 327] (category: pedestrian)\n- G3: [593, 290, 604, 311] (category: pedestrian)\n- G4: [673, 290, 703, 345] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A5: [401, 382, 414, 395] (category: pedestrian)\n- A2: [538, 361, 582, 459] (category: vehicle)\n- A4: [360, 305, 382, 325] (category: pedestrian)\n- A1: [903, 22, 948, 47] (category: pedestrian)\n- A3: [887, 546, 935, 574] (category: pedestrian)\n- A6: [730, 517, 782, 575] (category: vehicle)\n\nAnswer options:\nA: G1->A4, G2->A5, G3->A1, G4->A3\nB: G1->A6, G2->A5, G3->A2, G4->A1\nC: G1->A2, G2->A3, G3->A6, G4->A4\nD: G1->A5, G2->A2, G3->A3, G4->A6\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d3558a936a65613cbbec820d11b51cc2ac958af87155f8146abf0217b9691467.jpg", "reference_inputs/images/71627e38503f8fa008a5c48ecbf7cb63719a3f00c0377de6c1e90fbd4fd13671.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000072", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000072", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the source targets.\nImage 2 is the AERIAL image and contains the candidate targets.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked 4 source targets in the ground view as G1, G2, G3, G4. In your own aerial view, there are 6 labeled candidate targets: A5, A4, A2, A6, A1, A3. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (GROUND):\n- G1: [531, 297, 596, 325] (category: vehicle)\n- G2: [198, 292, 210, 331] (category: pedestrian)\n- G3: [880, 295, 903, 355] (category: pedestrian)\n- G4: [588, 313, 762, 369] (category: vehicle)\n\nCandidate labeled bboxes in Image 2 (AERIAL):\n- A5: [542, 0, 642, 47] (category: vehicle)\n- A4: [550, 291, 658, 347] (category: vehicle)\n- A2: [562, 408, 670, 458] (category: vehicle)\n- A6: [752, 436, 774, 454] (category: pedestrian)\n- A1: [547, 94, 646, 143] (category: vehicle)\n- A3: [173, 236, 202, 253] (category: pedestrian)\n\nAnswer options:\nA: G1->A6, G2->A5, G3->A4, G4->A3\nB: G1->A1, G2->A3, G3->A6, G4->A2\nC: G1->A4, G2->A3, G3->A6, G4->A1\nD: G1->A1, G2->A6, G3->A2, G4->A4\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/570aa2e1dcd552d6dd226263519545311ebc7a1e09b440e10a71bb754cf35d99.jpg", "reference_inputs/images/ff808de7f4377d64ed2599cf41e244006e312f326c271da6d10d4ea92c0320fb.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_matching/Town1_ground_aerial_v2/000073", "task_name": "multi_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000073", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the source targets.\nImage 2 is the GROUND image and contains the candidate targets.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked 4 source targets in the aerial view as A1, A2, A3, A4. In your own ground view, there are 6 labeled candidate targets: G4, G3, G2, G5, G6, G1. Exactly 4 of them correspond to the source targets, and the remaining candidates are distractors. Determine the full one-to-one correspondence from the source labels to the correct target labels.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n- Labels such as A1, A2, G1, G2 are symbolic IDs for boxes, not ranking or order hints.\n\nSource labeled bboxes in Image 1 (AERIAL):\n- A1: [70, 303, 106, 315] (category: pedestrian)\n- A2: [270, 298, 363, 407] (category: vehicle)\n- A3: [614, 148, 709, 258] (category: vehicle)\n- A4: [969, 90, 1016, 114] (category: pedestrian)\n\nCandidate labeled bboxes in Image 2 (GROUND):\n- G4: [490, 296, 506, 333] (category: pedestrian)\n- G3: [752, 302, 795, 318] (category: vehicle)\n- G2: [559, 292, 563, 310] (category: pedestrian)\n- G5: [830, 287, 837, 307] (category: pedestrian)\n- G6: [655, 310, 705, 334] (category: vehicle)\n- G1: [702, 296, 726, 306] (category: vehicle)\n\nAnswer options:\nA: A1->G4, A2->G1, A3->G2, A4->G6\nB: A1->G2, A2->G3, A3->G6, A4->G5\nC: A1->G4, A2->G6, A3->G3, A4->G5\nD: A1->G5, A2->G4, A3->G2, A4->G1\n\nTask: determine the complete one-to-one correspondence between the labeled boxes in Image 1 and the labeled boxes in Image 2, then choose the single option whose full mapping is entirely correct.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2f5b9c5803ca4db366c311011221b5a28998dfc282cf3612aa0248beaeaaa7a9.jpg", "reference_inputs/images/a54b9a361aecb04345c29fc9d533f9763dfc3b55579870f0ad6398d05d7a8e84.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000001", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [352, 522, 381, 547]\n- T1: [140, 542, 158, 563]\n- T2: [244, 503, 271, 530]\n\nImage 2 (GROUND) labeled targets:\n- R: [319, 278, 386, 302]\n- T1: [684, 262, 703, 297]\n- T2: [528, 274, 602, 309]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0afbd9401679e8a5f9f739e851e2d6565ca0847966b9c870ff25e6bd5f0227c3.jpg", "reference_inputs/images/a37278d6f61a09212963b84767982a4ff7065abe6dae528d356c579d594d5b87.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000002", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [178, 453, 197, 472]\n- T1: [156, 423, 176, 436]\n- T2: [246, 453, 277, 500]\n\nImage 2 (GROUND) labeled targets:\n- R: [470, 278, 489, 292]\n- T1: [454, 282, 470, 290]\n- T2: [537, 250, 579, 294]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/65bc64984b2063a131f48fe526cf6a7b07da1d78c38097687558bd599f26a609.jpg", "reference_inputs/images/326e309493b74656ab397a5c7fe4a7d782f7c238e5d6a88d3395323169ed9146.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000003", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [486, 418, 498, 430]\n- T1: [506, 526, 532, 562]\n- T2: [404, 543, 438, 575]\n\nImage 2 (GROUND) labeled targets:\n- R: [482, 281, 495, 290]\n- T1: [500, 268, 538, 298]\n- T2: [407, 282, 414, 298]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2dc7235177eab99e00758787c60e4230c1c1263feedc79745fdcd2eae1d0ed46.jpg", "reference_inputs/images/9ad8a50cccc69d4d19587748ca75a29e617773e0bff87ecdaec8ddeb22d583f1.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000004", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [616, 494, 622, 515]\n- T1: [643, 550, 654, 566]\n- T2: [642, 482, 646, 501]\n\nImage 2 (GROUND) labeled targets:\n- R: [334, 218, 361, 318]\n- T1: [797, 231, 848, 318]\n- T2: [264, 234, 283, 307]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bea921d528598e67ce2c2c61b17817b00fbeb16fe4424686f86ef1b39ff8552c.jpg", "reference_inputs/images/0434ee60b659113cb97e064670e8b02e9955eeff7f24ccedb8f970345927aa67.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000005", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [371, 114, 428, 227]\n- T1: [161, 388, 197, 406]\n- T2: [135, 78, 182, 100]\n\nImage 2 (GROUND) labeled targets:\n- R: [551, 241, 1023, 399]\n- T1: [245, 243, 272, 312]\n- T2: [710, 239, 730, 270]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6cc5f2a32deb5f6d718ce65a2f84ff9863030ca7431c9373a3d6d1bd209ce13d.jpg", "reference_inputs/images/7c40dda52bb9866996ccd7d539527860ee133d45134812dab45eff791b1e2824.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000006", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [487, 318, 538, 429]\n- T1: [512, 0, 564, 93]\n- T2: [667, 402, 730, 514]\n\nImage 2 (GROUND) labeled targets:\n- R: [318, 278, 354, 296]\n- T1: [374, 279, 399, 294]\n- T2: [399, 282, 434, 298]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f6a79034b298f8ff2f1b93aaa565a24635361b61b7454da8eb86027d450c4efa.jpg", "reference_inputs/images/136f851f4ce6fede76cc50fdc3fc6048d01d9a5ec12763732777c13fbb606be9.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000008", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 4 shared labeled targets across the two images, including 3 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2, T3.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [280, 538, 297, 561]\n- T1: [632, 510, 637, 531]\n- T2: [534, 448, 550, 465]\n- T3: [679, 565, 686, 575]\n\nImage 2 (GROUND) labeled targets:\n- R: [291, 270, 302, 292]\n- T1: [628, 272, 633, 291]\n- T2: [534, 280, 550, 290]\n- T3: [673, 268, 679, 293]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1 -> T3\nC: T3 -> T2 -> T1\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a4cf592a72e7c039e89f2ce68a4f098467de07ca020316c2c156372cb9a1cfc5.jpg", "reference_inputs/images/9762b261fdab1ae82f5704a59390f7a7a5f789c680e4cf37c3c11ba8d2d2fc0f.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000009", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [499, 388, 517, 412]\n- T1: [195, 548, 205, 574]\n- T2: [300, 546, 306, 564]\n\nImage 2 (GROUND) labeled targets:\n- R: [500, 270, 517, 289]\n- T1: [205, 268, 218, 292]\n- T2: [307, 275, 314, 292]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a4cf592a72e7c039e89f2ce68a4f098467de07ca020316c2c156372cb9a1cfc5.jpg", "reference_inputs/images/9762b261fdab1ae82f5704a59390f7a7a5f789c680e4cf37c3c11ba8d2d2fc0f.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000010", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [394, 463, 398, 482]\n- T1: [406, 557, 439, 575]\n- T2: [510, 570, 532, 575]\n\nImage 2 (GROUND) labeled targets:\n- R: [250, 257, 262, 297]\n- T1: [0, 257, 188, 352]\n- T2: [485, 233, 753, 396]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/54a45b98db6cf9dcae184341b11c3ada9efb63b5c8c0222a21ea269ead7746f1.jpg", "reference_inputs/images/22e82f3f5da3ca8e0a41507973a1054fc81a0436bd364afb7118546fe3ec6147.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000011", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [466, 391, 490, 399]\n- T1: [480, 355, 492, 372]\n- T2: [446, 446, 463, 464]\n\nImage 2 (GROUND) labeled targets:\n- R: [446, 282, 480, 290]\n- T1: [469, 271, 486, 289]\n- T2: [381, 274, 420, 298]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/54a45b98db6cf9dcae184341b11c3ada9efb63b5c8c0222a21ea269ead7746f1.jpg", "reference_inputs/images/22e82f3f5da3ca8e0a41507973a1054fc81a0436bd364afb7118546fe3ec6147.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000012", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [454, 428, 470, 447]\n- T1: [494, 430, 516, 466]\n- T2: [458, 468, 480, 490]\n\nImage 2 (GROUND) labeled targets:\n- R: [390, 267, 424, 285]\n- T1: [469, 223, 529, 299]\n- T2: [343, 273, 424, 306]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0338e6668ccc2980e70ebd4a9e32d2deccdc84389a69907b91810ac1184821af.jpg", "reference_inputs/images/a4ff943138300e14df5dde6b5b4d48e1550533ac77ffe671ed28e02e96353024.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000013", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [468, 399, 480, 408]\n- T1: [430, 458, 464, 520]\n- T2: [510, 466, 526, 485]\n\nImage 2 (GROUND) labeled targets:\n- R: [434, 282, 454, 292]\n- T1: [13, 152, 344, 330]\n- T2: [502, 270, 558, 309]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/09d6bbd65d5060eae93487145b4ecd97a46eec4da32909caff0ec56456d840de.jpg", "reference_inputs/images/b44ca937189d4b9529c968b0f3bea457bb9b34d3c8f1a0061bfd35bda0a95e8c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000015", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [328, 180, 351, 194]\n- T1: [322, 230, 372, 331]\n- T2: [162, 121, 205, 139]\n\nImage 2 (GROUND) labeled targets:\n- R: [398, 214, 440, 347]\n- T1: [0, 250, 310, 358]\n- T2: [539, 235, 560, 311]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0270d8b538279fc1cc2d42738a8e13de18ac9681608ba8b8b64e1d8f6abef7bd.jpg", "reference_inputs/images/7bb72823dd1ff767bcfa0af18ffdf886134eb00fa75108eafdef971bd5ab925b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000017", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [395, 494, 444, 575]\n- T1: [646, 557, 651, 575]\n- T2: [517, 459, 537, 474]\n\nImage 2 (GROUND) labeled targets:\n- R: [210, 206, 374, 310]\n- T1: [862, 241, 884, 306]\n- T2: [520, 280, 554, 296]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d0c7a1e600af97e4214a01d83668ccb9038b75752431c9384f9b153978205fed.jpg", "reference_inputs/images/6c5bcc96e9ec88143a2af460a2c36c62a89566b5824bb951b130bc5915679c3b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000018", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [753, 274, 786, 287]\n- T1: [748, 23, 778, 51]\n- T2: [822, 434, 857, 452]\n\nImage 2 (GROUND) labeled targets:\n- R: [557, 232, 586, 318]\n- T1: [58, 216, 98, 316]\n- T2: [800, 238, 822, 310]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ab2188839eca588e14ececd553c05ab309b3db806f153f7d78edfdf9250a16d7.jpg", "reference_inputs/images/4ea6834c85dd9da4d0e393b1fbd27477d525a255f2c04252747fb6893e6cca2b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000019", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [410, 64, 463, 177]\n- T1: [318, 360, 376, 473]\n- T2: [782, 302, 810, 315]\n\nImage 2 (GROUND) labeled targets:\n- R: [467, 275, 491, 293]\n- T1: [390, 280, 431, 297]\n- T2: [655, 268, 662, 292]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b355ce4a747caad0252734b0d1a30745dc9d99b13180da815406c6aa9dbfb60e.jpg", "reference_inputs/images/64aea7386a93b8d411c31c3c476f955f4bb9d442dea86752dfdbfd3a90eec182.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000020", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [54, 482, 110, 502]\n- T1: [452, 221, 537, 430]\n- T2: [151, 180, 193, 198]\n\nImage 2 (GROUND) labeled targets:\n- R: [111, 257, 125, 298]\n- T1: [475, 234, 526, 299]\n- T2: [310, 265, 317, 293]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/de07fc500f433a6eb9993bf376f2d200ba9871b557f96ad81813d6e0f4570870.jpg", "reference_inputs/images/c75e576ecbd789e91ef2be95caf7d32345bf6782da6c9167586dc89ad790e35c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000021", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [170, 463, 176, 479]\n- T1: [89, 470, 94, 487]\n- T2: [52, 508, 118, 529]\n\nImage 2 (GROUND) labeled targets:\n- R: [742, 280, 762, 304]\n- T1: [686, 279, 691, 301]\n- T2: [596, 287, 632, 303]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7585f41fe0b84cff321aa7f677c0cca23c67433ce4c183eb9baafe39b873f4d3.jpg", "reference_inputs/images/b829ddf4c587e79a67ffe16fe25873fe19448bfe353db00f3602571b6511c0a3.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000022", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [328, 180, 351, 194]\n- T1: [162, 121, 205, 139]\n- T2: [10, 205, 58, 222]\n\nImage 2 (GROUND) labeled targets:\n- R: [1, 216, 46, 350]\n- T1: [330, 237, 350, 312]\n- T2: [294, 252, 312, 305]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0270d8b538279fc1cc2d42738a8e13de18ac9681608ba8b8b64e1d8f6abef7bd.jpg", "reference_inputs/images/3e7836c5acb2dcdcaa8ae2df4de0058e570a9f85222862e2d16221447f8c8257.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000024", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000024", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [721, 282, 746, 296]\n- T1: [752, 311, 777, 331]\n- T2: [781, 179, 811, 197]\n\nImage 2 (GROUND) labeled targets:\n- R: [440, 210, 473, 322]\n- T1: [521, 225, 549, 318]\n- T2: [275, 224, 306, 313]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1bd832572b4b448ad338c0f4e1a79903adb20e89675c701d2d4d8178f28548e0.jpg", "reference_inputs/images/16b98ca976b6de186d99c34f240152f162895e25eef46ec2f005aaf3034c6c1a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000025", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [163, 362, 200, 376]\n- T1: [122, 262, 162, 276]\n- T2: [158, 162, 191, 182]\n\nImage 2 (GROUND) labeled targets:\n- R: [426, 246, 445, 307]\n- T1: [513, 250, 526, 303]\n- T2: [620, 251, 631, 300]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2cc8a54174351d0059fbddcddde72408fcaefe0c3a706b17b11c8bb9852967bf.jpg", "reference_inputs/images/53f6413a5aece953d3d3f19040ae4fcc8e6ef179c92e51b1341438a643b7e920.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000027", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [567, 284, 652, 500]\n- T1: [139, 361, 174, 382]\n- T2: [402, 31, 453, 146]\n\nImage 2 (GROUND) labeled targets:\n- R: [573, 162, 749, 317]\n- T1: [2, 238, 29, 303]\n- T2: [424, 271, 466, 301]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0fc31c1a19b1bb2cdd002913a35bc79ef3f253358915b124d93a788326e3387b.jpg", "reference_inputs/images/e6f52202e903210d25243e3735afe40fa075f6ed21c211af9dc0feaa58a7cb34.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000028", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [523, 362, 534, 379]\n- T1: [490, 400, 499, 410]\n- T2: [485, 349, 494, 363]\n\nImage 2 (GROUND) labeled targets:\n- R: [536, 250, 579, 294]\n- T1: [334, 265, 435, 320]\n- T2: [447, 264, 474, 290]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c39ffc71392900cb722b2de9e5fc7e0bfa3f44928702031c7009afbfa7b4d8e9.jpg", "reference_inputs/images/1f4e5bfc1dc07dd2fc6ab971db43666d4bee8f303a3194ecb836bf23599c8aa9.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000029", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [414, 325, 460, 426]\n- T1: [106, 422, 143, 453]\n- T2: [822, 434, 857, 452]\n\nImage 2 (GROUND) labeled targets:\n- R: [434, 275, 472, 300]\n- T1: [178, 259, 190, 297]\n- T2: [779, 257, 791, 296]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ab2188839eca588e14ececd553c05ab309b3db806f153f7d78edfdf9250a16d7.jpg", "reference_inputs/images/72cd0a8adb1640bf26b9f9f7b03503233379e3d16e3754adda473802969f4c0f.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000030", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [550, 470, 556, 488]\n- T1: [712, 524, 773, 549]\n- T2: [385, 505, 432, 525]\n\nImage 2 (GROUND) labeled targets:\n- R: [82, 271, 88, 290]\n- T1: [216, 280, 233, 290]\n- T2: [130, 279, 162, 295]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7585f41fe0b84cff321aa7f677c0cca23c67433ce4c183eb9baafe39b873f4d3.jpg", "reference_inputs/images/6b5ca664a0bec76fcb4e83634ea93c39a1cd48cb0e2e19cac5000ed55fa74d96.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000031", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000031", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [570, 0, 612, 83]\n- T1: [146, 413, 180, 436]\n- T2: [42, 446, 91, 467]\n\nImage 2 (GROUND) labeled targets:\n- R: [542, 275, 570, 296]\n- T1: [166, 256, 172, 299]\n- T2: [36, 250, 58, 298]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/80411caf2c694c923da91acf9ce65757d6d0b3cc4e9f32a494dc213f2e0988f9.jpg", "reference_inputs/images/9ee52d4b590978577fe3a21b750b8093f46ed3b4efce315bf8098e1feb0e6325.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000032", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [320, 178, 378, 291]\n- T1: [145, 450, 179, 476]\n- T2: [189, 59, 231, 81]\n\nImage 2 (GROUND) labeled targets:\n- R: [216, 259, 617, 360]\n- T1: [111, 242, 143, 310]\n- T2: [721, 237, 750, 312]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d47c774f2e316eb76158b6303c0ca7e885e89db572b755d0cb89296d52521cfc.jpg", "reference_inputs/images/65afc4ef610accaede802be45ab069bc460e54708884cbaffae2455de23764a6.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000033", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000033", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [555, 530, 590, 575]\n- T1: [438, 530, 443, 558]\n- T2: [410, 517, 417, 542]\n\nImage 2 (GROUND) labeled targets:\n- R: [163, 281, 230, 298]\n- T1: [56, 266, 65, 293]\n- T2: [38, 267, 47, 293]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cfd3ae4956d0d912e42e8b069a600698afb87d5f906b5079bc3f973feb58a839.jpg", "reference_inputs/images/8f6e1d91122c91578b18dffedaa92d09784d3e28552882ecd25afe2c34326083.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000035", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000035", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [454, 428, 470, 447]\n- T1: [549, 555, 576, 575]\n- T2: [650, 502, 659, 524]\n\nImage 2 (GROUND) labeled targets:\n- R: [443, 277, 462, 291]\n- T1: [563, 272, 606, 300]\n- T2: [690, 266, 703, 293]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0338e6668ccc2980e70ebd4a9e32d2deccdc84389a69907b91810ac1184821af.jpg", "reference_inputs/images/34d027dbab11397f568f524e93a1358bb23b7502da274b0c04a1dba07548b837.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000036", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000036", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [462, 410, 477, 418]\n- T1: [534, 425, 549, 440]\n- T2: [530, 402, 543, 411]\n\nImage 2 (GROUND) labeled targets:\n- R: [274, 275, 367, 307]\n- T1: [686, 202, 1000, 350]\n- T2: [577, 277, 637, 302]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1c3d4109a38cb72dd2d349b03eb89d2366a1cfb46285ae129959aae657de9447.jpg", "reference_inputs/images/e8f86a144ff07fab1b6ed4595caeffbf95d4782f342c62d61c804499bb0065bf.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000037", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000037", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [534, 490, 555, 514]\n- T1: [635, 562, 641, 575]\n- T2: [650, 519, 655, 542]\n\nImage 2 (GROUND) labeled targets:\n- R: [740, 276, 763, 295]\n- T1: [891, 258, 902, 297]\n- T2: [914, 262, 924, 294]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6d76248e7087fb5fcf6decf03cdcb4e685fdbb05d11f13983b35973a1a64e154.jpg", "reference_inputs/images/74f386330c7d9af7076f742a7d66db3df65f2fcc5027d9e6f0c2bb2d6c20ab41.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000038", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000038", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [505, 392, 521, 416]\n- T1: [476, 457, 494, 482]\n- T2: [510, 448, 525, 465]\n\nImage 2 (GROUND) labeled targets:\n- R: [494, 239, 539, 298]\n- T1: [0, 171, 350, 426]\n- T2: [497, 247, 615, 334]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3e21d07491e09cd36408a4cbcdacf1e77d7462660023e5288805dc80014a0207.jpg", "reference_inputs/images/423edee79517840c130357a8a1a08ab8664be9cec297a266fc992caea0152d92.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000039", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000039", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [82, 211, 126, 226]\n- T1: [106, 262, 148, 276]\n- T2: [90, 46, 130, 75]\n\nImage 2 (GROUND) labeled targets:\n- R: [354, 247, 372, 307]\n- T1: [280, 249, 298, 309]\n- T2: [560, 248, 573, 306]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8a36ca2e5fe4225b02d5b48a0cbef01de524cc951818afcfdff8047020c582b5.jpg", "reference_inputs/images/91613324c1c5a055df08a78e89bc88cb93a4e51fd5b21ca12316859c761a71ab.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000040", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000040", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [686, 549, 691, 574]\n- T1: [434, 472, 457, 501]\n- T2: [441, 515, 464, 546]\n\nImage 2 (GROUND) labeled targets:\n- R: [906, 243, 924, 301]\n- T1: [388, 269, 422, 289]\n- T2: [368, 274, 422, 303]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/369d8c5a94fad57873a397e17382502d73f237070752a22f9e7d59767de6f1e1.jpg", "reference_inputs/images/52169b7a0b342f2d63699bcc716256cbbe071a240289254b98442a1483faace4.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000041", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000041", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [500, 406, 551, 518]\n- T1: [52, 151, 97, 170]\n- T2: [122, 155, 167, 171]\n\nImage 2 (GROUND) labeled targets:\n- R: [495, 247, 573, 315]\n- T1: [166, 258, 178, 290]\n- T2: [213, 262, 224, 296]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/686afe55c2fa06299ea2faa81211ce9eaa6045b06db4b14b5dbd4117d79dc52f.jpg", "reference_inputs/images/176edff7c377ecfb0a85bbb5fe6f909debb69e63dd867a4d9757b6b09591aec1.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000042", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000042", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [82, 211, 126, 226]\n- T1: [90, 46, 130, 75]\n- T2: [106, 262, 148, 276]\n\nImage 2 (GROUND) labeled targets:\n- R: [182, 248, 204, 308]\n- T1: [389, 249, 402, 307]\n- T2: [102, 250, 122, 310]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8a36ca2e5fe4225b02d5b48a0cbef01de524cc951818afcfdff8047020c582b5.jpg", "reference_inputs/images/1dd8da1c6e720f3f329f418fddf85c33c24dc0d7eddb073383e93040bd79026b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000043", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000043", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [466, 391, 490, 399]\n- T1: [674, 554, 683, 575]\n- T2: [406, 557, 439, 575]\n\nImage 2 (GROUND) labeled targets:\n- R: [462, 283, 482, 289]\n- T1: [729, 259, 740, 295]\n- T2: [364, 278, 417, 299]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/54a45b98db6cf9dcae184341b11c3ada9efb63b5c8c0222a21ea269ead7746f1.jpg", "reference_inputs/images/389a42f43ad667172d6855c4182d3f7f74d9b73cea54b3b04b85012cce008507.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000045", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000045", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 4 shared labeled targets across the two images, including 3 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2, T3.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [512, 449, 529, 472]\n- T1: [422, 522, 429, 546]\n- T2: [554, 425, 567, 436]\n- T3: [501, 501, 522, 530]\n\nImage 2 (GROUND) labeled targets:\n- R: [522, 275, 534, 293]\n- T1: [389, 264, 397, 294]\n- T2: [562, 281, 578, 291]\n- T3: [497, 274, 526, 297]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T3 -> T1 -> T2\nC: unable to determine\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/29a6541a1c03ccfac6d073b89565f4d67321d2e2acdf67f5fe0a474b8c507135.jpg", "reference_inputs/images/314d01bbfb4d254b346e7f6bc4fb62da813832c57ce8b979e13a5d0aae67e2bd.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000046", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000046", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [494, 431, 518, 469]\n- T1: [326, 512, 335, 535]\n- T2: [345, 467, 356, 483]\n\nImage 2 (GROUND) labeled targets:\n- R: [488, 254, 522, 294]\n- T1: [216, 258, 228, 295]\n- T2: [274, 269, 288, 292]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ee076bd854493700b9fab044dd65fda55b8ff5f246f0fb232ded9339d9b3fc6c.jpg", "reference_inputs/images/15efbf7d748243da6419b1350bea59113e625f891f627c70f979a4ee745a68a8.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000047", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000047", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [599, 75, 647, 178]\n- T1: [95, 216, 134, 230]\n- T2: [500, 470, 551, 575]\n\nImage 2 (GROUND) labeled targets:\n- R: [561, 278, 590, 296]\n- T1: [257, 266, 266, 294]\n- T2: [499, 258, 554, 306]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/73b89e2237847f5cca0725df1591c03e7b3ef51668e61483992f6a44fab9033a.jpg", "reference_inputs/images/267140f4919911df6db0b7cf8ed7c04c6d81dac27a9d963632b49ae48d93239f.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000048", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000048", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [554, 425, 567, 436]\n- T1: [422, 522, 429, 546]\n- T2: [587, 524, 619, 553]\n\nImage 2 (GROUND) labeled targets:\n- R: [408, 284, 421, 293]\n- T1: [152, 270, 158, 292]\n- T2: [327, 282, 368, 297]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/29a6541a1c03ccfac6d073b89565f4d67321d2e2acdf67f5fe0a474b8c507135.jpg", "reference_inputs/images/006e4b97e40fb9fb2f043974ca480bf6e4fbbcfac6f82f579c3ab98b45c6c1f4.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000049", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000049", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [147, 526, 180, 559]\n- T1: [818, 76, 858, 97]\n- T2: [823, 518, 856, 541]\n\nImage 2 (GROUND) labeled targets:\n- R: [100, 250, 118, 302]\n- T1: [694, 266, 700, 292]\n- T2: [865, 248, 881, 299]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b6e69bb6b20e30c6251d467415a026ad170366486c78c77a0625195e38cab5b5.jpg", "reference_inputs/images/214450fcb9ec8f693510b00eb2718a24d336b00136b94d6785fc63ff6db98864.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000050", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000050", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [873, 396, 918, 413]\n- T1: [761, 55, 789, 82]\n- T2: [874, 125, 920, 142]\n\nImage 2 (GROUND) labeled targets:\n- R: [906, 247, 934, 308]\n- T1: [446, 220, 478, 316]\n- T2: [556, 245, 572, 306]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/94a42c416d039b03a48924175f0fe6b51d54b3a84c1a26fee68d14316a486a57.jpg", "reference_inputs/images/2506433c82e3e48ac5f1db08b9f7e3f477f766a7bfcdea59cc70da55f5e29026.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000051", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000051", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [775, 52, 808, 77]\n- T1: [363, 417, 412, 519]\n- T2: [315, 198, 374, 307]\n\nImage 2 (GROUND) labeled targets:\n- R: [709, 260, 718, 295]\n- T1: [217, 264, 356, 322]\n- T2: [326, 262, 382, 305]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/19b97407fa1dc9b3bb12ab0247fd4fa6b09c797c0452b5a47ff0ca4a0408ea8c.jpg", "reference_inputs/images/e071a3d5ffba4f7bb6addc5eae0d5ad70da90edbd8c3be2635e149f0f653c9c0.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000053", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000053", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [301, 518, 310, 540]\n- T1: [310, 528, 317, 553]\n- T2: [350, 500, 360, 521]\n\nImage 2 (GROUND) labeled targets:\n- R: [304, 254, 322, 306]\n- T1: [243, 251, 259, 308]\n- T2: [380, 242, 402, 310]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f6f61ba102d1b190255546c7c2a3a4ec5ac6658a821aaae3580ee11e9c1267e5.jpg", "reference_inputs/images/a71e45c547e3afc57b1c6f30cf4d9fc977114d99cb3e9ae0d66203b12df4ad69.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_20m/000054", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000054", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [766, 0, 795, 29]\n- T1: [398, 87, 454, 201]\n- T2: [812, 295, 846, 313]\n\nImage 2 (GROUND) labeled targets:\n- R: [650, 267, 657, 292]\n- T1: [439, 279, 477, 297]\n- T2: [746, 262, 757, 296]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9b07624335b4fa633b68a1113aa88dc561c5ac1dcdfee58fdfa373ca97d99437.jpg", "reference_inputs/images/67d3c4dfaf5334baa46fdb18857f233075e29e32dcbad3903bb91dcb5950e373.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000001", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [418, 385, 456, 458]\n- T1: [290, 564, 302, 575]\n- T2: [270, 370, 293, 379]\n\nImage 2 (GROUND) labeled targets:\n- R: [0, 245, 498, 404]\n- T1: [48, 244, 74, 284]\n- T2: [525, 240, 545, 310]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/edbffda0f2b67c463fd64cae867842b22130d1b4b709eeaf66f07cd968060492.jpg", "reference_inputs/images/f387777464b99b654288c0f5691bec7740a4e22e19afc495f2252450b68ef7d1.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000003", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [422, 152, 459, 227]\n- T1: [244, 169, 261, 182]\n- T2: [676, 166, 688, 178]\n\nImage 2 (GROUND) labeled targets:\n- R: [494, 282, 529, 299]\n- T1: [345, 270, 352, 295]\n- T2: [718, 271, 725, 297]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/65ab20dcbba598480d730324f3c9ceffe765f5fccfc0bf381aeb02abec648f0f.jpg", "reference_inputs/images/9c6b3f29f4d9176113b4baac3a280ea580c2ac94190eae813932e337951513d2.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000004", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [274, 541, 294, 558]\n- T1: [137, 432, 161, 446]\n- T2: [237, 436, 258, 448]\n\nImage 2 (GROUND) labeled targets:\n- R: [198, 246, 235, 311]\n- T1: [474, 256, 494, 302]\n- T2: [454, 250, 477, 306]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/beb4a4da294a37a291b7007f52317ea5ec3c945d61ec032b2127c546ed28efa1.jpg", "reference_inputs/images/17f870c3e80b4ea658ab9fad190bfd9efb07496fc221e0102bd1988432097401.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000006", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [446, 367, 479, 440]\n- T1: [281, 205, 296, 215]\n- T2: [697, 205, 716, 215]\n\nImage 2 (GROUND) labeled targets:\n- R: [254, 235, 423, 333]\n- T1: [181, 255, 198, 298]\n- T2: [782, 255, 797, 298]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c236d9347f55c0c3ec5ba178d174cdfed7be3be0fb69abec27fac1b60b6cc6cf.jpg", "reference_inputs/images/64ecb14e96931fa12d63d7d52e97e6551d9ab8794bec873526ad42f1cf64fd39.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000007", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [475, 478, 490, 493]\n- T1: [499, 450, 514, 476]\n- T2: [470, 447, 481, 462]\n\nImage 2 (GROUND) labeled targets:\n- R: [37, 261, 332, 344]\n- T1: [429, 170, 538, 314]\n- T2: [314, 262, 386, 304]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6dbfd15899345da323d2701eef481ffd7f42d3c6927e0fa020b0dae7b21106da.jpg", "reference_inputs/images/dd84ec2bafb16bb213460e7a124814e17d0f7d9c46fed3791c34b8ea12eba9b2.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000008", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [473, 437, 484, 446]\n- T1: [448, 508, 473, 562]\n- T2: [510, 511, 523, 528]\n\nImage 2 (GROUND) labeled targets:\n- R: [425, 281, 442, 293]\n- T1: [0, 18, 257, 336]\n- T2: [498, 264, 580, 320]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3b817beb70079de70a053c49a0c92810b0f54beca2a9e32a535454ce3b66b03f.jpg", "reference_inputs/images/34591fc1fa22610d2519ee5814ed778a1c518ffb2f2dfb3ab15e3acf8d937d50.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000010", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [504, 43, 512, 60]\n- T1: [0, 250, 125, 314]\n- T2: [558, 474, 566, 498]\n\nImage 2 (GROUND) labeled targets:\n- R: [916, 218, 926, 246]\n- T1: [646, 227, 674, 255]\n- T2: [510, 238, 519, 265]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9387fb07c1ea73e8176f603aa4bc4630b33ca3cd16a48c5c1865d623fa341a4d.jpg", "reference_inputs/images/a0ccf0f5e696281230600960e2603ce51a994ea07386f34ba1410ded71be1b4c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000011", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [694, 231, 707, 245]\n- T1: [286, 77, 307, 90]\n- T2: [214, 203, 242, 212]\n\nImage 2 (GROUND) labeled targets:\n- R: [753, 258, 765, 298]\n- T1: [304, 267, 320, 294]\n- T2: [152, 261, 163, 296]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5b432ec76c85bb5a57956a33adb0353060181098b08c679dfb6e3d264a7c7ff0.jpg", "reference_inputs/images/9a6e5f9f30ee237528c1c8bea12e3e2df67813b2f7eaf7df22f916e035be8caa.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000012", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 4 shared labeled targets across the two images, including 3 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2, T3.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [234, 246, 254, 254]\n- T1: [239, 139, 257, 155]\n- T2: [250, 279, 270, 287]\n- T3: [231, 238, 252, 246]\n\nImage 2 (GROUND) labeled targets:\n- R: [516, 246, 534, 306]\n- T1: [719, 247, 740, 306]\n- T2: [450, 248, 467, 308]\n- T3: [531, 250, 546, 306]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1 -> T3\nC: T1 -> T2 -> T3\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3c008430886d13cc06b3af1ca92b44323c9fdc224aa91af5bf1a822dea3e92b5.jpg", "reference_inputs/images/9bd1be9135b490482662595471ecb13389cc21f0185e9afd81e9b1cf2152433c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000013", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [548, 288, 602, 425]\n- T1: [270, 340, 286, 354]\n- T2: [440, 122, 473, 196]\n\nImage 2 (GROUND) labeled targets:\n- R: [544, 235, 607, 298]\n- T1: [269, 263, 278, 294]\n- T2: [458, 277, 482, 294]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7178e1b2d214b3c18b7b708a46724106443903040c88c0642dfca51bb5d6326b.jpg", "reference_inputs/images/aa247e83cbfa6a55e87f479c5d2e650e3f2266f2b3bb012623eb5ab140aa8602.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000014", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [668, 494, 694, 509]\n- T1: [447, 526, 478, 575]\n- T2: [664, 341, 682, 350]\n\nImage 2 (GROUND) labeled targets:\n- R: [706, 262, 738, 295]\n- T1: [414, 273, 464, 304]\n- T2: [654, 266, 662, 293]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a1d1537e3be8850fe6f3702a692d2159724b057ca4725db2e331af7a19f6bdbf.jpg", "reference_inputs/images/fc2f455a3a134b67bdb86f8967c4e888682308e77817c4a1c0ede8e7b7ffb7cb.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000015", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [919, 384, 952, 394]\n- T1: [290, 117, 302, 126]\n- T2: [668, 450, 690, 462]\n\nImage 2 (GROUND) labeled targets:\n- R: [534, 260, 545, 299]\n- T1: [19, 232, 66, 310]\n- T2: [766, 230, 810, 318]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ea0b07874c421a30ce3bde9476a6b09e6eeb1f527b7870a89b564f5017a1476a.jpg", "reference_inputs/images/d396ca39d542235fbcc69b5858e5f5a773de1152513842df5f369577757d93ca.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000016", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [714, 558, 718, 575]\n- T1: [886, 537, 915, 575]\n- T2: [660, 470, 674, 484]\n\nImage 2 (GROUND) labeled targets:\n- R: [418, 270, 436, 292]\n- T1: [478, 184, 574, 311]\n- T2: [730, 281, 794, 294]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f6477f2676e348ed6ac8c5e36ad5d03777705ee4cd07bf1376693b6eee8e1d2d.jpg", "reference_inputs/images/31c6abcf7eb4aa46f118f145da3d0ff14ff0c62da2378150b791ad7d61f0b10b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000017", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 4 shared labeled targets across the two images, including 3 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2, T3.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [58, 530, 86, 546]\n- T1: [302, 360, 324, 367]\n- T2: [287, 222, 302, 236]\n- T3: [330, 448, 343, 463]\n\nImage 2 (GROUND) labeled targets:\n- R: [200, 266, 210, 301]\n- T1: [275, 238, 290, 314]\n- T2: [607, 242, 642, 310]\n- T3: [0, 240, 10, 318]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2 -> T3\nC: T2 -> T3 -> T1\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5503db5f5ae4225d2c6f3add36d2f86fa8ee0bd3d0e0cce90ad74ba26b5a2990.jpg", "reference_inputs/images/ee43fb8141aa4576e3ebdf22ef02f663d6e11cac7f7e49ef76189736638c9d2a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000018", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [626, 26, 636, 47]\n- T1: [504, 43, 512, 60]\n- T2: [545, 174, 622, 215]\n\nImage 2 (GROUND) labeled targets:\n- R: [350, 264, 370, 298]\n- T1: [248, 263, 258, 297]\n- T2: [67, 276, 162, 309]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9387fb07c1ea73e8176f603aa4bc4630b33ca3cd16a48c5c1865d623fa341a4d.jpg", "reference_inputs/images/16652bd8a0749e2b8903fd22753cff94fa7d4fe455aeb70555cc2ffe81a55955.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000020", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [286, 197, 305, 205]\n- T1: [188, 251, 210, 262]\n- T2: [386, 266, 418, 333]\n\nImage 2 (GROUND) labeled targets:\n- R: [304, 264, 312, 294]\n- T1: [178, 262, 188, 294]\n- T2: [359, 275, 406, 300]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/01bec1e6063f2abdaa5a2a82f0cc79873e8527ba96cf4d200ccc3916f0c1bc18.jpg", "reference_inputs/images/502fe2b7bf5d1fe5c44e96178c57cdc8f6cda7fad2cf013c6853ede35a442838.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000021", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [373, 533, 410, 575]\n- T1: [665, 461, 678, 476]\n- T2: [646, 559, 664, 575]\n\nImage 2 (GROUND) labeled targets:\n- R: [358, 278, 408, 298]\n- T1: [646, 266, 653, 293]\n- T2: [654, 266, 667, 294]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8f8c0dec3dcd919ee36e796fe4394daf6f9e24b899b50a0dcda096047b7b9f4d.jpg", "reference_inputs/images/0a637c542cd55efcd7db57209cb296cd5b3b8122a25a1081e21a75d2467f1eec.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000022", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [242, 57, 258, 69]\n- T1: [212, 46, 237, 62]\n- T2: [116, 77, 143, 91]\n\nImage 2 (GROUND) labeled targets:\n- R: [618, 262, 629, 306]\n- T1: [622, 250, 635, 292]\n- T2: [553, 258, 564, 300]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f21e2f3ad9c76b4b7cd93f585cefa4c1a8e7d0563290f1be9515c9b7766eb3ff.jpg", "reference_inputs/images/d1a2e0d8a7b1234553fb0f2b6b49098b855cc90c859e55d2b0ab0d19a2abd575.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000023", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [681, 110, 695, 122]\n- T1: [698, 398, 711, 409]\n- T2: [712, 301, 729, 312]\n\nImage 2 (GROUND) labeled targets:\n- R: [166, 222, 208, 314]\n- T1: [962, 238, 1008, 318]\n- T2: [695, 240, 726, 313]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f64871178c49e7c522c6b88ab61796420ba48e16d787c5ab3deaac380c90cc48.jpg", "reference_inputs/images/a4819079d3821c574871437de2ea57a167191c3f99cd65d3aac541cb6ba59da1.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000025", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [283, 398, 299, 411]\n- T1: [280, 187, 295, 198]\n- T2: [417, 356, 455, 430]\n\nImage 2 (GROUND) labeled targets:\n- R: [454, 268, 461, 295]\n- T1: [468, 274, 473, 294]\n- T2: [569, 283, 606, 300]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b4d0eda699256237ed11a4a95c83ee5937f0838e00ca027a698437dc64ed5d94.jpg", "reference_inputs/images/47f9e862659644a60fa5865a905e3616ae64a6438419abb7aae6d257478f5f7c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000028", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [850, 178, 874, 194]\n- T1: [880, 280, 954, 311]\n- T2: [670, 88, 682, 101]\n\nImage 2 (GROUND) labeled targets:\n- R: [546, 256, 570, 302]\n- T1: [663, 278, 722, 306]\n- T2: [322, 225, 350, 318]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0fef7bf43103cdecd4e2c7ba577149ffb74ee952c482f428f478d115ae197001.jpg", "reference_inputs/images/9dabade18b21c6aa67c58127d7a894f94f432a931037529764f20ab13e6c9cdc.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000030", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [548, 499, 581, 575]\n- T1: [383, 94, 419, 168]\n- T2: [235, 70, 250, 99]\n\nImage 2 (GROUND) labeled targets:\n- R: [564, 268, 621, 305]\n- T1: [423, 274, 450, 294]\n- T2: [333, 270, 350, 292]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2942222141645b774a4b3bd6e5b7422af344083fdde6423f2b414362c1d9f968.jpg", "reference_inputs/images/75bb9f3ac1921d7fed1cce6bc9bbfa9956c3988c4fcf366d0c757f6c48b81399.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000031", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000031", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [334, 243, 396, 306]\n- T1: [434, 22, 445, 40]\n- T2: [338, 70, 354, 89]\n\nImage 2 (GROUND) labeled targets:\n- R: [178, 271, 310, 307]\n- T1: [585, 258, 598, 296]\n- T2: [470, 263, 482, 294]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/17e8acdee8139296e74054ad46f0b8e4ac3f31e20e58c6ad512a0f49c246c007.jpg", "reference_inputs/images/29fc872b749e944ac734776f6892fe8e4600cf609cbd0a71f1ff2dda5f3dc4e4.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000032", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [474, 458, 522, 575]\n- T1: [654, 279, 666, 288]\n- T2: [280, 434, 300, 449]\n\nImage 2 (GROUND) labeled targets:\n- R: [449, 198, 534, 308]\n- T1: [647, 265, 654, 294]\n- T2: [220, 256, 226, 297]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b54de1266305b9b20f5b628edb6cc3be8bf303f5eaaa4ba4235d674b186ebab7.jpg", "reference_inputs/images/4183a0b6de334d0e851eef8ba64a5005f51b443781882f4bdd69e0315019eef6.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000033", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000033", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [530, 466, 543, 477]\n- T1: [534, 503, 549, 521]\n- T2: [462, 476, 477, 488]\n\nImage 2 (GROUND) labeled targets:\n- R: [577, 277, 637, 302]\n- T1: [686, 202, 1000, 350]\n- T2: [274, 275, 367, 307]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4e33cdd8ea49b40c1dc79aa8a47c3247f0846a5bfc6bf23c11356d4b495d047c.jpg", "reference_inputs/images/e8f86a144ff07fab1b6ed4595caeffbf95d4782f342c62d61c804499bb0065bf.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000034", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000034", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [215, 425, 242, 436]\n- T1: [279, 226, 299, 237]\n- T2: [474, 253, 528, 387]\n\nImage 2 (GROUND) labeled targets:\n- R: [238, 266, 251, 294]\n- T1: [358, 270, 362, 291]\n- T2: [486, 249, 519, 295]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9800dd1985605e5316d25a285f75017cbfaced8522231d9c47b6dd8504e11fe3.jpg", "reference_inputs/images/955cbee518632600ad55301eeb8bda54362a5a47bf884e33762d6db37d7a90c0.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000035", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000035", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [448, 466, 478, 534]\n- T1: [1, 180, 33, 191]\n- T2: [554, 0, 600, 98]\n\nImage 2 (GROUND) labeled targets:\n- R: [334, 265, 435, 320]\n- T1: [26, 263, 38, 294]\n- T2: [536, 250, 579, 294]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6e4af1e6834a23f481da3fb5481f4f060494ac2b58b1035dff9f3ad5ad3f2661.jpg", "reference_inputs/images/1f4e5bfc1dc07dd2fc6ab971db43666d4bee8f303a3194ecb836bf23599c8aa9.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000036", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000036", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [750, 0, 766, 15]\n- T1: [822, 108, 848, 122]\n- T2: [673, 50, 691, 64]\n\nImage 2 (GROUND) labeled targets:\n- R: [70, 238, 99, 305]\n- T1: [354, 251, 371, 302]\n- T2: [27, 225, 74, 314]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d23def54bc38c6919f30f1cf05d51e32af4daf8d2413d5881837e1e5da81ad0b.jpg", "reference_inputs/images/eb97a3246d862464d2b51b720f73d209b8d3a46bb24e54d9e5a2b762b5dfa240.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000037", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000037", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [506, 159, 538, 227]\n- T1: [554, 498, 608, 575]\n- T2: [222, 306, 239, 316]\n\nImage 2 (GROUND) labeled targets:\n- R: [506, 276, 540, 299]\n- T1: [299, 222, 414, 305]\n- T2: [98, 253, 125, 298]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fa77697dfc4085fced112433ab375a16fb84c526f445ccec94d4d1edbce39c59.jpg", "reference_inputs/images/5a50298408b5a44380b325f06a63c9e6f92a85c81613f17465ad2b8e5b4836a3.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000038", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000038", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 4 shared labeled targets across the two images, including 3 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2, T3.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [214, 92, 242, 104]\n- T1: [234, 31, 254, 50]\n- T2: [249, 66, 269, 80]\n- T3: [273, 251, 289, 266]\n\nImage 2 (GROUND) labeled targets:\n- R: [572, 252, 589, 304]\n- T1: [690, 245, 706, 306]\n- T2: [630, 247, 649, 306]\n- T3: [232, 240, 250, 309]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: unable to determine\nC: T3 -> T1 -> T2\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e4457e1a53e081868651ab4f5bdb7954e2e6197ba1a7b0805272ee9e12908fcd.jpg", "reference_inputs/images/fb65f7995f334114946a52d3172118daba99133ed34921956d2365e3e709e2a6.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000039", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000039", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [687, 220, 700, 229]\n- T1: [754, 401, 774, 411]\n- T2: [677, 359, 690, 370]\n\nImage 2 (GROUND) labeled targets:\n- R: [52, 221, 97, 313]\n- T1: [556, 245, 572, 306]\n- T2: [446, 220, 478, 316]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7312f49d0b403f90f880885c10c3f564ca6b1d2fa5ce2b847dca574c103831d4.jpg", "reference_inputs/images/2506433c82e3e48ac5f1db08b9f7e3f477f766a7bfcdea59cc70da55f5e29026.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000040", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000040", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [298, 210, 313, 220]\n- T1: [285, 63, 300, 77]\n- T2: [244, 234, 261, 241]\n\nImage 2 (GROUND) labeled targets:\n- R: [218, 233, 246, 314]\n- T1: [574, 241, 594, 268]\n- T2: [239, 248, 254, 308]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/223cbee8041d6f1a55407a61b86a321c281765434511cd82b74d02a28bad5c15.jpg", "reference_inputs/images/5c64e995683b1de21c283c671349dac82351d1b3a64375df3ccdb6bffaedb802.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000041", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000041", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [347, 224, 476, 344]\n- T1: [994, 395, 1023, 413]\n- T2: [849, 467, 922, 536]\n\nImage 2 (GROUND) labeled targets:\n- R: [34, 245, 103, 297]\n- T1: [657, 253, 678, 302]\n- T2: [349, 272, 470, 317]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9a637268ccc66c4e35c4ed2e56daf521b49a57fefec9c91252cab14f54e236a3.jpg", "reference_inputs/images/6f0f1e07119f53abe8f08c28c43962d83b3debe85de2c64a6e89e21ebb95a02c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000042", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000042", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [274, 425, 293, 435]\n- T1: [380, 0, 435, 128]\n- T2: [548, 390, 602, 535]\n\nImage 2 (GROUND) labeled targets:\n- R: [360, 272, 366, 290]\n- T1: [453, 266, 478, 290]\n- T2: [534, 254, 570, 294]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b5beace03cbf4076653701c80e5956fa795b10d04db5e3655a7d1fff5e4a8dfd.jpg", "reference_inputs/images/6e73dd999c827994b8f9751f6509c2a233f365e7d66ae156e647207f6714bc11.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000043", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000043", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [382, 265, 421, 338]\n- T1: [289, 522, 307, 538]\n- T2: [257, 397, 274, 407]\n\nImage 2 (GROUND) labeled targets:\n- R: [544, 258, 945, 358]\n- T1: [58, 245, 100, 314]\n- T2: [390, 247, 408, 309]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e5b411a68a7ef2e6b0baef87b94099e01703403ccf54d36ff3073d57e76ece61.jpg", "reference_inputs/images/6bde287cf09eccf13f1a7282d5f5bbecd74d66d92d3f8b16409fa292a063c87d.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000044", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000044", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [254, 274, 271, 284]\n- T1: [302, 148, 318, 158]\n- T2: [329, 242, 342, 252]\n\nImage 2 (GROUND) labeled targets:\n- R: [425, 242, 454, 308]\n- T1: [730, 234, 758, 311]\n- T2: [489, 222, 520, 316]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ef9daea12006a2e51e5c982a5b280c67b50689e9b2b397d8028d0f4166e33386.jpg", "reference_inputs/images/82e49c092a52ba14ad4d6e963fee05e3b260013b8eddacfdcbdcbc606ff915ef.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000045", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000045", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [42, 540, 72, 564]\n- T1: [38, 148, 69, 166]\n- T2: [239, 158, 262, 170]\n\nImage 2 (GROUND) labeled targets:\n- R: [146, 266, 166, 301]\n- T1: [598, 263, 615, 298]\n- T2: [646, 242, 668, 306]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0fef7bf43103cdecd4e2c7ba577149ffb74ee952c482f428f478d115ae197001.jpg", "reference_inputs/images/a6bbb87b287f5401a21be4253d6a9ba6d1b1fb7dd5a4446eec287e86133e81c1.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000046", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000046", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 4 shared labeled targets across the two images, including 3 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2, T3.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [286, 262, 310, 271]\n- T1: [721, 167, 736, 177]\n- T2: [686, 553, 702, 574]\n- T3: [719, 456, 734, 468]\n\nImage 2 (GROUND) labeled targets:\n- R: [284, 266, 304, 294]\n- T1: [694, 266, 700, 292]\n- T2: [962, 226, 994, 310]\n- T3: [865, 248, 881, 299]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: T2 -> T3 -> T1\nC: unable to determine\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cec587f7f71d59f91723be411e06f634b2c38d94eef8cefcfe273dea4be0b7c4.jpg", "reference_inputs/images/214450fcb9ec8f693510b00eb2718a24d336b00136b94d6785fc63ff6db98864.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000047", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000047", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [490, 310, 525, 382]\n- T1: [674, 78, 686, 93]\n- T2: [275, 468, 290, 482]\n\nImage 2 (GROUND) labeled targets:\n- R: [478, 265, 529, 301]\n- T1: [635, 270, 641, 292]\n- T2: [100, 250, 118, 302]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cec587f7f71d59f91723be411e06f634b2c38d94eef8cefcfe273dea4be0b7c4.jpg", "reference_inputs/images/214450fcb9ec8f693510b00eb2718a24d336b00136b94d6785fc63ff6db98864.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000048", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000048", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [506, 456, 522, 485]\n- T1: [475, 554, 494, 575]\n- T2: [510, 537, 526, 560]\n\nImage 2 (GROUND) labeled targets:\n- R: [494, 239, 539, 298]\n- T1: [0, 171, 350, 426]\n- T2: [497, 247, 615, 334]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5f78e355e5e15fb436335e1dcade9259514f85cdeef39a550afcc8bcc9b80649.jpg", "reference_inputs/images/423edee79517840c130357a8a1a08ab8664be9cec297a266fc992caea0152d92.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000049", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000049", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [522, 426, 560, 501]\n- T1: [666, 248, 678, 255]\n- T2: [649, 350, 662, 360]\n\nImage 2 (GROUND) labeled targets:\n- R: [789, 268, 819, 281]\n- T1: [704, 254, 710, 278]\n- T2: [722, 258, 730, 276]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5dbeddd1c3c7318355a2bef03ca4c4171d13aa2262ccc2af44a86fbae0d714a3.jpg", "reference_inputs/images/db28015dd827c7a87eca54c8bed03ee4194eed92a3b8f6f40966f1f3a8f4b38c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000051", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000051", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [674, 291, 690, 299]\n- T1: [714, 397, 734, 406]\n- T2: [664, 130, 683, 142]\n\nImage 2 (GROUND) labeled targets:\n- R: [762, 256, 795, 298]\n- T1: [986, 234, 1014, 305]\n- T2: [679, 262, 688, 294]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/acba93faab2b2b94bd8f121833dcdc0686ddae44c5819cbb8c7ea4fc08b05082.jpg", "reference_inputs/images/51c4243b98d1a3e1b646c00d0570aefc32d478d1128648c52c2f84bc12b8f8ff.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000052", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000052", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [283, 215, 298, 224]\n- T1: [260, 281, 278, 290]\n- T2: [374, 61, 409, 134]\n\nImage 2 (GROUND) labeled targets:\n- R: [579, 245, 594, 304]\n- T1: [446, 245, 462, 306]\n- T2: [860, 250, 1023, 318]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e15b5ab4486d18e49579e63edec4137ea576208c7b3ed60f09684c21f8bbd4b2.jpg", "reference_inputs/images/83de54604647f8550f7370570918454c1f86c25ce7ef1fa73fb95060dbcecc4f.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000054", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000054", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [242, 57, 258, 69]\n- T1: [116, 77, 143, 91]\n- T2: [212, 46, 237, 62]\n\nImage 2 (GROUND) labeled targets:\n- R: [452, 262, 464, 307]\n- T1: [438, 259, 450, 301]\n- T2: [464, 253, 483, 305]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f21e2f3ad9c76b4b7cd93f585cefa4c1a8e7d0563290f1be9515c9b7766eb3ff.jpg", "reference_inputs/images/a60763d0d1b0ce193d9f444a7d18b463b8dd9af80e65adad392a15122348c50c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000055", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000055", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [616, 163, 654, 237]\n- T1: [390, 206, 399, 216]\n- T2: [541, 466, 579, 539]\n\nImage 2 (GROUND) labeled targets:\n- R: [538, 283, 570, 298]\n- T1: [361, 271, 367, 294]\n- T2: [451, 279, 508, 307]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/973b4ee6cb6be27d8382e93778c62382c19fd1bcba16f5d79f44a33baeb49871.jpg", "reference_inputs/images/0712923cb9ef7bc88dbc73ca5969dd9ffecc54a415e0f4d2e3a28e04b7f8a7a7.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000056", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000056", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [469, 450, 481, 460]\n- T1: [460, 490, 473, 506]\n- T2: [531, 470, 543, 484]\n\nImage 2 (GROUND) labeled targets:\n- R: [445, 282, 465, 291]\n- T1: [417, 279, 442, 294]\n- T2: [543, 275, 566, 293]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d4add35c59b208cf3073b24e073aa526911f29639774756b227a864960cb28c1.jpg", "reference_inputs/images/9c4e551e68a3a6d6aa8c608c169cf03df9762fc2789534aeba9a6a5bf69ba888.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000057", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000057", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [390, 429, 401, 443]\n- T1: [606, 175, 616, 185]\n- T2: [798, 548, 821, 566]\n\nImage 2 (GROUND) labeled targets:\n- R: [464, 275, 475, 298]\n- T1: [765, 266, 776, 302]\n- T2: [150, 243, 180, 316]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/917c46fc5eb1879e3fc37c74683cf256f05979d46d41df406d52e1334f750549.jpg", "reference_inputs/images/0aa9e9877595e7ea42312c409755117570b7d8ce02794676566ebbbedb330209.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000059", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000059", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [471, 439, 481, 450]\n- T1: [458, 493, 474, 510]\n- T2: [510, 498, 522, 514]\n\nImage 2 (GROUND) labeled targets:\n- R: [400, 275, 432, 295]\n- T1: [115, 269, 306, 323]\n- T2: [491, 260, 600, 330]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d244fd3db364087e2ec3d7b95f0825a29b0801934b47170eb9f4cba1fe921459.jpg", "reference_inputs/images/b719686ac3ce75d89888dbdd46c328c0141d8802f7ef17467032f5f9c3aea3c0.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000060", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000060", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 4 shared labeled targets across the two images, including 3 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2, T3.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [471, 442, 482, 450]\n- T1: [438, 552, 466, 575]\n- T2: [602, 552, 606, 570]\n- T3: [506, 560, 526, 575]\n\nImage 2 (GROUND) labeled targets:\n- R: [438, 282, 460, 292]\n- T1: [0, 30, 268, 333]\n- T2: [913, 232, 945, 307]\n- T3: [484, 269, 593, 323]\n\nAnswer options:\nA: T3 -> T1 -> T2\nB: T2 -> T3 -> T1\nC: T3 -> T2 -> T1\nD: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e9067dfc8620208fff7788da69543ac530ef3e75011512f0184c3a8e854a456f.jpg", "reference_inputs/images/ce6b08db9a41daef36af00206c758b526312c6cf41bef6347f4d72c299e9eb2b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000061", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000061", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [238, 493, 254, 506]\n- T1: [230, 273, 254, 283]\n- T2: [384, 506, 419, 575]\n\nImage 2 (GROUND) labeled targets:\n- R: [542, 253, 558, 294]\n- T1: [958, 242, 983, 303]\n- T2: [164, 207, 549, 358]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/043861481c93694c02127ab7d41d2e19419e9b7d4039311d46ba9a2f73117327.jpg", "reference_inputs/images/05653c6340d7df55442d03ed3d8119b93e8374d103577b227acac495e750863c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000062", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000062", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [236, 528, 256, 548]\n- T1: [550, 279, 578, 352]\n- T2: [698, 422, 711, 438]\n\nImage 2 (GROUND) labeled targets:\n- R: [388, 267, 394, 292]\n- T1: [642, 278, 664, 293]\n- T2: [773, 267, 780, 291]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4ab1ce6d791179a724a0abda241d57be67280c3d399e30581a826bced7e1b7b5.jpg", "reference_inputs/images/b3574f508469643224817c0968375f2dbc9b8fb5695e469f277f5de5349a47a5.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000063", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000063", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [650, 498, 662, 515]\n- T1: [690, 436, 704, 450]\n- T2: [672, 515, 683, 538]\n\nImage 2 (GROUND) labeled targets:\n- R: [749, 211, 782, 324]\n- T1: [515, 226, 545, 314]\n- T2: [786, 227, 818, 320]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3c008430886d13cc06b3af1ca92b44323c9fdc224aa91af5bf1a822dea3e92b5.jpg", "reference_inputs/images/7b76f625b0e713b81c3b2fd9005caf85291cd0af2b37d0152ae222104a3abd57.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000064", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000064", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [283, 215, 298, 224]\n- T1: [374, 61, 409, 134]\n- T2: [260, 281, 278, 290]\n\nImage 2 (GROUND) labeled targets:\n- R: [620, 251, 631, 300]\n- T1: [840, 259, 947, 309]\n- T2: [513, 250, 526, 303]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e15b5ab4486d18e49579e63edec4137ea576208c7b3ed60f09684c21f8bbd4b2.jpg", "reference_inputs/images/53f6413a5aece953d3d3f19040ae4fcc8e6ef179c92e51b1341438a643b7e920.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000065", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000065", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 4 shared labeled targets across the two images, including 3 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2, T3.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [670, 306, 682, 314]\n- T1: [673, 106, 684, 124]\n- T2: [692, 12, 706, 32]\n- T3: [879, 488, 953, 524]\n\nImage 2 (GROUND) labeled targets:\n- R: [826, 227, 858, 320]\n- T1: [212, 223, 237, 317]\n- T2: [2, 222, 35, 311]\n- T3: [850, 280, 894, 303]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: T2 -> T3 -> T1\nC: T3 -> T1 -> T2\nD: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/54fe527ada2d1060e9abe51e500c229043388ec7321da0fcfb22ea61353b10b0.jpg", "reference_inputs/images/794cfbcbe7aebf0fd07de502de177b39301e6da824fb87e6bb1e4aa716ca8342.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000066", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000066", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 4 shared labeled targets across the two images, including 3 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2, T3.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [490, 517, 525, 575]\n- T1: [279, 492, 297, 508]\n- T2: [719, 379, 735, 388]\n- T3: [673, 292, 686, 298]\n\nImage 2 (GROUND) labeled targets:\n- R: [463, 248, 543, 314]\n- T1: [120, 246, 137, 301]\n- T2: [771, 258, 781, 295]\n- T3: [678, 265, 686, 294]\n\nAnswer options:\nA: T2 -> T1 -> T3\nB: unable to determine\nC: T1 -> T2 -> T3\nD: T1 -> T3 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1a0ea3e0c43e1f8333c7b5ac26d86b0d4127ca6debff07ec7d82fe47ce562749.jpg", "reference_inputs/images/c6777d37805bc1e28ba259a064593108c7c27c287b70aceb3538e5e222e83c53.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000067", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000067", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [674, 36, 690, 51]\n- T1: [678, 370, 689, 384]\n- T2: [502, 350, 539, 425]\n\nImage 2 (GROUND) labeled targets:\n- R: [612, 273, 626, 290]\n- T1: [692, 262, 702, 295]\n- T2: [498, 278, 548, 300]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ef325bedc98b528fdeca5d4476bb350fd621ccf27ffa94819def6fde3cedd910.jpg", "reference_inputs/images/42ee3d96da42feea79428db2a7aaa9f59bcc8eab925e72270f906ef0ce0d1823.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000068", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000068", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 4 shared labeled targets across the two images, including 3 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2, T3.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [756, 191, 775, 200]\n- T1: [674, 392, 686, 403]\n- T2: [678, 146, 690, 159]\n- T3: [755, 370, 775, 378]\n\nImage 2 (GROUND) labeled targets:\n- R: [385, 243, 401, 304]\n- T1: [918, 232, 971, 322]\n- T2: [197, 218, 230, 314]\n- T3: [741, 246, 761, 306]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1 -> T3\nC: T3 -> T2 -> T1\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/425ae2a6e520d6325a6ef5ca44e168b8600759bb84f76528dbe67ab98b445533.jpg", "reference_inputs/images/3b84b5d82519d128b30a19c0cdb9e0532e599f2540e6a0fb617747af66d36169.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000069", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000069", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [243, 456, 261, 469]\n- T1: [570, 360, 601, 427]\n- T2: [447, 54, 478, 123]\n\nImage 2 (GROUND) labeled targets:\n- R: [257, 266, 266, 294]\n- T1: [561, 278, 590, 296]\n- T2: [475, 280, 493, 292]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/79637357a794d114c87fa20d24a320beb4983fb07bc013db84fecd8b57b8d6ee.jpg", "reference_inputs/images/267140f4919911df6db0b7cf8ed7c04c6d81dac27a9d963632b49ae48d93239f.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_30m/000070", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000070", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [240, 345, 258, 358]\n- T1: [246, 190, 270, 201]\n- T2: [232, 447, 253, 458]\n\nImage 2 (GROUND) labeled targets:\n- R: [389, 249, 402, 307]\n- T1: [701, 242, 721, 306]\n- T2: [206, 252, 222, 308]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4c684bc0fac895f0282015d43db3defc102f937d9368b066abb088876dbfb397.jpg", "reference_inputs/images/1dd8da1c6e720f3f329f418fddf85c33c24dc0d7eddb073383e93040bd79026b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_50m/000001", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [470, 190, 489, 234]\n- T1: [434, 12, 455, 58]\n- T2: [536, 289, 562, 367]\n\nImage 2 (GROUND) labeled targets:\n- R: [470, 278, 489, 292]\n- T1: [454, 282, 470, 290]\n- T2: [537, 250, 579, 294]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1336f727d14e0a43920ba0e565b7a3a66a5ef82e0946e2742228982971657d7b.jpg", "reference_inputs/images/326e309493b74656ab397a5c7fe4a7d782f7c238e5d6a88d3395323169ed9146.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_50m/000002", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [511, 294, 532, 337]\n- T1: [574, 458, 596, 503]\n- T2: [570, 162, 589, 203]\n\nImage 2 (GROUND) labeled targets:\n- R: [341, 290, 366, 305]\n- T1: [261, 295, 295, 304]\n- T2: [310, 292, 335, 308]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6335503211ddf6b2368041e62c88cd8c440ca0c151d187f98a89d26e416d6c09.jpg", "reference_inputs/images/606dd5fea012af2756b373ee21b6a9f4a7ff6eca37a4eb356fb07f1cb3476f3e.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_50m/000003", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [435, 1, 458, 46]\n- T1: [437, 381, 464, 458]\n- T2: [509, 334, 527, 374]\n\nImage 2 (GROUND) labeled targets:\n- R: [425, 281, 442, 293]\n- T1: [0, 18, 257, 336]\n- T2: [498, 264, 580, 320]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e4ff7a204c5127ccacc2cdbba88bc5213586c8fae62904c9c188f0a3e8929b50.jpg", "reference_inputs/images/34591fc1fa22610d2519ee5814ed778a1c518ffb2f2dfb3ab15e3acf8d937d50.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_50m/000005", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [496, 197, 522, 276]\n- T1: [530, 442, 551, 486]\n- T2: [574, 454, 597, 498]\n\nImage 2 (GROUND) labeled targets:\n- R: [481, 247, 523, 302]\n- T1: [467, 281, 482, 290]\n- T2: [430, 276, 450, 286]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f1b1fabf0f694b63c2ad925139ff55b74e5134f92f21fc72456f28796a8d42c4.jpg", "reference_inputs/images/30e6bd80fec7eeb677a4b663c71af085dbdb8ae9f65814a556f247652e14e01b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_50m/000007", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [472, 559, 487, 575]\n- T1: [434, 385, 456, 429]\n- T2: [507, 365, 527, 407]\n\nImage 2 (GROUND) labeled targets:\n- R: [0, 210, 334, 399]\n- T1: [331, 277, 393, 302]\n- T2: [500, 264, 544, 301]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5279243a189dac5df5abe787abc68a73bb0ecaf5b22deee4090ef8daf12ab644.jpg", "reference_inputs/images/876e52177c54d2f1a987ee7d6b60aef790c909d21871bccc19c29ad50f3249e4.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_50m/000009", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [437, 279, 459, 323]\n- T1: [434, 27, 453, 73]\n- T2: [506, 464, 527, 507]\n\nImage 2 (GROUND) labeled targets:\n- R: [406, 280, 442, 295]\n- T1: [448, 280, 464, 290]\n- T2: [500, 258, 562, 306]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8bc139be26f110733fffacdf00532d34cea4f956a85c5c19c599c4ebbb253eba.jpg", "reference_inputs/images/4b219c2e1a5ddc8d0638c9babe4ad1646d1f1623be3f4678702dfca40e824bbc.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_50m/000010", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [435, 322, 457, 366]\n- T1: [473, 206, 493, 249]\n- T2: [542, 435, 562, 478]\n\nImage 2 (GROUND) labeled targets:\n- R: [438, 282, 461, 292]\n- T1: [482, 278, 498, 290]\n- T2: [547, 274, 573, 294]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8e830185c444506afc6719687ec1c17624b01db7c88652e33518d29d0f62951b.jpg", "reference_inputs/images/0babe2b74ff3bef3cdac98263fe4524186cb1fd3c2bd927891a0e7a8de627389.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_50m/000011", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [499, 162, 527, 239]\n- T1: [511, 402, 526, 446]\n- T2: [473, 453, 493, 496]\n\nImage 2 (GROUND) labeled targets:\n- R: [337, 241, 417, 302]\n- T1: [425, 281, 446, 294]\n- T2: [472, 280, 484, 294]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d78eac01e2212aeff0e4dc193dad35c2de808cf46bddb2f94cf15c2e06f99339.jpg", "reference_inputs/images/772ae829aab097e813edc986df6337aa53a4dfb46188b630cff56caf74228b46.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_50m/000012", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [534, 226, 565, 302]\n- T1: [434, 0, 465, 70]\n- T2: [474, 520, 492, 561]\n\nImage 2 (GROUND) labeled targets:\n- R: [545, 244, 590, 296]\n- T1: [444, 262, 471, 291]\n- T2: [154, 250, 390, 352]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9825ac2cf22175ea4b1eccdc5ca6a7fbc946d54fd5eba1875ac2cbbcef46b1e9.jpg", "reference_inputs/images/f715261c31086ae963af6c95167737fcc81cedadfdf0f8460aaa5262d2a8969f.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_50m/000014", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [698, 346, 742, 368]\n- T1: [883, 392, 928, 414]\n- T2: [913, 318, 962, 338]\n\nImage 2 (GROUND) labeled targets:\n- R: [532, 281, 559, 294]\n- T1: [563, 282, 583, 290]\n- T2: [504, 282, 519, 290]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bf4e96f4b7d94ff33c01530a446d0874327259742aea607e7efd739a863ebae4.jpg", "reference_inputs/images/912001ea44deb2e90793bbecd6ab0407711e1c2fc6b14c92a7fff3b001b742e5.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_50m/000015", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [437, 538, 457, 575]\n- T1: [428, 183, 450, 227]\n- T2: [518, 245, 540, 289]\n\nImage 2 (GROUND) labeled targets:\n- R: [341, 265, 402, 302]\n- T1: [439, 282, 460, 291]\n- T2: [518, 282, 539, 292]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c0754c3fd9ed92cc477ef99dfe6d1ab619e52fe47f3d8eb86cfd86884958c025.jpg", "reference_inputs/images/121ca0a666b27ee44d92209d4b2bcecda33df73f1f6332b6dff01271e6619a5e.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_50m/000016", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [434, 27, 457, 71]\n- T1: [506, 452, 528, 496]\n- T2: [437, 490, 464, 567]\n\nImage 2 (GROUND) labeled targets:\n- R: [438, 282, 460, 292]\n- T1: [484, 269, 593, 323]\n- T2: [0, 30, 268, 333]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e2d3d48257038542e95080942944bed57b0dd925a0658170946298c77bad1501.jpg", "reference_inputs/images/ce6b08db9a41daef36af00206c758b526312c6cf41bef6347f4d72c299e9eb2b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_50m/000017", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [502, 301, 522, 344]\n- T1: [530, 515, 551, 558]\n- T2: [570, 37, 590, 78]\n\nImage 2 (GROUND) labeled targets:\n- R: [629, 275, 654, 293]\n- T1: [709, 274, 787, 305]\n- T2: [677, 279, 694, 289]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9e6c6109eafac788101cfd3b1b7088160bb23bf14895659d84c525d5f69ff798.jpg", "reference_inputs/images/2b6fff964b0a7b426d3c5321d16ca75da087b49c9e9fb3414269f22cf638b5e2.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_50m/000018", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [539, 523, 558, 569]\n- T1: [493, 260, 518, 337]\n- T2: [456, 344, 478, 388]\n\nImage 2 (GROUND) labeled targets:\n- R: [629, 251, 784, 328]\n- T1: [482, 245, 524, 296]\n- T2: [409, 278, 454, 298]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/435fb3438a5939da084e7fb0ad694d3be840746dfe99be4dbd160aa60bc58719.jpg", "reference_inputs/images/bb5fb5ea9e5337cd8439330bc758a7b429296eb13e3a974b49d53f06abdc07c5.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_50m/000019", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [539, 523, 558, 569]\n- T1: [456, 344, 478, 388]\n- T2: [493, 260, 518, 337]\n\nImage 2 (GROUND) labeled targets:\n- R: [545, 277, 570, 294]\n- T1: [463, 282, 483, 291]\n- T2: [496, 265, 517, 290]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/435fb3438a5939da084e7fb0ad694d3be840746dfe99be4dbd160aa60bc58719.jpg", "reference_inputs/images/bcdc1e69c0a36fdd6f81b6dc138bac8948640e1f401b454a6715a6941bd5a4a5.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_50m/000020", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [437, 296, 456, 338]\n- T1: [518, 2, 540, 46]\n- T2: [534, 536, 553, 575]\n\nImage 2 (GROUND) labeled targets:\n- R: [374, 269, 419, 298]\n- T1: [518, 282, 537, 291]\n- T2: [680, 214, 1023, 390]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/81f35adf479a310e3e4657e23a0d8b894b30bb03197b5adc41b6e6f750abcede.jpg", "reference_inputs/images/f3d0a4d31d16c204ff9a76a17bcb947c6cb4b7bbb5cf7f59f6e1d7ad3080d41a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_50m/000023", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [430, 440, 472, 482]\n- T1: [435, 126, 463, 208]\n- T2: [750, 553, 791, 575]\n\nImage 2 (GROUND) labeled targets:\n- R: [399, 280, 460, 295]\n- T1: [454, 266, 477, 290]\n- T2: [994, 277, 1023, 295]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5cc168c4f41abb4da2660f9d26603948f951d0fc700b54023fa10e550294abe6.jpg", "reference_inputs/images/d805d4a2c5ab72663e1777357194e745b5bdfd434697b3d4f6a88ba3b05e17b3.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/city6_ground_aerial_v2_50m/000024", "task_name": "multi_target_relation", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000024", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [436, 126, 458, 171]\n- T1: [506, 317, 527, 359]\n- T2: [475, 304, 493, 344]\n\nImage 2 (GROUND) labeled targets:\n- R: [383, 278, 406, 299]\n- T1: [464, 173, 702, 370]\n- T2: [161, 250, 392, 351]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d1c99ae970acc5d983e9425d72c5654176ba542ae95b32114c33ffcf0ce51a4c.jpg", "reference_inputs/images/1f14b718b819a632f51cf4ea6d5486cc08d1760dcb059aa48113fb4a35adc9bd.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000002", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [586, 402, 598, 412]\n- T1: [653, 429, 674, 440]\n- T2: [613, 389, 626, 399]\n\nImage 2 (GROUND) labeled targets:\n- R: [245, 322, 413, 374]\n- T1: [728, 315, 828, 344]\n- T2: [84, 299, 219, 343]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ef914f0e8dd6ab8d097e08449220a058e4bf7d703d7db775bb65dc43e21a8c4c.jpg", "reference_inputs/images/7a0bf23ec2ab312ed51cb9da293f10c6cf147291121dfec77e840170a2766309.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000003", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 4 shared labeled targets across the two images, including 3 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2, T3.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [565, 472, 570, 490]\n- T1: [450, 430, 466, 444]\n- T2: [504, 440, 518, 456]\n- T3: [707, 472, 713, 490]\n\nImage 2 (GROUND) labeled targets:\n- R: [623, 288, 632, 324]\n- T1: [410, 298, 439, 311]\n- T2: [498, 295, 523, 314]\n- T3: [912, 288, 924, 324]\n\nAnswer options:\nA: T2 -> T3 -> T1\nB: T2 -> T1 -> T3\nC: unable to determine\nD: T3 -> T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/532a7bb18fb47c2c7a6afad29eee2e22fe5703751da3632aa370b49a67654041.jpg", "reference_inputs/images/bb1b3b3955a3f2a6491e5176b2d74a97d2bd6453e48df238d57ef60e4f23c296.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000004", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [970, 125, 1023, 224]\n- T1: [628, 309, 645, 323]\n- T2: [654, 159, 670, 174]\n\nImage 2 (GROUND) labeled targets:\n- R: [358, 311, 465, 345]\n- T1: [674, 300, 743, 510]\n- T2: [141, 330, 194, 461]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9d4861e27c7af67f311d1114d48865240984cf9383313830a45ab9c71394e6ae.jpg", "reference_inputs/images/b28e4de0f5b59dbe34d9977ae7924f1d6f127b56999d871fee4297beb2c3d38a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000006", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [401, 511, 451, 541]\n- T1: [528, 438, 562, 454]\n- T2: [651, 565, 658, 575]\n\nImage 2 (GROUND) labeled targets:\n- R: [398, 291, 450, 310]\n- T1: [530, 290, 563, 298]\n- T2: [658, 289, 666, 314]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/17b62a74f74c1ea060246e106301a4bc1105d82173d30d693bf2f9d1dd4a700a.jpg", "reference_inputs/images/c050fcbf6a2fbb7ae73cec88f53be285420ba47b9cda22afa8774910d8ad59db.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000007", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 4 shared labeled targets across the two images, including 3 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2, T3.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [623, 445, 654, 458]\n- T1: [506, 547, 512, 573]\n- T2: [719, 412, 750, 426]\n- T3: [822, 542, 831, 568]\n\nImage 2 (GROUND) labeled targets:\n- R: [405, 290, 450, 307]\n- T1: [125, 290, 154, 338]\n- T2: [522, 289, 554, 297]\n- T3: [655, 287, 666, 320]\n\nAnswer options:\nA: T3 -> T2 -> T1\nB: unable to determine\nC: T2 -> T1 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d5bfe9a4746a03c92cbeee7518807d18504fafcbc4319a64d05fe881f42e9aab.jpg", "reference_inputs/images/daafd56b7f9b0584e99977588f276d6d9f561cd5540ec6e91fc46c16fc19f642.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000009", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [507, 122, 518, 142]\n- T1: [562, 420, 584, 434]\n- T2: [574, 510, 589, 539]\n\nImage 2 (GROUND) labeled targets:\n- R: [508, 286, 516, 317]\n- T1: [562, 294, 596, 338]\n- T2: [598, 290, 617, 355]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/98588555a9656d7cfc9f63266d2fcf2b6eb3064482278c14ddac2c450e16d2c2.jpg", "reference_inputs/images/c32b20010a9f85d7f75228c727ce166db866aa2ad664200bc62120323d137ad8.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000010", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [301, 252, 361, 360]\n- T1: [486, 373, 534, 482]\n- T2: [654, 179, 674, 197]\n\nImage 2 (GROUND) labeled targets:\n- R: [538, 302, 574, 319]\n- T1: [668, 298, 722, 325]\n- T2: [726, 287, 731, 311]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d9e96895d6c2797f50c2aea2b79ee35dc34e37c5d5b54b575d4db8cd5807d9e2.jpg", "reference_inputs/images/7e335e03c825200cce7a9e42c6bbfaa0e960ed5feec24233d18da53e6340c06c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000011", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [186, 478, 218, 502]\n- T1: [289, 421, 310, 436]\n- T2: [366, 371, 383, 392]\n\nImage 2 (GROUND) labeled targets:\n- R: [726, 289, 741, 331]\n- T1: [840, 290, 857, 340]\n- T2: [959, 295, 990, 350]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/04ae5b63297ef9faa39e0c51d2a61dcbc1fd82e72a57a0281f9362ae3f704f66.jpg", "reference_inputs/images/e6a930cc13c9435d48975c505cbff4ed5d6ffbc32bde3d87755e13d702c854ea.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000012", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [422, 474, 463, 494]\n- T1: [354, 522, 359, 542]\n- T2: [389, 548, 442, 575]\n\nImage 2 (GROUND) labeled targets:\n- R: [438, 291, 474, 305]\n- T1: [354, 289, 360, 308]\n- T2: [390, 298, 442, 314]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8d623e668ea8bf1fb46683001b988f91917220a05ed970bf04e08492689ac461.jpg", "reference_inputs/images/f598fb5a843da633d25b3a73385b1f3542e3af3c5ae2199248979d89ad1ff582.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000013", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [414, 429, 614, 575]\n- T1: [806, 530, 836, 561]\n- T2: [707, 100, 822, 187]\n\nImage 2 (GROUND) labeled targets:\n- R: [606, 270, 674, 322]\n- T1: [516, 293, 526, 330]\n- T2: [896, 317, 1023, 353]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ecc4acc96adf7c3979c31ffb91951a9a19ba78253b731c04299d813779925993.jpg", "reference_inputs/images/d81a3b74561f4bd0c14ad133335196ab668e38d1ba1c358f915caf128e382d2a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000015", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [591, 406, 605, 418]\n- T1: [588, 525, 594, 546]\n- T2: [581, 511, 586, 534]\n\nImage 2 (GROUND) labeled targets:\n- R: [626, 292, 648, 304]\n- T1: [714, 294, 729, 350]\n- T2: [677, 290, 691, 342]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5decbba8daa7c5df5a7fff7c056b20766744bd23a17ecce6a98368f20a0ea9b9.jpg", "reference_inputs/images/b85a6b29f399b960fda0517b94c5f99319b190ff4a324961395b1f408498e61b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000016", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [543, 466, 594, 575]\n- T1: [366, 307, 383, 328]\n- T2: [725, 182, 780, 288]\n\nImage 2 (GROUND) labeled targets:\n- R: [522, 297, 604, 344]\n- T1: [506, 290, 513, 317]\n- T2: [770, 292, 826, 328]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3d91c45a797db089579e59426df5d2694d63f16a52b3e2a02525e4ff3e3856ec.jpg", "reference_inputs/images/76ed8b4b58e6d6a91d059e0a8dce2a568ffed197666f9d6c66a6ff036a79e92d.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000017", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [330, 538, 338, 562]\n- T1: [383, 531, 440, 562]\n- T2: [408, 504, 456, 530]\n\nImage 2 (GROUND) labeled targets:\n- R: [950, 298, 983, 347]\n- T1: [975, 325, 1023, 351]\n- T2: [805, 309, 876, 339]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/45161aa0655e931e7924983dd2f4d431231d378aad482002b7e8acb1c3c8ed8a.jpg", "reference_inputs/images/dce70d33d8f9602df8bca856a63b807f674d48cc5394e78f6422eb83339193aa.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000021", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [506, 0, 620, 55]\n- T1: [653, 330, 676, 346]\n- T2: [694, 376, 720, 397]\n\nImage 2 (GROUND) labeled targets:\n- R: [254, 331, 362, 391]\n- T1: [761, 293, 783, 352]\n- T2: [798, 294, 814, 346]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a86b10af6f61a2902a8ff2549f52ec7d4696fe04c8625701878839ce4cffcbbc.jpg", "reference_inputs/images/b222eeecc0a0b3660e5632c382b7f4bdbff21bc456f06d0165822026551dda65.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000022", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [332, 432, 347, 445]\n- T1: [347, 413, 362, 422]\n- T2: [298, 453, 334, 494]\n\nImage 2 (GROUND) labeled targets:\n- R: [740, 296, 826, 328]\n- T1: [927, 325, 1023, 369]\n- T2: [641, 273, 697, 311]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2d9c0d21686be8efdd9a86bf727dc62b7458f9c9b66e2b26748d02e03988a651.jpg", "reference_inputs/images/696db644fc0e650348f4be016e8e5fd9c181eee2de39d8b0f2f71ba4a8fde652.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000023", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [595, 383, 618, 389]\n- T1: [511, 385, 533, 392]\n- T2: [698, 382, 721, 390]\n\nImage 2 (GROUND) labeled targets:\n- R: [717, 298, 778, 311]\n- T1: [529, 298, 568, 313]\n- T2: [977, 294, 1023, 310]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fc020544619452bceeb4c01d861680999eff0a57983a98dd5efea5e57ebef25d.jpg", "reference_inputs/images/6c1faeec1b212385a6473aba22bc10edfcccad8ee4a4d132b5fa6116fc167924.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000026", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 4 shared labeled targets across the two images, including 3 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2, T3.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [575, 0, 589, 22]\n- T1: [899, 211, 938, 226]\n- T2: [658, 364, 678, 379]\n- T3: [871, 98, 908, 118]\n\nImage 2 (GROUND) labeled targets:\n- R: [493, 290, 498, 311]\n- T1: [340, 294, 349, 322]\n- T2: [306, 291, 311, 314]\n- T3: [419, 291, 426, 320]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T3 -> T2\nC: T1 -> T2 -> T3\nD: T2 -> T1 -> T3\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/96282862df9fed1e54bdafbd1cefb080486775ae3b35c0d1cf0a80741f19ab67.jpg", "reference_inputs/images/6c19efe1630f6eca1709b8bf989300cf0ebc534fdb36b20b60aeb358ef605f59.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000027", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [797, 118, 853, 217]\n- T1: [626, 551, 643, 575]\n- T2: [970, 370, 1023, 468]\n\nImage 2 (GROUND) labeled targets:\n- R: [679, 299, 722, 318]\n- T1: [638, 288, 662, 338]\n- T2: [888, 305, 985, 333]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b8200a86bbc7cdf8a4e6f13a73c77702bd6d6aef0e6249858945696f82d3c305.jpg", "reference_inputs/images/8e90901750b16872fdaa935ea78ac05d94b56c4f0c1df5921fade7be28e20261.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000028", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [961, 496, 1023, 575]\n- T1: [652, 297, 760, 382]\n- T2: [769, 289, 875, 368]\n\nImage 2 (GROUND) labeled targets:\n- R: [665, 377, 790, 423]\n- T1: [0, 360, 202, 465]\n- T2: [157, 335, 383, 410]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0671ae8ee169440e6627d92e84b2ba53644dd471a3dc9097fb5bbae57d101527.jpg", "reference_inputs/images/dc50adabea5c473de95da4c3e3bf5ddb25880f2dd43b3eb3b1e235dc6a75a193.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000029", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [506, 455, 524, 474]\n- T1: [410, 519, 436, 553]\n- T2: [445, 429, 459, 443]\n\nImage 2 (GROUND) labeled targets:\n- R: [657, 294, 705, 320]\n- T1: [572, 310, 646, 361]\n- T2: [526, 294, 548, 310]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/eecc5fd6027c2498737cfd1fb08d36e7c6fe4d6927d160f242cd9a1edbb2dc0e.jpg", "reference_inputs/images/0270e6c49c8feef3c5b6875c7c99e75229547dceb165bee3d8a659b59ba4fe54.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000031", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000031", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [297, 9, 320, 42]\n- T1: [336, 266, 355, 285]\n- T2: [393, 384, 415, 401]\n\nImage 2 (GROUND) labeled targets:\n- R: [814, 290, 822, 317]\n- T1: [747, 289, 757, 326]\n- T2: [728, 293, 748, 337]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fb55ae6fe0a5e87f2a974b92788564c804ef2b673d2f3f9e41ec8f55f4f06477.jpg", "reference_inputs/images/8acb44882ad014a1644004c081aaf072d078f577d9277171e12f3dedf544739b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000032", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [629, 465, 654, 485]\n- T1: [658, 435, 679, 450]\n- T2: [602, 423, 618, 436]\n\nImage 2 (GROUND) labeled targets:\n- R: [758, 306, 839, 330]\n- T1: [774, 295, 828, 314]\n- T2: [664, 294, 698, 311]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c391ab20637d039794ce863c1b9db7c92b6d6325ce8db865c9cab7217923f7de.jpg", "reference_inputs/images/fa38e654e7d24f439480ca85078b43de575e08da51ef24adf0cb83b923b47e43.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000033", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000033", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [149, 414, 163, 426]\n- T1: [208, 367, 222, 385]\n- T2: [172, 378, 183, 389]\n\nImage 2 (GROUND) labeled targets:\n- R: [530, 302, 566, 325]\n- T1: [577, 276, 609, 303]\n- T2: [528, 286, 550, 303]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d5147459caeae67ce54f9d4151e6266c145f10f639a335c96f481cd5daf5010b.jpg", "reference_inputs/images/700b82317f7b7799736a2401b76a6c47038f11fbd151a0c02396195b2dcfde77.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000034", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000034", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [434, 470, 454, 490]\n- T1: [506, 437, 519, 453]\n- T2: [504, 498, 523, 522]\n\nImage 2 (GROUND) labeled targets:\n- R: [413, 294, 438, 311]\n- T1: [503, 289, 522, 302]\n- T2: [502, 298, 527, 315]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/58535b94fbff0387d3d936599a30966b8283f1f79f82b2557b1e6cb15cb992b7.jpg", "reference_inputs/images/d364b0e3541250a939545692adbf2b7f5dd4eda4eed7a82b53526764b5620522.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000035", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000035", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 4 shared labeled targets across the two images, including 3 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2, T3.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [522, 475, 543, 497]\n- T1: [413, 504, 418, 524]\n- T2: [605, 474, 626, 493]\n- T3: [450, 498, 455, 517]\n\nImage 2 (GROUND) labeled targets:\n- R: [408, 297, 436, 310]\n- T1: [254, 287, 262, 311]\n- T2: [505, 291, 528, 308]\n- T3: [306, 289, 311, 310]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: T1 -> T3 -> T2\nC: T3 -> T1 -> T2\nD: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/07ca824b99b314bd94d57eda78a95cebe85ce30bc8e7d8698d800d83e704a7c3.jpg", "reference_inputs/images/1075720d899aa88c5aa14cb7eeb20344e150199624b9bde5c85be648db3d26cb.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000037", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000037", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [155, 302, 186, 314]\n- T1: [314, 361, 370, 470]\n- T2: [151, 436, 183, 457]\n\nImage 2 (GROUND) labeled targets:\n- R: [510, 290, 522, 318]\n- T1: [584, 302, 641, 334]\n- T2: [469, 292, 478, 325]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c97a9f8f7de6a094a1f743d885b8fb4e23676ff22a3149bfbf2818ff7fc8494b.jpg", "reference_inputs/images/1634592f4f73b92a835b205ab6e70cd9d32472082667dd9479c0f2f0c1eb1a71.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000038", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000038", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [707, 266, 737, 279]\n- T1: [507, 197, 518, 213]\n- T2: [577, 276, 590, 288]\n\nImage 2 (GROUND) labeled targets:\n- R: [625, 290, 632, 304]\n- T1: [510, 287, 516, 310]\n- T2: [550, 292, 559, 313]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/38b14bfd106225a2e22f062250eb7d51bddb58c2ae67d4a7ab3f274540ea8a91.jpg", "reference_inputs/images/57ae96560ad6020664e5a95327c1992dcf7337d3980ecf843a71447b788d5bfd.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000039", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000039", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [903, 22, 948, 47]\n- T1: [401, 382, 414, 395]\n- T2: [887, 546, 935, 574]\n\nImage 2 (GROUND) labeled targets:\n- R: [583, 289, 592, 308]\n- T1: [251, 288, 262, 319]\n- T2: [636, 289, 655, 330]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/71627e38503f8fa008a5c48ecbf7cb63719a3f00c0377de6c1e90fbd4fd13671.jpg", "reference_inputs/images/18af91d51cab25cd269148c24de3e6c462001bcbe4cdf2ad120569133edaefec.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000040", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000040", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [371, 435, 393, 446]\n- T1: [624, 535, 671, 575]\n- T2: [621, 518, 659, 546]\n\nImage 2 (GROUND) labeled targets:\n- R: [461, 334, 557, 391]\n- T1: [811, 298, 872, 312]\n- T2: [760, 302, 815, 317]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d76465ef7fbd5baaf0fe865c9d44135d8dba217611268bbefb6586e6993147b2.jpg", "reference_inputs/images/c52d3ed15c8ad4699b4bc938831c5b06416414b65cdf6d09847efe920d0a1925.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000041", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000041", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [506, 446, 518, 468]\n- T1: [710, 521, 722, 530]\n- T2: [577, 520, 590, 543]\n\nImage 2 (GROUND) labeled targets:\n- R: [507, 288, 516, 321]\n- T1: [674, 292, 698, 316]\n- T2: [565, 295, 576, 326]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/671c05560af838f8e7ce8cfb9c707b838fff0c85439a5c7e4b8e942e3e784aab.jpg", "reference_inputs/images/0e21b79909dbcad1e395e42f1bc899edc1a61134bc233babed9b8aeb536d73de.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000042", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000042", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [242, 550, 271, 575]\n- T1: [575, 550, 590, 574]\n- T2: [618, 353, 634, 369]\n\nImage 2 (GROUND) labeled targets:\n- R: [325, 287, 332, 310]\n- T1: [561, 294, 567, 320]\n- T2: [573, 286, 578, 312]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/40d0ce2fd4a56089ea8837e202fe35b54cc6d6a9d77031cfb4521ab135148a32.jpg", "reference_inputs/images/f00be5f5431d5f216408980ab71683bd8405197facf8cec2f1e73396ba223dd5.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000043", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000043", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [434, 283, 456, 294]\n- T1: [700, 43, 723, 70]\n- T2: [557, 519, 590, 540]\n\nImage 2 (GROUND) labeled targets:\n- R: [423, 290, 428, 311]\n- T1: [263, 294, 271, 312]\n- T2: [353, 290, 363, 308]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ad6fe422531a93b9485098c6c3e4bc860f8aeb03907c9f838ade835d139050c2.jpg", "reference_inputs/images/d2fdc4683ecf48ad9dd5cf6d1739017f93cacc73846c598a25716dd0416a30a8.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000044", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000044", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [588, 374, 608, 379]\n- T1: [347, 505, 354, 525]\n- T2: [308, 530, 314, 553]\n\nImage 2 (GROUND) labeled targets:\n- R: [670, 291, 696, 297]\n- T1: [394, 290, 410, 338]\n- T2: [297, 289, 316, 357]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5bc11b7f2e38ac137619a4800f8fe2d0896a00695683ac0192ea35c976912819.jpg", "reference_inputs/images/ef1c9e15657d46d314b74cd595a8bf5a6aa2686d5fedec13ec8318fc99ba7f6c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000045", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000045", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [731, 166, 840, 269]\n- T1: [917, 504, 962, 535]\n- T2: [965, 14, 1003, 59]\n\nImage 2 (GROUND) labeled targets:\n- R: [608, 310, 835, 409]\n- T1: [998, 294, 1018, 350]\n- T2: [420, 297, 444, 342]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1aedfd6862af453cdd58f1227e627eccd860b652ccd30b1767d50432e1f9439a.jpg", "reference_inputs/images/9d7dd89769e249a4eddd5467312146be0428ac4007260c0ea7d8f0137e5d7c62.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000046", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000046", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [618, 486, 635, 511]\n- T1: [498, 139, 546, 247]\n- T2: [639, 58, 666, 81]\n\nImage 2 (GROUND) labeled targets:\n- R: [656, 294, 673, 352]\n- T1: [505, 299, 539, 327]\n- T2: [605, 287, 609, 316]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/927a6db3097148ac7af46cf66cc1d6625a7eb825690a7f4628d4b84905873f8a.jpg", "reference_inputs/images/839e7480594c685cc455a57a7f79881f7053bafcfd91d1204a3a1244770b77d3.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000047", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000047", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [582, 435, 599, 450]\n- T1: [530, 466, 547, 486]\n- T2: [610, 494, 634, 517]\n\nImage 2 (GROUND) labeled targets:\n- R: [0, 310, 190, 402]\n- T1: [502, 405, 1023, 575]\n- T2: [854, 336, 1023, 405]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/654c3d84068fdb72b52ef550344c62e961a4fa99698a25e2abbf8a2d86719353.jpg", "reference_inputs/images/e01531ab97172b6d6fc487648c467a07105722883b7d20cb602749df943d7e66.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000048", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000048", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [546, 254, 556, 266]\n- T1: [347, 158, 367, 178]\n- T2: [856, 361, 872, 367]\n\nImage 2 (GROUND) labeled targets:\n- R: [540, 288, 544, 325]\n- T1: [400, 290, 408, 319]\n- T2: [840, 288, 857, 332]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bd8e20d934fbc45d672e81c0db3b04487695cb9e586a6a8c0598d67457e2bca8.jpg", "reference_inputs/images/1ce871a55fdec50c152f440c6573c7c01bf5491413c70e4e355dfc4ae58391eb.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000049", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000049", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [958, 455, 1000, 473]\n- T1: [918, 0, 951, 17]\n- T2: [633, 458, 649, 475]\n\nImage 2 (GROUND) labeled targets:\n- R: [385, 293, 403, 322]\n- T1: [203, 287, 213, 312]\n- T2: [51, 295, 75, 346]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fce96307898600f8aa5eac561a501e327b9732a3917600d3ba991883de1c00b7.jpg", "reference_inputs/images/23d606339b44ef900e4f4b8fd734f19f809273819df9d7e7aac8b402b5896432.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000050", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000050", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [694, 450, 717, 468]\n- T1: [609, 208, 639, 223]\n- T2: [785, 497, 846, 575]\n\nImage 2 (GROUND) labeled targets:\n- R: [624, 286, 638, 315]\n- T1: [562, 288, 569, 308]\n- T2: [702, 298, 803, 324]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f3f337c634a9cc8fb5161273331a95970b7f1312fe185c12bb1a6b550e321af6.jpg", "reference_inputs/images/be43c524ae2c0a4f5cba49ce67e3a33f7ca881c5b79e6c1eb61e6f8372beb563.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000051", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000051", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [694, 473, 724, 490]\n- T1: [594, 421, 611, 433]\n- T2: [457, 517, 502, 538]\n\nImage 2 (GROUND) labeled targets:\n- R: [765, 299, 817, 314]\n- T1: [615, 294, 638, 303]\n- T2: [414, 304, 496, 325]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/09e4d63e55bd5f22dd2cec7666efaac4ee4d83856b562dc1c9726bfdc0600797.jpg", "reference_inputs/images/8c822065e6a4a3137d25a6abc18bb3c2c267e21f7d088cad1a3a17dbee7b5c2a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000052", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000052", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [156, 530, 162, 552]\n- T1: [335, 484, 358, 506]\n- T2: [255, 545, 286, 575]\n\nImage 2 (GROUND) labeled targets:\n- R: [362, 288, 368, 307]\n- T1: [532, 290, 562, 309]\n- T2: [408, 292, 447, 318]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1b51b7b6b84ff1a9d8e198eaf5e87d3a7a4fa6656b294051d6264f679cc47f33.jpg", "reference_inputs/images/d31f007add278cf876ac86158736bcb1f97f1402eaef053294693069319ac645.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000053", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000053", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [799, 486, 807, 505]\n- T1: [629, 465, 654, 485]\n- T2: [658, 435, 679, 450]\n\nImage 2 (GROUND) labeled targets:\n- R: [638, 291, 648, 325]\n- T1: [426, 325, 592, 371]\n- T2: [173, 305, 297, 342]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c391ab20637d039794ce863c1b9db7c92b6d6325ce8db865c9cab7217923f7de.jpg", "reference_inputs/images/cbdd1a6ed8b5b494d10165aa541a404c08ab908f6e9e23b313eeebf05ec1f2b2.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000054", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000054", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [459, 446, 478, 464]\n- T1: [302, 187, 330, 201]\n- T2: [214, 83, 244, 110]\n\nImage 2 (GROUND) labeled targets:\n- R: [631, 292, 639, 330]\n- T1: [583, 291, 589, 315]\n- T2: [562, 290, 567, 311]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a4968ce1b55bf0be69a034655155c8cadb8cd1afd0efe8ff7198b89dd500688a.jpg", "reference_inputs/images/3a6c37ec5821918390e2d7fe5f5eced40254021cdba472145ae816c701d4959b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000055", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000055", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [930, 247, 1023, 303]\n- T1: [114, 122, 222, 178]\n- T2: [52, 546, 100, 575]\n\nImage 2 (GROUND) labeled targets:\n- R: [474, 294, 496, 301]\n- T1: [394, 300, 422, 311]\n- T2: [580, 289, 586, 309]\n\nAnswer options:\nA: T2 -> T1\nB: unable to determine\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b31cbbae1559d1aa0a7a38ebb3cbe9cf1fc7511cd0a6c9b152204e9b2d52ad7d.jpg", "reference_inputs/images/23d32f7a6776881bab4902b18c060313973797d4d8e2e4d0cb1a413c41e67a4c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000056", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000056", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 4 shared labeled targets across the two images, including 3 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2, T3.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [301, 122, 356, 228]\n- T1: [653, 66, 677, 90]\n- T2: [89, 122, 131, 140]\n- T3: [649, 333, 667, 346]\n\nImage 2 (GROUND) labeled targets:\n- R: [396, 292, 429, 314]\n- T1: [583, 286, 587, 309]\n- T2: [305, 287, 312, 310]\n- T3: [602, 288, 608, 317]\n\nAnswer options:\nA: T1 -> T2 -> T3\nB: unable to determine\nC: T2 -> T1 -> T3\nD: T2 -> T3 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b93ab76ddb5a25b217c5c906a07d5efb3a35760afac1f9e5f0eaea7e0b6e0003.jpg", "reference_inputs/images/30142c91b4a377fd8bb01708b286c070258fd5c4a84e7771a9bec6c4b1def4c8.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000057", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000057", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [658, 562, 674, 575]\n- T1: [695, 122, 721, 140]\n- T2: [138, 486, 174, 510]\n\nImage 2 (GROUND) labeled targets:\n- R: [649, 292, 667, 330]\n- T1: [608, 287, 618, 310]\n- T2: [222, 288, 231, 325]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b93ab76ddb5a25b217c5c906a07d5efb3a35760afac1f9e5f0eaea7e0b6e0003.jpg", "reference_inputs/images/30142c91b4a377fd8bb01708b286c070258fd5c4a84e7771a9bec6c4b1def4c8.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000058", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000058", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [730, 417, 839, 520]\n- T1: [959, 277, 998, 302]\n- T2: [382, 374, 482, 476]\n\nImage 2 (GROUND) labeled targets:\n- R: [735, 296, 813, 342]\n- T1: [838, 290, 854, 321]\n- T2: [388, 296, 488, 336]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/593aba1a5f50088dcad24abfed915ae977e9337ecee2638c883220b5be6434b8.jpg", "reference_inputs/images/c78a0378bceef9047b848116311ccaa5e549a203c8ea138f175b0306b5e4c7b7.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000059", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000059", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [862, 466, 899, 491]\n- T1: [750, 119, 806, 227]\n- T2: [890, 38, 938, 62]\n\nImage 2 (GROUND) labeled targets:\n- R: [665, 298, 689, 358]\n- T1: [16, 322, 270, 398]\n- T2: [109, 290, 133, 324]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/dab276950498d69db9d901ba43a79e5efbd434f50527250358ba058437224aa7.jpg", "reference_inputs/images/c94dbab8a34ba2e3a307856aa254444e784933a90da7257f8df60d4f34f2733b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000060", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000060", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [644, 430, 676, 442]\n- T1: [730, 490, 767, 510]\n- T2: [632, 475, 654, 494]\n\nImage 2 (GROUND) labeled targets:\n- R: [610, 306, 662, 320]\n- T1: [833, 302, 889, 315]\n- T2: [869, 307, 954, 324]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d4e07913a5a14259f4b28417bcd90f7478a358c22747ce6cf3e6cfa47b710b39.jpg", "reference_inputs/images/6aeff7f4a3fe7015e7da04d5c9d75664c3247c67da7bc0db28bb73528a27a828.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000061", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000061", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [561, 158, 581, 173]\n- T1: [870, 362, 906, 374]\n- T2: [898, 470, 935, 488]\n\nImage 2 (GROUND) labeled targets:\n- R: [543, 290, 564, 318]\n- T1: [862, 289, 875, 331]\n- T2: [990, 294, 1009, 345]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4f1d592425f43ddd44ef33551b00627e3bfa2b20112e736d3202b2b1e08e70c9.jpg", "reference_inputs/images/5aa2727bedc35e1a6443c7568f420d63beeb311a931ccbf27122807cd271e2ca.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000062", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000062", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [569, 170, 585, 187]\n- T1: [551, 437, 562, 453]\n- T2: [820, 307, 879, 415]\n\nImage 2 (GROUND) labeled targets:\n- R: [562, 286, 574, 327]\n- T1: [573, 302, 587, 362]\n- T2: [866, 300, 1022, 358]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d0048a3c1c294547ebe61642ac6915062a9d96f63aae94ff790904dbd788c32e.jpg", "reference_inputs/images/8f4940264acc56f4ce551a90fee909e68ecf5da8ccb78f8c6385b3774c7c049d.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000063", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000063", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [716, 217, 742, 229]\n- T1: [151, 440, 192, 461]\n- T2: [0, 147, 34, 162]\n\nImage 2 (GROUND) labeled targets:\n- R: [453, 287, 460, 310]\n- T1: [12, 289, 39, 327]\n- T2: [1, 290, 17, 315]\n\nAnswer options:\nA: unable to determine\nB: T1 -> T2\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/62b21a2b843cd73f959a114c90ba7721502bba41770522608b3cf2779d5083c2.jpg", "reference_inputs/images/ba1dd72c489a349f7b0ce72aff1a61fec69dba3647c2dca0d283dd44fc578db2.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000064", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000064", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [744, 453, 771, 479]\n- T1: [561, 126, 626, 238]\n- T2: [295, 390, 356, 500]\n\nImage 2 (GROUND) labeled targets:\n- R: [736, 288, 742, 306]\n- T1: [620, 297, 642, 306]\n- T2: [558, 296, 579, 311]\n\nAnswer options:\nA: T2 -> T1\nB: T1 -> T2\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5c0e4b7fd28d747f1c4c10d28e50b31f60ce5a6fbd644a44f8b265ae00c19b94.jpg", "reference_inputs/images/9a8ae0b3d983add7ddedd2d91a1213471f9aed9b1ff95a593c93278990174473.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000065", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000065", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [386, 440, 400, 466]\n- T1: [406, 134, 419, 150]\n- T2: [359, 254, 385, 269]\n\nImage 2 (GROUND) labeled targets:\n- R: [162, 296, 190, 356]\n- T1: [597, 293, 617, 359]\n- T2: [430, 293, 458, 352]\n\nAnswer options:\nA: T1 -> T2\nB: unable to determine\nC: T2 -> T1\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0238f2f24037ed86b818cbce1240597128732f2ff9a788e59d4209d2b8c6fcfc.jpg", "reference_inputs/images/e376e3cca0dbc79868cc3dfd93002d9db47d9f17f23b2659d83b8ba6ab8e5f26.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000066", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000066", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [115, 372, 223, 429]\n- T1: [578, 470, 590, 494]\n- T2: [589, 146, 602, 166]\n\nImage 2 (GROUND) labeled targets:\n- R: [173, 307, 274, 331]\n- T1: [576, 294, 585, 336]\n- T2: [558, 286, 565, 314]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/40f61ea9edb503112601c62b3f52b58bfbfc3755589baafaa08d713e3b776977.jpg", "reference_inputs/images/12ff10e2b317b9f7c0aff4ccda2d4142e0e5b16c61dab7dfee15d3580d0725a9.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000067", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000067", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from nearest to farthest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [658, 438, 681, 454]\n- T1: [485, 422, 506, 432]\n- T2: [630, 406, 647, 414]\n\nImage 2 (GROUND) labeled targets:\n- R: [499, 292, 528, 304]\n- T1: [389, 300, 421, 314]\n- T2: [345, 295, 381, 303]\n\nAnswer options:\nA: T1 -> T2\nB: T2 -> T1\nC: unable to determine\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/49d6755483780222f4b2c746fdedcb4f0a357cfbfc7fee28d788d62ac63643b6.jpg", "reference_inputs/images/0c6b408cdf2cefd100f4a37aefd42ebf656c8b482363c1a627695c1bb2fcbf1d.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "multi_target_relation/town1_ground_aerial_v2/000068", "task_name": "multi_target_relation", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/multi_target_relation/town1_ground_aerial_v2/all_questions.json", "question_id": "000068", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Multi-Target Relation.\nYou will receive exactly two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nThere are 3 shared labeled targets across the two images, including 2 comparison targets.\nThe reference target is labeled R. The comparison targets are labeled T1, T2.\nDetermine the order of the comparison targets from farthest to nearest relative to the reference target.\nUse the cross-view layout, scene geometry, target spacing cues, and context visible in both images.\nIf the ordering truly cannot be determined from the provided evidence, choose the 'unable to determine' option.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- The same label refers to the same shared target across both images.\n- Label R is the reference target.\n\nImage 1 (AERIAL) labeled targets:\n- R: [517, 375, 537, 388]\n- T1: [589, 146, 602, 166]\n- T2: [578, 470, 590, 494]\n\nImage 2 (GROUND) labeled targets:\n- R: [522, 308, 562, 392]\n- T1: [594, 288, 607, 338]\n- T2: [730, 314, 774, 468]\n\nAnswer options:\nA: unable to determine\nB: T2 -> T1\nC: T1 -> T2\n\nTask: choose the option that gives the correct ordering of the comparison targets.\nBenchmark mode instruction: respond with only one option letter from A, B, C. Do not output any explanation or extra text.", "images": ["reference_inputs/images/40f61ea9edb503112601c62b3f52b58bfbfc3755589baafaa08d713e3b776977.jpg", "reference_inputs/images/c1af4763b62bb1c4e18c6f2872b14ca427b73e8e71ef4b9b97e5add7b49f9e43.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000002", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [902, 241, 927, 306].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/e4d923360efb0a11f4dcd6ff2db5b67792540a723ae2ce828c607a4bb8d7e216.jpg", "reference_inputs/images/f74c703367de7b2fc1295542d39d0c937f0335d7cc3c1046ad03f832839e7815.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000003", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [389, 250, 442, 295].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/4d94ae80e8ba04008e857d425712d1b5485e3b7c1903015d02007d4e2f266833.jpg", "reference_inputs/images/ff9f872d28e676633b1aaec5c24f8e3560e55512fa951d2b82602e48a167bf6b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000004", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [827, 247, 845, 303].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/9d12e0d7e2f66043454da17dd4d6bf2b971c6c1620f747bf4be58db60909059e.jpg", "reference_inputs/images/334b4e4d96aa2cc5e2ef412bf4ca996a78d03fc955f3ccf5b8987a427f0f5fb7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000006", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [497, 406, 515, 435].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/4c4ce9a2a452eb6f170a53a1fc9424444112c57777f71b6b9064acc133562968.jpg", "reference_inputs/images/f4a7098cbe94ae907eafc0f5997b8a099436c210eb839ecea4f7bac29d27b2a5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000007", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [746, 247, 778, 305].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/6668b014712035c16c55c9d346ec9d65f9aea0505cf77f2f90be5d06ec528252.jpg", "reference_inputs/images/903023c225c149cedd651170969d99e1abf3ce6bbe19bfa15fba8caa56e5ca42.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000008", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [851, 183, 910, 336].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/6f8268f4a1b4de318f12616aabc34ebe6f28ba8ef563703d6abe227cc94b8015.jpg", "reference_inputs/images/976da8dfa3423941abd9793286db98d6bd73430d1dc81d0128956bb2ae23436b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000011", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [490, 244, 680, 366].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/1c86cc5951e53a469d0dfa1d49811abb2200a75b723703e0b03968751a5c81cc.jpg", "reference_inputs/images/5efc52003cf953f94e3968827080733c9c7d3c87a5365c1c13b75ff2f9a21f03.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000012", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [492, 275, 557, 307].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/4361054b280cd9d73394edc6e371afc818b989aed9f9327cd8d6243c79133116.jpg", "reference_inputs/images/09d6bbd65d5060eae93487145b4ecd97a46eec4da32909caff0ec56456d840de.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000015", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [841, 253, 857, 298].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/93b9920d0da98bcb158997ea9b6aee139f8868b3e2df6767cd5a601a3e6c1e43.jpg", "reference_inputs/images/e0f97d1dbbb520d258aa80b8177f8e8a06c5310d6dbb6a555b54fa0664011e66.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000016", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [769, 277, 802, 298].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/f84e51c89c7aaeb4f164fd1217f5b807fcbcb5695f418ece8dc950712607074b.jpg", "reference_inputs/images/8bc8e89b82da28dd3eb342c5c2d8ae3aa6bd8756a9bcf90ee17dd356cdad3e46.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000017", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [955, 262, 966, 294].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/cc432e0bbc20b1fa31132acb84df4fcdde4420d63cdec092f34992c12a82e331.jpg", "reference_inputs/images/302afa3fe401318929f9af4cedaf36e9720f28d6ca384fec6c40c86710b61e35.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000019", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [497, 247, 615, 334].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/423edee79517840c130357a8a1a08ab8664be9cec297a266fc992caea0152d92.jpg", "reference_inputs/images/3e21d07491e09cd36408a4cbcdacf1e77d7462660023e5288805dc80014a0207.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000020", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [505, 526, 532, 555].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/6a5e13694f5f116e2e93f37ae27001d2424ab7e44cf25ed4c982edd446cbccd4.jpg", "reference_inputs/images/34591fc1fa22610d2519ee5814ed778a1c518ffb2f2dfb3ab15e3acf8d937d50.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000022", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [558, 203, 670, 306].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/4d1429f8a559ffa76ff6dd6181113d642a7d6ccac748309b027172a63950138b.jpg", "reference_inputs/images/2dc7235177eab99e00758787c60e4230c1c1263feedc79745fdcd2eae1d0ed46.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000023", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [484, 423, 497, 436].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/fd1674d77c460465cf29ec46b38e30477c75afa39928d0787c1612eadc6c74a8.jpg", "reference_inputs/images/165cc78f38c0f249eb17768c4ed0e3d7f315364b213c61a5d41731ef443b90ce.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000024", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000024", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [464, 173, 702, 370].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/1f14b718b819a632f51cf4ea6d5486cc08d1760dcb059aa48113fb4a35adc9bd.jpg", "reference_inputs/images/155fc91be7a97a5fdc159453201df0ee7518aad3bfa301d6c7114bad16e55843.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000025", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [574, 182, 723, 312].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/861fd07722f02e7775960ebcbb39ddd697e58e8a2aa057c62f68cc8e41bceecf.jpg", "reference_inputs/images/7fe81ffe48ac7047ad4182035e62772d238bb9f1dfef4032e169297c0ac7b451.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000026", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [431, 239, 451, 308].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/b16a03557227e8f95de9c50c426c05f1c1350b29fff4a4bc8e35879fa14681f1.jpg", "reference_inputs/images/32f5b5d0605f2bbeca2a3a4b5930b7b534552a65ced65cb6367ae6531b6d7770.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000027", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [237, 274, 246, 290].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/6b7b77f4c99dfd37df994d2154d5b52d6924e2b46395660b62e6dc3ce09556c7.jpg", "reference_inputs/images/0d940f51c7d196226a92b50c9e8d77e7521d744213237d2caff682872d867ddc.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000029", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [722, 486, 758, 508].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/1a5875388e9777c9da6a86358adce81ccf7c461253ed1c37b5bfb0d1e954f77e.jpg", "reference_inputs/images/4e30652027712adc0c0e87a4e16e0b9c27e3b8e6ef2558e50ef70cecd05a5453.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000032", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [580, 534, 588, 558].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/e0381b03e29afa40e6d18828c0f36bc0d89bb5a74d329b1a9516c1a79e4b51a4.jpg", "reference_inputs/images/0712923cb9ef7bc88dbc73ca5969dd9ffecc54a415e0f4d2e3a28e04b7f8a7a7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000033", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000033", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [546, 277, 552, 298].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/78b710e10880884ed52f424e0282646d1ef457426d2c7796aa7d1474956fae5b.jpg", "reference_inputs/images/faf59fea4cf18c3dcd8db23e2cd5efe84410f212d9a2d99bcc7eaeaa65a8e2be.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000034", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000034", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [355, 275, 426, 318].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/5a92023fc6c9aea8e39fc3d35221fd1e4628638df8ebee477cde0cef714a3e78.jpg", "reference_inputs/images/bacd0526867e3021b3476ecac5fd7a0986826a48bc56cf026195ba716071923d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000035", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000035", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [368, 460, 376, 478].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/4c4ce9a2a452eb6f170a53a1fc9424444112c57777f71b6b9064acc133562968.jpg", "reference_inputs/images/214151f8fd0558933c35e9f922e57e66b1faaf84c138a73d2da854872e21955d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000036", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000036", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [114, 254, 132, 300].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/fc6f32bd6db38d02e41ac4f11d99c28059ead881cada91e34ed65f18dd1e8cf1.jpg", "reference_inputs/images/cc654abf3625cebb4fb1ef067790e12e02268773b72d9deb070fed6671b3d91a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000037", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000037", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [477, 268, 601, 327].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/d805d4a2c5ab72663e1777357194e745b5bdfd434697b3d4f6a88ba3b05e17b3.jpg", "reference_inputs/images/5409c8630a60faefc71d747f11c0f0ab2d0019e1d05d89cad3dfb0b1ba66be1c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000038", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000038", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [462, 182, 583, 341].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/c5b7c80344326eadf236c095b74caab02fec7f079c683b63408b21f0a4fb239d.jpg", "reference_inputs/images/619f3705bc8caa217fc274311cfbdfd18340a5ff504a9675a8f779bd4fe7f099.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000039", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000039", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [753, 258, 765, 298].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/9a6e5f9f30ee237528c1c8bea12e3e2df67813b2f7eaf7df22f916e035be8caa.jpg", "reference_inputs/images/37193519644ede3cffe2c34e62b72d4eca36138768a5b768ea8675d7056bb134.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000040", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000040", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [488, 254, 522, 294].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/15efbf7d748243da6419b1350bea59113e625f891f627c70f979a4ee745a68a8.jpg", "reference_inputs/images/ee076bd854493700b9fab044dd65fda55b8ff5f246f0fb232ded9339d9b3fc6c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000041", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000041", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [537, 506, 580, 574].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/9e3c55173c9e66d07f914d2a2284cea3ba284452cb7b3dc166d8cad737c5426c.jpg", "reference_inputs/images/42d29e7610dbd9d7ed53070484ca26768cc0f73c1d9e32dcd9365b6a0432c197.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000042", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000042", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [690, 262, 705, 294].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/9ee52d4b590978577fe3a21b750b8093f46ed3b4efce315bf8098e1feb0e6325.jpg", "reference_inputs/images/80411caf2c694c923da91acf9ce65757d6d0b3cc4e9f32a494dc213f2e0988f9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000045", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000045", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [493, 250, 645, 351].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/4186c556f100d93e53855131b95c8bead2b9379f008b39b0b666c3427e84b2f4.jpg", "reference_inputs/images/155fc91be7a97a5fdc159453201df0ee7518aad3bfa301d6c7114bad16e55843.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000046", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000046", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [512, 421, 526, 438].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/35717c2dd7162b69eeec0d1cc1f01170ab4debdc09307139ab4bc23ae14cb0cf.jpg", "reference_inputs/images/21f2c320c1e570119f3c51a4276d427d1289f16ef62a1d359f7a1f4beaf17ca3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000047", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000047", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [562, 452, 579, 467].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/ebfa3e4e0b1e369bd197de8803340f89b0e019810fdd19183074acb8416c2de1.jpg", "reference_inputs/images/c493e53abb3faed8b4199f002c0950d1071d71ae57f73613920948372a33197c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000049", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000049", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [430, 452, 466, 510].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/ac4775a98cec7476dfac01427aa354d079ad37508abae07202865e79af92f1ae.jpg", "reference_inputs/images/4bef85cf9114015576227a64dbbbb0fb2abfbf1cc0ae73d509da424e5b2b3f87.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000050", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000050", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [505, 515, 531, 542].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/8fd402abf4edbff749a00e1626fa8df9dd2e0f2e68c3b40d39f362cc7f7a842a.jpg", "reference_inputs/images/237069dce3d0317d10e096791b5d633b982014b0bd077bcd2d5a32473f3250ae.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000051", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000051", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [907, 238, 942, 309].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/66691a578a4928bbc428083db134a66fa0f01535601cf27c059d0ec3adb0b665.jpg", "reference_inputs/images/ee568e58e8260c2a49a21cd116e1e6065a266da834f0bc842f073d0ee0e5da43.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000053", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000053", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [330, 486, 340, 507].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/09d6bbd65d5060eae93487145b4ecd97a46eec4da32909caff0ec56456d840de.jpg", "reference_inputs/images/b44ca937189d4b9529c968b0f3bea457bb9b34d3c8f1a0061bfd35bda0a95e8c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000054", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000054", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [655, 266, 662, 292].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/52169b7a0b342f2d63699bcc716256cbbe071a240289254b98442a1483faace4.jpg", "reference_inputs/images/369d8c5a94fad57873a397e17382502d73f237070752a22f9e7d59767de6f1e1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000056", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000056", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [26, 237, 270, 331].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/c1c4c16c5c4f7375ec69c2f13d00eb7c2ce0de9fe9e8b6fabc682275254caa50.jpg", "reference_inputs/images/334b4e4d96aa2cc5e2ef412bf4ca996a78d03fc955f3ccf5b8987a427f0f5fb7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000057", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000057", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [963, 222, 1003, 319].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/092ab14103d7daee4393542be420b50621530d158741b01471fd752dac1a81a2.jpg", "reference_inputs/images/80411caf2c694c923da91acf9ce65757d6d0b3cc4e9f32a494dc213f2e0988f9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000058", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000058", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [331, 282, 378, 298].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/67ae8f2087945e69f4bbf4dbed43553acfd163ecc892980fd3d02960b23ccd4f.jpg", "reference_inputs/images/ff6cbea361193bea986b927cfcb03bef6765e8a580f5e75078971de887869eaf.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000059", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000059", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [733, 257, 748, 298].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/f6c484dcb9b6020582b2c92fbdb472b4796aeaa6ee3b54cbc4c7fb823701544f.jpg", "reference_inputs/images/977acfa7137e557053b937cdac0f694f0400cda381f38b0fcb1b44864ca8defa.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000060", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000060", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [604, 280, 626, 295].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/98e184ef25113e99d3acb81d2b999b2dc5d3a8d18056ae3ea6167479bc81beba.jpg", "reference_inputs/images/1c3d4109a38cb72dd2d349b03eb89d2366a1cfb46285ae129959aae657de9447.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000063", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000063", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [945, 493, 988, 571].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/1c089b7bf6df9995574907af48586f64b88c52634432467756c6559107cfb606.jpg", "reference_inputs/images/0b04d2c048d9ec3220836158a321c7d88654a550351b69a7abad236dadc214fa.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000064", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000064", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [506, 474, 526, 492].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/a28580c0acd048e5e50dca36b438dbaa11ae1aafe0cfb9ab9367bcfe807cc7bf.jpg", "reference_inputs/images/affe958434ac9600d59e4e439ed10d5a41e9ed75ffaf1bb9084b9fe73c2ace8c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000065", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000065", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [963, 262, 978, 295].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/29fc872b749e944ac734776f6892fe8e4600cf609cbd0a71f1ff2dda5f3dc4e4.jpg", "reference_inputs/images/eea0ba64382eff8b05e91d81cc247d4594f546377d9a5972b7845affdcbb97dd.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000067", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000067", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [530, 274, 613, 312].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/88604a8a195412bd016ea404f6c5a296199e3ec9b21d70fb637c712eab67e52e.jpg", "reference_inputs/images/0ac9f845326119cc41a5d02d1566a611233a2f94f696b99198332a27e2440ec9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000068", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000068", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [755, 26, 781, 54].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/725ced7b4e8267a06b6a001f736a03604d3db982e8bd0e8a825da0fb4cf6c4d4.jpg", "reference_inputs/images/bc63985550a18559e0aeab6a28a3017dda43dd184d3fae61673d03527e58bbcc.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000069", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000069", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [334, 522, 350, 543].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/9e3c55173c9e66d07f914d2a2284cea3ba284452cb7b3dc166d8cad737c5426c.jpg", "reference_inputs/images/ce6b08db9a41daef36af00206c758b526312c6cf41bef6347f4d72c299e9eb2b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000070", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000070", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [394, 271, 456, 309].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/6e73dd999c827994b8f9751f6509c2a233f365e7d66ae156e647207f6714bc11.jpg", "reference_inputs/images/61e5799ec4365ae66e47f25111b7621c86ca0ffc24894e6fee6505c7db389c9c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000074", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000074", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [710, 496, 758, 515].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/9c7047075779686d932e90770dddff6b67c27ed2467b2c66c6e22ded141760a6.jpg", "reference_inputs/images/7d0ada91494a357afe62a3dcacc31e5200312464c1c5d46f241f69acbb88744a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000075", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000075", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [623, 264, 633, 294].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/54274a9910117326d71f3e76d951de9920b2bf270df85509c72aa19302d45c54.jpg", "reference_inputs/images/27579f644325a8ad7d9b4eb3774ab00c1617860d245edf1122ec84095e57b823.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000077", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000077", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [398, 87, 454, 201].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/9b07624335b4fa633b68a1113aa88dc561c5ac1dcdfee58fdfa373ca97d99437.jpg", "reference_inputs/images/67d3c4dfaf5334baa46fdb18857f233075e29e32dcbad3903bb91dcb5950e373.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000078", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000078", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [38, 246, 64, 304].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/5900d4d9d0b61f173161a1afe333eb888a59d2158498195106ef162bf8526830.jpg", "reference_inputs/images/a2c94439294592ac645c2eebdc4815dacd145ee16cf456363ed94e9886ac961c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000079", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000079", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [672, 261, 681, 295].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/76fc88162067a28cc7dd4115331e07bee33e6792d008798bfb64fff5a9690e4b.jpg", "reference_inputs/images/295ad840ca29e6224f5cb5df428f61e9d838e7b3c8dff1eaa17dea5a4c1d89be.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000083", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000083", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [926, 244, 947, 304].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/3a78694109d6f4e2aa6c3201e224184d3c1c7c4736b8adac3b14a0233f1f3ea6.jpg", "reference_inputs/images/077e994a9c05afe1d801c4fa13a982412196817966c2f8922b7c12be72d925b1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000084", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000084", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [490, 227, 698, 366].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/41e3b47493e7cef3d7bdf3f8545edf826aef809002ba8b49baaeb5b7323e94ac.jpg", "reference_inputs/images/8d8c75d9f111c8e3347293dbc4f5899a1254a3711441e4ff69bbe385932d7afc.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000085", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000085", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [413, 10, 460, 116].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/f2f80da1a6b22f4d549f719e57b16815e0a24c91e341527ac187a7ca38be6e9f.jpg", "reference_inputs/images/feda76232f2405ca0f902f5e80fccb76864fe71c636e873aaad62eb1c082cd77.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000086", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000086", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [926, 482, 958, 502].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/477961a7c2fcfa75245d3681f251e6db5b6e409ba433b4c7c755c533979915bb.jpg", "reference_inputs/images/d6c7b56b517be5f9269a8f041fc150e7eb5f171fcdeee107c90b0307faf41b76.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000087", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000087", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [355, 494, 365, 514].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/7abe8bf0e8316b5c50cf7a1af93b7def29fa22d68678887ccb5eace7bbf814b0.jpg", "reference_inputs/images/1ac8ae79665ede165e788f39ca756a98d62b27c8dc1a0d5c67b311d5992b3466.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000088", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000088", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [722, 275, 823, 314].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/8d37f800ffbf71905c3386e8f60ca28c2dadd52fc67793a1c4e7b790f4df4fcb.jpg", "reference_inputs/images/efecc6ea655d2baf9c731021df4e0e26e6ddae097bd52949d1ab78c194d2649f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000089", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000089", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [85, 262, 277, 326].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/0babe2b74ff3bef3cdac98263fe4524186cb1fd3c2bd927891a0e7a8de627389.jpg", "reference_inputs/images/3e21d07491e09cd36408a4cbcdacf1e77d7462660023e5288805dc80014a0207.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000090", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000090", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [712, 524, 773, 549].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/7585f41fe0b84cff321aa7f677c0cca23c67433ce4c183eb9baafe39b873f4d3.jpg", "reference_inputs/images/6b5ca664a0bec76fcb4e83634ea93c39a1cd48cb0e2e19cac5000ed55fa74d96.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000091", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000091", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [532, 243, 554, 309].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/a71e45c547e3afc57b1c6f30cf4d9fc977114d99cb3e9ae0d66203b12df4ad69.jpg", "reference_inputs/images/f6f61ba102d1b190255546c7c2a3a4ec5ac6658a821aaae3580ee11e9c1267e5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000092", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000092", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [236, 262, 255, 297].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/a37278d6f61a09212963b84767982a4ff7065abe6dae528d356c579d594d5b87.jpg", "reference_inputs/images/0afbd9401679e8a5f9f739e851e2d6565ca0847966b9c870ff25e6bd5f0227c3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000093", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000093", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [829, 243, 846, 301].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/64ecb14e96931fa12d63d7d52e97e6551d9ab8794bec873526ad42f1cf64fd39.jpg", "reference_inputs/images/35e859c28bf029e3e97f5365979050dd06717e2c31f0c6b35566c29e99c74a4e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000094", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000094", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [814, 246, 846, 302].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/525985cacf9e6d8ff8c7962a50ef0a037d1804093274c55cd33d83c121e048a7.jpg", "reference_inputs/images/ee076bd854493700b9fab044dd65fda55b8ff5f246f0fb232ded9339d9b3fc6c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000095", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000095", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [354, 275, 402, 300].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/0eb73e9b59314fe88798044774dc7b2c640adb41215b0ea735d4ac8fa5c6b73d.jpg", "reference_inputs/images/bf8b4830a17135711744d612169e36a7ce3c7f1e60ba36a3cf23a84c26612ee6.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000097", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000097", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [404, 272, 459, 306].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/06ecb69a21b3792b108f8f5945a2d8ba57f1cf4758d640de4aaa28301ab4558b.jpg", "reference_inputs/images/bfbff7adbebc42ae595d39371347b63219639483ac4c74142b2bc61623dabbd5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000098", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000098", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [641, 230, 817, 328].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/fe8e110d6bb38794d1bebd55d8ac04077ab3b8a2e90fe6f56803526d46685019.jpg", "reference_inputs/images/8d8c75d9f111c8e3347293dbc4f5899a1254a3711441e4ff69bbe385932d7afc.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000099", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000099", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [288, 260, 374, 306].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/81046b24d8b0d3e578bdfbf2026c08137a814732d76eb9b298a0b6178fe200d5.jpg", "reference_inputs/images/9f402f3e87b729a243a89e1c0c80ee300787da4c6b9d8811cd50f3d5b0071d9f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000100", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000100", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [500, 268, 538, 298].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/9ad8a50cccc69d4d19587748ca75a29e617773e0bff87ecdaec8ddeb22d583f1.jpg", "reference_inputs/images/2dc7235177eab99e00758787c60e4230c1c1263feedc79745fdcd2eae1d0ed46.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000101", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000101", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [508, 497, 528, 521].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/5409c8630a60faefc71d747f11c0f0ab2d0019e1d05d89cad3dfb0b1ba66be1c.jpg", "reference_inputs/images/ef33bb0f5b8406fa4590a64cf002e11bf44d3fc2863193c752ca295ac6bb401d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000102", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000102", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [507, 539, 532, 572].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/205e468a9b3ce22bd6e26f5375ed3a57a3422b806adaba4c88961f27d3d22be2.jpg", "reference_inputs/images/efbffbf5cc31f2b35479861871fbf8dc5c049c8a9a6f5acc5915be6086160d5f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000103", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000103", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [862, 238, 897, 306].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/948eb1874b239d46c2fb2d23b88f7be52db81266aea190904fab7655e01e84d4.jpg", "reference_inputs/images/01a33378cb0d1c10ec0222f6b5d05e9e374d1a4bae5fd14ef92f9858de4d88ff.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000105", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000105", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [764, 262, 776, 300].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/8d417a17c18faf40fc765c8b23d0f70cca61632b0e3f1a324e97c9d544bb2988.jpg", "reference_inputs/images/6866c21286157ed1a98a7a54f7f47485a2c683f40fea7eb1b587906f1433733e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000108", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000108", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [403, 272, 471, 306].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/16a8e2767ddc2533679e06b8f287131d7eb896b588ca165bca382bb4df8e8d3c.jpg", "reference_inputs/images/07ce94b08f3eb198a9dff4281ea74631308b0765bb341ac4c720a3b229986266.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000112", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000112", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [881, 269, 888, 291].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/43b0a3fe42f2acc793e99dea4dd8389f524b8f3d47f09236481fb1cc1befeb13.jpg", "reference_inputs/images/a60a62748a3f2e935195f4fb999c581d2c0846a7a84952173733ba07423749b3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000114", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000114", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [26, 263, 38, 294].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/1f4e5bfc1dc07dd2fc6ab971db43666d4bee8f303a3194ecb836bf23599c8aa9.jpg", "reference_inputs/images/c39ffc71392900cb722b2de9e5fc7e0bfa3f44928702031c7009afbfa7b4d8e9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000117", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000117", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [575, 240, 595, 309].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/a5d0e7cbf9767cb7cc710572ec696712ed5a45fc51ec05aba895560f75129c58.jpg", "reference_inputs/images/903023c225c149cedd651170969d99e1abf3ce6bbe19bfa15fba8caa56e5ca42.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000119", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000119", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [423, 481, 449, 512].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/6d76248e7087fb5fcf6decf03cdcb4e685fdbb05d11f13983b35973a1a64e154.jpg", "reference_inputs/images/74f386330c7d9af7076f742a7d66db3df65f2fcc5027d9e6f0c2bb2d6c20ab41.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000120", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000120", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [498, 186, 608, 337].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/502fe2b7bf5d1fe5c44e96178c57cdc8f6cda7fad2cf013c6853ede35a442838.jpg", "reference_inputs/images/cc654abf3625cebb4fb1ef067790e12e02268773b72d9deb070fed6671b3d91a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000122", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000122", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [493, 261, 519, 291].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/ed2189f53b3f6d1c05bceb524c0005a7508a1ff0431f3713d530ae623c5b2642.jpg", "reference_inputs/images/8ff5853e0bd8ef95d064df272a19356897f599fdd94098d87eba1a836939d593.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000123", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000123", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [338, 266, 346, 293].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/dbdbafd1873f8374fc2da3b5dc8e39767781e722be8ed25fdcf2da41eac4a224.jpg", "reference_inputs/images/1c4a984fa871acdb7a4fcedb40fb895e0a1f4c3d65635ac5eb19c51fb3d8efe8.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000124", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000124", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [316, 294, 374, 403].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/1045dc14fabe226b9f5e057b367632be473563d2f1157262f947e0220d68634a.jpg", "reference_inputs/images/05653c6340d7df55442d03ed3d8119b93e8374d103577b227acac495e750863c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000125", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000125", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [819, 279, 886, 297].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/57e4694e7f958f730076926ae88c4407181e8cbbd683ad2fdc003377784f44e2.jpg", "reference_inputs/images/e0381b03e29afa40e6d18828c0f36bc0d89bb5a74d329b1a9516c1a79e4b51a4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000126", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000126", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [876, 234, 937, 311].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/6a7f9e051b900b178f7d7c6eefc92cdd3bb4cf1baf907fdc77313333c8ba0243.jpg", "reference_inputs/images/0afbd9401679e8a5f9f739e851e2d6565ca0847966b9c870ff25e6bd5f0227c3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000129", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000129", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [19, 236, 54, 305].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/75bb9f3ac1921d7fed1cce6bc9bbfa9956c3988c4fcf366d0c757f6c48b81399.jpg", "reference_inputs/images/302afa3fe401318929f9af4cedaf36e9720f28d6ca384fec6c40c86710b61e35.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000130", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000130", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [779, 66, 830, 86].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/32ec1cfd3248914be2bb9127bbec2417bc640c23321fbce528d484cc3448edf0.jpg", "reference_inputs/images/bc9ecf1a01f58e317cba1408870ca8cd9224a3287a4cfcfb65985c9815fc847b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000131", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000131", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [706, 262, 738, 295].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/fc2f455a3a134b67bdb86f8967c4e888682308e77817c4a1c0ede8e7b7ffb7cb.jpg", "reference_inputs/images/38ccc557c2c7f36d3afb0b4ec64f79eadd0fd34a1ffbaab0a18d60eaf685afb4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000132", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000132", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [638, 536, 652, 559].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/205e468a9b3ce22bd6e26f5375ed3a57a3422b806adaba4c88961f27d3d22be2.jpg", "reference_inputs/images/f807901de3dee9c2d7c7f64e832960a87933faac7c63f54bc8101d31810c5d2d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000137", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000137", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [65, 173, 115, 339].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/47f9e862659644a60fa5865a905e3616ae64a6438419abb7aae6d257478f5f7c.jpg", "reference_inputs/images/eb1a89bc11789e99b19e77585fd46e9e36fae0d3273a75e9ebfa9d805ddb8203.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000138", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000138", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [653, 478, 659, 496].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/38ccc557c2c7f36d3afb0b4ec64f79eadd0fd34a1ffbaab0a18d60eaf685afb4.jpg", "reference_inputs/images/3dbe00d1c39eb299cc2669b61de63e39aa84e832b904cccc4990c3b6a995c0ef.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000139", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000139", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [470, 490, 490, 514].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/bfbff7adbebc42ae595d39371347b63219639483ac4c74142b2bc61623dabbd5.jpg", "reference_inputs/images/d58c9ae08eef29500a5bf2870e2b136cf317de0992e07d5be2a8fc5ab35b65a2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000140", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000140", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [949, 241, 1002, 304].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/a13aec45195756770182b27499d472f47b160ff507dee6ba042827d5c22fdc1c.jpg", "reference_inputs/images/9a9a407934174b36a0f55a6072776286defe626b8cd00878ae1f471077a5f307.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000142", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000142", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [754, 254, 767, 299].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/135d5add10ec73a8f8dccd1efdf99084b9cf438829994e55973c598984e0abbc.jpg", "reference_inputs/images/d0c7a1e600af97e4214a01d83668ccb9038b75752431c9384f9b153978205fed.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000143", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000143", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [520, 280, 551, 295].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/140c37b0862350acdc31ccd5823d77a3406f34d820549a19c0ff03742502e98b.jpg", "reference_inputs/images/79c7088a1f08249b13cfb452a3aa80cf0acb81937c2f6e8724e40fe57ddd7298.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000144", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000144", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [111, 253, 138, 302].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/326e309493b74656ab397a5c7fe4a7d782f7c238e5d6a88d3395323169ed9146.jpg", "reference_inputs/images/65bc64984b2063a131f48fe526cf6a7b07da1d78c38097687558bd599f26a609.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000145", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000145", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [541, 479, 575, 534].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/a00ea25c32d5a5a8e9818c7c7d02f9e95b2ee0792e92f82461c929232b25ea4e.jpg", "reference_inputs/images/b800d7a40eea13b4f7e2a8f5023f5f5668249c4876513c6be9737791b96918ee.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000146", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000146", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [14, 266, 25, 293].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/1c045cc9d56ac5597dd52c6857718f20cd7716c7605e6a518351ebcde2bc3b2d.jpg", "reference_inputs/images/6a5e13694f5f116e2e93f37ae27001d2424ab7e44cf25ed4c982edd446cbccd4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000147", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000147", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [460, 282, 472, 291].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/5656546a0d9e30f3637c2bbbabe2b6fcd0e76f79f385253514533b4f71a004c0.jpg", "reference_inputs/images/8771c85ad7e20bee39b6eb3be777a3478801bbd3fe47bf6dc3ea6bcf510f6924.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000148", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000148", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [274, 275, 367, 307].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/e8f86a144ff07fab1b6ed4595caeffbf95d4782f342c62d61c804499bb0065bf.jpg", "reference_inputs/images/1c3d4109a38cb72dd2d349b03eb89d2366a1cfb46285ae129959aae657de9447.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000149", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000149", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [459, 422, 475, 434].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/51980d29e747112f647b997b5ef6601453caad7f89bcd907c86d16752df0158b.jpg", "reference_inputs/images/b683797c51a05883b7e76e4ff54291fc1eb45bd00558f4d84c85fd38534d3947.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000150", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000150", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [318, 270, 398, 310].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/b4d21a7e3f68d56255291f70c6384a51e8da3d9f3012b231848b23a57471c549.jpg", "reference_inputs/images/760df4126276baeb1783940be483f8faf7c7cad7e4581beac622ddf0980c251b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000151", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000151", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [195, 548, 205, 574].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/a4cf592a72e7c039e89f2ce68a4f098467de07ca020316c2c156372cb9a1cfc5.jpg", "reference_inputs/images/38b794c38af6958c62c3bf8c175f24dd96ded835eb611aef8431b16ece19e7c4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000152", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000152", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [304, 387, 358, 500].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/1dd5e7838e2c03695637a18be71970132c07aa954e052efb5afe736e1b8ede06.jpg", "reference_inputs/images/44d016d6bc7786c0ccb7737f4ed263b47be79d3f688d9d39c0375e18dd37137d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000153", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000153", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [921, 274, 937, 284].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/45e3c91fbcd4ccfd9412b8b98482ff8cc51d1559ca7f3948f165faadf6dd1976.jpg", "reference_inputs/images/ff6cbea361193bea986b927cfcb03bef6765e8a580f5e75078971de887869eaf.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000154", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000154", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [220, 256, 226, 297].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/4183a0b6de334d0e851eef8ba64a5005f51b443781882f4bdd69e0315019eef6.jpg", "reference_inputs/images/295ad840ca29e6224f5cb5df428f61e9d838e7b3c8dff1eaa17dea5a4c1d89be.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000155", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000155", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [429, 490, 454, 512].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/f74c703367de7b2fc1295542d39d0c937f0335d7cc3c1046ad03f832839e7815.jpg", "reference_inputs/images/b84ea74379a0dc35ed4efa117f79ffd3eeacf5644d4810946141066cea215ba7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000157", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000157", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [784, 501, 813, 533].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/5de87c83d655713d33a81d292bf084f128ce53e77ca6e352e4e381bbced8fb64.jpg", "reference_inputs/images/fbdc39cc7686c6098f4439103a97584a6a002a045d6e03bb084cdca90a855d25.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000158", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000158", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [472, 245, 594, 338].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/360617ebd1ddf8668083527a4aeebbdda1341224f852b2297b1aa4e0edd6dbf3.jpg", "reference_inputs/images/ce8efefe6296b719d57192eb8ddd3beade6a537c2ce5be71960d4a9c54866577.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000159", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000159", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [798, 532, 805, 555].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/477961a7c2fcfa75245d3681f251e6db5b6e409ba433b4c7c755c533979915bb.jpg", "reference_inputs/images/a4cb032f62d08649d215a14f7f4615f0350ca5a9e04fde3f92f65683205efc94.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000160", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000160", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [756, 246, 790, 306].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/45da49fdb274749d63e970bd37daf93437d2d5567cbf72ecf6ffa776e5544b3f.jpg", "reference_inputs/images/55945b428279ead34d400d4671a3514cbc80f1a9f5fd26f6f94de1b2a28bba2d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000161", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000161", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [727, 254, 738, 296].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/6735f98cc46d273f973863ea4bee7b2468bdd114361efff6508e8164bd3036ee.jpg", "reference_inputs/images/27579f644325a8ad7d9b4eb3774ab00c1617860d245edf1122ec84095e57b823.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000163", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000163", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [954, 476, 965, 491].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/65bc64984b2063a131f48fe526cf6a7b07da1d78c38097687558bd599f26a609.jpg", "reference_inputs/images/876792b574311b760babd853433624869dbf4494f45b87e39af7b6aa66402447.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000164", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000164", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [509, 402, 520, 412].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/29a6541a1c03ccfac6d073b89565f4d67321d2e2acdf67f5fe0a474b8c507135.jpg", "reference_inputs/images/314d01bbfb4d254b346e7f6bc4fb62da813832c57ce8b979e13a5d0aae67e2bd.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000165", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000165", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [654, 268, 660, 292].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/d6ed40f9a99760bbce21fd56ff793cb50ab6368c56c4fb93d57ab5a08b3cb049.jpg", "reference_inputs/images/e0f97d1dbbb520d258aa80b8177f8e8a06c5310d6dbb6a555b54fa0664011e66.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000168", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000168", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [955, 239, 994, 309].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/0a637c542cd55efcd7db57209cb296cd5b3b8122a25a1081e21a75d2467f1eec.jpg", "reference_inputs/images/9a9a407934174b36a0f55a6072776286defe626b8cd00878ae1f471077a5f307.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000169", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000169", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [342, 262, 424, 314].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/bf6e1f1858c88fb977f0dbc32936d9c13cdd676f51a1e07b64daf08ce71bb42b.jpg", "reference_inputs/images/c18ed58b26d234fb469fd862a56269881f2ef4151f5fe299c88cbab19df890eb.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000170", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000170", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [728, 254, 764, 299].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/daa4685c900c7485f0b3274935690a92a3a0e3d43a08f6d210f5627453dcfb6d.jpg", "reference_inputs/images/55945b428279ead34d400d4671a3514cbc80f1a9f5fd26f6f94de1b2a28bba2d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000172", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000172", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [793, 206, 894, 334].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/c5b600c09a6d37cd46165ccf13dda6e2f27b7d463018639951817900fbdef907.jpg", "reference_inputs/images/4dabf19c91624ce95cd33715a4547c90755ab3181ee934441d405dcb02c76b3b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000176", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000176", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [506, 502, 529, 531].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/07ce94b08f3eb198a9dff4281ea74631308b0765bb341ac4c720a3b229986266.jpg", "reference_inputs/images/8156fbafe831b86bac65b3a21863b8beeddbe10ef22d7a56993d4274368b123e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000178", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000178", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [791, 109, 828, 126].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/9244160e155b63bdbce63c0444e48588782c5415701fca6dadd7b04750cdf985.jpg", "reference_inputs/images/4bfe30390fd98ddc3c05448d9d9ec8a97771e1404cd960a70bdff1890009516f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000179", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000179", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [694, 262, 700, 294].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/6c5bcc96e9ec88143a2af460a2c36c62a89566b5824bb951b130bc5915679c3b.jpg", "reference_inputs/images/d0c7a1e600af97e4214a01d83668ccb9038b75752431c9384f9b153978205fed.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000181", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000181", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [730, 262, 744, 294].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/9e191a9c833a6d0191d5f681a95a3241003ac2ef139ea07eb111a9690c44cf99.jpg", "reference_inputs/images/c1bd0271eb6687fc02d0d636a157471d657f873c3affd73cdb9593443aab7b57.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000183", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000183", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [2, 455, 44, 503].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/0d940f51c7d196226a92b50c9e8d77e7521d744213237d2caff682872d867ddc.jpg", "reference_inputs/images/aa247e83cbfa6a55e87f479c5d2e650e3f2266f2b3bb012623eb5ab140aa8602.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000184", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000184", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [334, 476, 348, 493].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/dfa601d61da56df004960eee7565aa9b513ce9efabd8a09e9b983a0845dcda7b.jpg", "reference_inputs/images/dd84ec2bafb16bb213460e7a124814e17d0f7d9c46fed3791c34b8ea12eba9b2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000186", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000186", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [509, 490, 529, 514].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/db9fc74c10b486ab63a0d92483697632bbdab6f70838191d395d4b86b21e4f3c.jpg", "reference_inputs/images/44166659bf064b7b7f17ebe1cd27d0ffcc7acc728691ab96f5c75bf2df465995.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000187", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000187", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [264, 275, 405, 308].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/bb1234fc75db8b2004f6e5dad03553677922756abe73a2bc76ac3db2608e9339.jpg", "reference_inputs/images/ad7e06d19be14c7572094fcd2a8656811827fd7f8c582feb9f85922ac8d4a358.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000189", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000189", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [17, 242, 42, 302].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/f67eec63d5c0c61458bbae0766c70b18ecdbfa0cb0affc46531d9447a2b74645.jpg", "reference_inputs/images/8ff5853e0bd8ef95d064df272a19356897f599fdd94098d87eba1a836939d593.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000190", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000190", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [545, 244, 590, 296].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/f715261c31086ae963af6c95167737fcc81cedadfdf0f8460aaa5262d2a8969f.jpg", "reference_inputs/images/5fcef0ed1e787fc5faa2ef566001887a3b83ebedf6e08c51a36e92a864bcc37c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000191", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000191", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [594, 274, 678, 309].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/199b8fbe4a152e9d5ee803958dac2f67a88f2cf0659fdff516308206c27adca2.jpg", "reference_inputs/images/35e859c28bf029e3e97f5365979050dd06717e2c31f0c6b35566c29e99c74a4e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000193", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000193", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [498, 277, 551, 303].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/46bb3bd981a5ae2e0219fc7f3f31f830d5b4bef526551466cbe8e18d63df974c.jpg", "reference_inputs/images/72ddf1a3292fe77737b11c17910cf5f5f03a6950dc2d9a10a5ece571bfaacb98.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000194", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000194", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [307, 547, 313, 574].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/977acfa7137e557053b937cdac0f694f0400cda381f38b0fcb1b44864ca8defa.jpg", "reference_inputs/images/bcdc1e69c0a36fdd6f81b6dc138bac8948640e1f401b454a6715a6941bd5a4a5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000195", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000195", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [218, 258, 229, 296].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/c0d2a3e08cf989cab00b16bae02eabdb9d196be80eb7ba851f154ec5d4a68f11.jpg", "reference_inputs/images/dfa601d61da56df004960eee7565aa9b513ce9efabd8a09e9b983a0845dcda7b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000196", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000196", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [853, 278, 930, 297].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/aafc73a7aa22bec3f1d4f00952a95fe97681a40c8c7274f4dc3e6d3f563faebf.jpg", "reference_inputs/images/38bce0541269914b1c0ed8a05ee895cfe10108d7ac23d6e020342e42a1c72cb4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000198", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000198", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [503, 456, 520, 476].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/35717c2dd7162b69eeec0d1cc1f01170ab4debdc09307139ab4bc23ae14cb0cf.jpg", "reference_inputs/images/ad26a1f82d612f9596c6e104842d9b22eb76e19fdd9cd98173f1cffb07c5df33.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000199", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000199", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [658, 264, 673, 294].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/c45f03204bfb05c79b1c7600d3582de3ac706beb732d0c7fa825c450a31e4ec8.jpg", "reference_inputs/images/bacd0526867e3021b3476ecac5fd7a0986826a48bc56cf026195ba716071923d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000201", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000201", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [414, 270, 420, 295].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/8f6e1d91122c91578b18dffedaa92d09784d3e28552882ecd25afe2c34326083.jpg", "reference_inputs/images/cfd3ae4956d0d912e42e8b069a600698afb87d5f906b5079bc3f973feb58a839.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000202", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000202", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [626, 504, 631, 526].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/8fd402abf4edbff749a00e1626fa8df9dd2e0f2e68c3b40d39f362cc7f7a842a.jpg", "reference_inputs/images/33205fcc688c463d26db6e85757732d996e12afdebcb5e4ff4d86f47f3b63602.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000204", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000204", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [810, 273, 817, 292].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/d9f7b70cbce2f492858416fca652a2814092d4f954affedf5bcbcae877e1df12.jpg", "reference_inputs/images/698a1f1fea025f2ef0067d251e43da911ee6143b00cb60b7e052abd12237e7ba.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000205", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000205", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [469, 370, 484, 391].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/6d76248e7087fb5fcf6decf03cdcb4e685fdbb05d11f13983b35973a1a64e154.jpg", "reference_inputs/images/9f8bd7afecd76acd880fa4f00ef12b38a4ea1e2d76f817d9fdbce48abab400e1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000207", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000207", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [156, 271, 166, 292].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/79eb13a479257c25e7f5b6897a0354841cbe1086dd11e973156ea1192b073fc2.jpg", "reference_inputs/images/beed69582179f7e4769d3cc45dd0fdff7a8efafbfa7a4eb2f9e0c8f4f69a69ca.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000208", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000208", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [499, 388, 517, 412].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/a4cf592a72e7c039e89f2ce68a4f098467de07ca020316c2c156372cb9a1cfc5.jpg", "reference_inputs/images/9762b261fdab1ae82f5704a59390f7a7a5f789c680e4cf37c3c11ba8d2d2fc0f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000209", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000209", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [718, 262, 736, 294].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/0183cfd2866b057e06464418777bcf14316141f2c7c9027b08baf55273d48d83.jpg", "reference_inputs/images/7a039d2d017126298f41fed2e5455af7f8cbb6030470c65d9b9c7e08bc14ab7f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000211", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000211", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [674, 493, 704, 511].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/142692d88c20575cf72bdb9f1042a2514c754d2a190fb17636a4b131aa1eaf10.jpg", "reference_inputs/images/7f94190665dd70f7d3a511d556daa5af6ea4f449ea5a64eaf2788cf6931abf62.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000213", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000213", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [750, 409, 776, 426].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/163d53839b6d2b76ccde8abff6da2726cb12953bdfcc951e7e20911117043a7f.jpg", "reference_inputs/images/cc8944aa5d435ca2466183a66169534e98db87d523d527b86a53f9a05a7ad589.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_20m/000214", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000214", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [364, 278, 417, 299].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/389a42f43ad667172d6855c4182d3f7f74d9b73cea54b3b04b85012cce008507.jpg", "reference_inputs/images/54a45b98db6cf9dcae184341b11c3ada9efb63b5c8c0222a21ea269ead7746f1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000002", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [46, 243, 74, 305].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/d58c9ae08eef29500a5bf2870e2b136cf317de0992e07d5be2a8fc5ab35b65a2.jpg", "reference_inputs/images/4bbaf7b875b927803278ab0e3bbe127dc986d7e632e71250f020a2395eb683e5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000003", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [154, 258, 303, 318].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/31c6abcf7eb4aa46f118f145da3d0ff14ff0c62da2378150b791ad7d61f0b10b.jpg", "reference_inputs/images/f6477f2676e348ed6ac8c5e36ad5d03777705ee4cd07bf1376693b6eee8e1d2d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000005", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [299, 222, 414, 305].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/5a50298408b5a44380b325f06a63c9e6f92a85c81613f17465ad2b8e5b4836a3.jpg", "reference_inputs/images/fa77697dfc4085fced112433ab375a16fb84c526f445ccec94d4d1edbce39c59.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000006", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [505, 68, 535, 137].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/72e6066675405adba0630b30e491e278d7c25335b22456974782fe3e844d5eda.jpg", "reference_inputs/images/b0382be12d2c85e51697193d9193414c662b366062cf8c11b124a8efee5b4bf9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000008", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [662, 276, 746, 306].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/2a4ced4ecbda3f0ad9bfde9034be5e0a49ec9bab9c712884affd3ad08ecb94fb.jpg", "reference_inputs/images/72e6066675405adba0630b30e491e278d7c25335b22456974782fe3e844d5eda.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000009", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [78, 249, 106, 302].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/a863c0425374d378735269f9a18b0e786addc41c4cfe1ec95b61f4e2c1feb0bd.jpg", "reference_inputs/images/dd757b41e3a6615c95ab1be539de5dbc5ff534d425852717977a24e6142ba69e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000010", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [906, 250, 928, 298].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/9ee52d4b590978577fe3a21b750b8093f46ed3b4efce315bf8098e1feb0e6325.jpg", "reference_inputs/images/334cb2b5cce36bd8cde7417ad4671597bacfb74d13eda21ecf0bfb3b49d6419b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000011", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [447, 206, 559, 310].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/44d016d6bc7786c0ccb7737f4ed263b47be79d3f688d9d39c0375e18dd37137d.jpg", "reference_inputs/images/f6477f2676e348ed6ac8c5e36ad5d03777705ee4cd07bf1376693b6eee8e1d2d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000012", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [391, 19, 401, 35].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/78c4d62d677825deee57612396c5c7a9551bb4faba10dc9c14391fb1ab84f3b3.jpg", "reference_inputs/images/564ec729fefa0968379ec0be485704b80be1d9bed12050c65437d7dbb5776981.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000013", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [510, 56, 541, 125].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/8c8c71d5b70713619640b57c796efadd1e32035dc8e87f9885b6dac3f7c4a2c2.jpg", "reference_inputs/images/cc8944aa5d435ca2466183a66169534e98db87d523d527b86a53f9a05a7ad589.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000014", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [492, 275, 557, 307].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/4361054b280cd9d73394edc6e371afc818b989aed9f9327cd8d6243c79133116.jpg", "reference_inputs/images/f6f1c52deeda02696e616062d3700b17ffeac1e2d963d9ec5a0d768c6fccd7b7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000016", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [843, 209, 933, 351].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/3f6b8fcca774c8e8aea257100830169913e8608cc0bcabc19239f3bc3545c962.jpg", "reference_inputs/images/10e0caccd78914ca0540a51e20bb43e216355caae6de0b951c41c6400442c2ba.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000017", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [154, 270, 365, 319].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/b01bcd72038777584cc4a4cc511b5af23e6bd0cf19996a6ed670e78ddfb4a5e1.jpg", "reference_inputs/images/3b817beb70079de70a053c49a0c92810b0f54beca2a9e32a535454ce3b66b03f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000018", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [890, 226, 921, 311].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/8fa263fd0a2bb14dc52206fb1d755d9a5f3fdea1c2da06223795028c853d0a3e.jpg", "reference_inputs/images/ce00400318ec5a563e2db6cb9d771e702629e1ec0c577a36a04cd47af22eec3e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000020", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [774, 255, 790, 299].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/c0d2a3e08cf989cab00b16bae02eabdb9d196be80eb7ba851f154ec5d4a68f11.jpg", "reference_inputs/images/b93a5eea86d8ed6ee483ca1841644a860eed2f66bd6ad287b7ff0aa36b80dc9f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000022", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [738, 263, 748, 294].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/d6ed40f9a99760bbce21fd56ff793cb50ab6368c56c4fb93d57ab5a08b3cb049.jpg", "reference_inputs/images/edaeca5eeae53e66ff713db1ed142af7f2ff6833c4cbeceae71b3dc182709b2a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000026", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [305, 267, 326, 292].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/50a8652b77bf5cbed4b0cd0bb8bab9494f07e35a1602248d716e193cd683d51a.jpg", "reference_inputs/images/bb8e9d2ac3d59e32364282074e4f3c9e8eb0c951cc7514e946fcebd5fc357468.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000027", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [490, 42, 562, 82].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/4d54ca25c075d2117a23d1a492c6e787b27d6949726c7e86152b21bbed22fa3a.jpg", "reference_inputs/images/ff6c1dc8e7cddbfa32a9b40096be7e72aaf3a09ae2f0b40094b712370828563c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000028", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [548, 182, 600, 313].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/4d4be13a5256b9afde9a01a0399e95853f4c8c40a360410da68a3d90e6d97fa3.jpg", "reference_inputs/images/5e787882a87d24e352cab13c717c715f81757643a7d3e3a57a774b3fed9ac3ff.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000033", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000033", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [506, 272, 542, 300].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/b84ea74379a0dc35ed4efa117f79ffd3eeacf5644d4810946141066cea215ba7.jpg", "reference_inputs/images/ca513ffc9b1c99c0f5ed3236a7524c984df0176f05fa92e5b3fe95b58184d6e6.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000034", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000034", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [145, 254, 166, 301].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/dbdbafd1873f8374fc2da3b5dc8e39767781e722be8ed25fdcf2da41eac4a224.jpg", "reference_inputs/images/645c9bf1b25cd76a2843ff2e95d1d9d7f6d29a09e1a67527a227fd74797152f7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000035", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000035", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [455, 463, 475, 501].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/a1c511f49696a445fff0f4cd0794fc2a08afe9322f7069f11eeea41c460a318f.jpg", "reference_inputs/images/88604a8a195412bd016ea404f6c5a296199e3ec9b21d70fb637c712eab67e52e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000036", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000036", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [498, 235, 526, 314].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/46a777e3a51d6c39f60dd75a2a9a9f7a3b523703d394b82437cbb2aba0121458.jpg", "reference_inputs/images/e52c16d1d979c88426a29b8a8a019a1af495fab894de56409f5f420abf1b6662.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000038", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000038", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [542, 425, 614, 497].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/b68b6dbbc1bbfab91556c7af06bfa9a3d5ce616d1538f94c6a4c7b73338b2daf.jpg", "reference_inputs/images/fc4a796869a6465d6668f9ee17b73e88031099fc46ca0dc1820dd139d0835b3f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000040", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000040", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [478, 546, 494, 570].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/a0615fc5d592de64278ec88960745f2edc782ab5e116e4a4c1da9eb5e026d467.jpg", "reference_inputs/images/b4d21a7e3f68d56255291f70c6384a51e8da3d9f3012b231848b23a57471c549.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000043", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000043", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [446, 502, 474, 554].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/8b4439705634f7cf79b30484806714f7f226ba06ade8af01d6630181b2a604d8.jpg", "reference_inputs/images/360617ebd1ddf8668083527a4aeebbdda1341224f852b2297b1aa4e0edd6dbf3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000045", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000045", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [485, 296, 534, 322].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/850fb0f29a14c4ee2c2467fbd9fb5993039d76575fc152877126bac4451a298e.jpg", "reference_inputs/images/f193bb00ef247ef9d88a96146d29e7f7a175c10da3e12d859889d7018f720d73.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000046", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000046", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [235, 11, 255, 28].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/f46502989374e8640303553bfb8dc805aa142c4a3c72827a9bf6ab93826cb737.jpg", "reference_inputs/images/1e39378ac5766e78a1ddbeac3d2435169317ebdbb78727242c19741fb2c1f858.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000050", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000050", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [246, 254, 342, 313].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/9d12e0d7e2f66043454da17dd4d6bf2b971c6c1620f747bf4be58db60909059e.jpg", "reference_inputs/images/a0615fc5d592de64278ec88960745f2edc782ab5e116e4a4c1da9eb5e026d467.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000051", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000051", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [854, 473, 870, 486].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/8965f22b2f72d587c9a9005be33de5ff477b593ae296956517dbaae4a1e4a5db.jpg", "reference_inputs/images/ece3bae0329b27511ceceb90a491585b4bb5d99c62cc3ac9adf979872926aec9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000052", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000052", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [498, 186, 608, 337].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/502fe2b7bf5d1fe5c44e96178c57cdc8f6cda7fad2cf013c6853ede35a442838.jpg", "reference_inputs/images/01bec1e6063f2abdaa5a2a82f0cc79873e8527ba96cf4d200ccc3916f0c1bc18.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000058", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000058", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [27, 239, 138, 299].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/1be0d45fc5eb88503ea3915f0b8fe008a69f91db0a9f5307220b605630c8732b.jpg", "reference_inputs/images/9a4b8a0ef75644fc6326ef5f14e66451e3488ddd11d148c7e81dea717a438b9b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000060", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000060", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [455, 205, 532, 304].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/59889975d1c2138d1cdf6b09ce66836cf17bb32052b25eec4eda07580915cf06.jpg", "reference_inputs/images/ba07ee4c302489b0010eac852c92f18584fa7e1048095e5bd08bbd8b03343869.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000061", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000061", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [406, 502, 416, 514].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/e9543968dff9f10b27bfebf5ed34688a0cb2ef488405812130d4d1ea16399eb8.jpg", "reference_inputs/images/f9bc503738885b01ca235207b496c464e9c81793a69812cd6c0afbc81583385f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000062", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000062", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [343, 273, 424, 306].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/a4ff943138300e14df5dde6b5b4d48e1550533ac77ffe671ed28e02e96353024.jpg", "reference_inputs/images/b93a5eea86d8ed6ee483ca1841644a860eed2f66bd6ad287b7ff0aa36b80dc9f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000063", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000063", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [507, 490, 520, 506].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/a1beb898158c1d9a3e2c39049b040ab4dadc0fe20a5a863cd02f1ef4e80bf859.jpg", "reference_inputs/images/46bb3bd981a5ae2e0219fc7f3f31f830d5b4bef526551466cbe8e18d63df974c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000065", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000065", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [342, 262, 424, 314].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/bf6e1f1858c88fb977f0dbc32936d9c13cdd676f51a1e07b64daf08ce71bb42b.jpg", "reference_inputs/images/aa2b8d84cbfc6eec2f9f80bd81595ceb78624e4ad064b2ac46b70eb7732e9d05.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000066", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000066", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [390, 462, 401, 482].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/8acf094b5d622819c031b510b0a37ae772303e14eb6cd58466b0dad41852fb44.jpg", "reference_inputs/images/b3e61a6479831944e0b4dbdb7dcc29239bd5d34ca3e7660f572cd61e6b90bf2d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000067", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000067", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [604, 280, 626, 295].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/98e184ef25113e99d3acb81d2b999b2dc5d3a8d18056ae3ea6167479bc81beba.jpg", "reference_inputs/images/c31d00eae2750b53ab399ff00ab6c130a7ae1a57468953c9467d116200a9fcf1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000070", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000070", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [872, 234, 896, 307].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/52169b7a0b342f2d63699bcc716256cbbe071a240289254b98442a1483faace4.jpg", "reference_inputs/images/3bde83c90aa2675fda8182f60cf5101e9a061fcede4f93aed9f93b53dea81928.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000073", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000073", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [563, 272, 606, 300].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/34d027dbab11397f568f524e93a1358bb23b7502da274b0c04a1dba07548b837.jpg", "reference_inputs/images/cfa18aef7c5f1fd61ebebff7dafc64d1918ac5e9d9bb3839c48b3dd33dba961b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000075", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000075", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [295, 262, 450, 326].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/ff8fa66cbab0447fe511d276663e38da9d137919753d5a15e75769603d65bef0.jpg", "reference_inputs/images/edaeca5eeae53e66ff713db1ed142af7f2ff6833c4cbeceae71b3dc182709b2a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000076", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000076", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [527, 437, 538, 447].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/bec159013db8b15522ef914027d5b4939fa6aa1ec865aea29e1d74b66d73bbf3.jpg", "reference_inputs/images/22e82f3f5da3ca8e0a41507973a1054fc81a0436bd364afb7118546fe3ec6147.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000080", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000080", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [46, 246, 94, 302].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/1c86cc5951e53a469d0dfa1d49811abb2200a75b723703e0b03968751a5c81cc.jpg", "reference_inputs/images/f221b3b69de7e07be0c49324d9ef559f3a8f8d4b2de7913d4e7f3da6c599413a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_30m/000081", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000081", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [526, 434, 540, 456].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/f5158663b6493a36ff24cee09cee5d0205729b56889b7518818f61b53f9e3b4b.jpg", "reference_inputs/images/efbffbf5cc31f2b35479861871fbf8dc5c049c8a9a6f5acc5915be6086160d5f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000001", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [533, 37, 562, 116].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/2f79ce4bfba6673ebcc4059065d875c12c7004a0bcb4ab1863efafaff622d998.jpg", "reference_inputs/images/33479d35b44aa18cd3869f33debd528efca86b2a06e5055948e2dd6a95cfdcf7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000003", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [540, 159, 562, 203].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/90d6c35940c03a21763073bbadcc743ee401c41b7327f8103194e6255a97f686.jpg", "reference_inputs/images/4c69c5d98fb3fd3e9d3c953f7532ffdfef243195db145b25042b014624479f9b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000004", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [688, 260, 1000, 350].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/2c4d8a2bd0304a00d3fcdc28715ad25aedb7c02628a93f4a466fc02ebe7ae0fe.jpg", "reference_inputs/images/d1ee783bb91590ea5128726b99ed4acd64eb536a4dfa2e81f2a4271e019ffe16.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000005", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [495, 247, 573, 315].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/176edff7c377ecfb0a85bbb5fe6f909debb69e63dd867a4d9757b6b09591aec1.jpg", "reference_inputs/images/8bc139be26f110733fffacdf00532d34cea4f956a85c5c19c599c4ebbb253eba.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000006", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [434, 202, 454, 245].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/5279243a189dac5df5abe787abc68a73bb0ecaf5b22deee4090ef8daf12ab644.jpg", "reference_inputs/images/e6f52202e903210d25243e3735afe40fa075f6ed21c211af9dc0feaa58a7cb34.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000009", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [507, 155, 526, 194].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/d6d5c7567398c19334c521f3f939a35c6222e35a97340cadca037bc7534d30dd.jpg", "reference_inputs/images/314d01bbfb4d254b346e7f6bc4fb62da813832c57ce8b979e13a5d0aae67e2bd.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000010", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [464, 173, 702, 370].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/1f14b718b819a632f51cf4ea6d5486cc08d1760dcb059aa48113fb4a35adc9bd.jpg", "reference_inputs/images/d1c99ae970acc5d983e9425d72c5654176ba542ae95b32114c33ffcf0ce51a4c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000011", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [227, 202, 392, 311].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/46bb3bd981a5ae2e0219fc7f3f31f830d5b4bef526551466cbe8e18d63df974c.jpg", "reference_inputs/images/88901f1b80fcc6a22add5b0413965bc44db5260cc9fff6217e99c9def73f08ba.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000012", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [809, 437, 854, 457].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/d1ee783bb91590ea5128726b99ed4acd64eb536a4dfa2e81f2a4271e019ffe16.jpg", "reference_inputs/images/64ecb14e96931fa12d63d7d52e97e6551d9ab8794bec873526ad42f1cf64fd39.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000015", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [505, 142, 524, 186].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/5d2f1767d27bd298c77346a506fc1f34a7c27bed6120a4c3a93840e6e303f7cb.jpg", "reference_inputs/images/4bef85cf9114015576227a64dbbbb0fb2abfbf1cc0ae73d509da424e5b2b3f87.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000018", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [321, 72, 331, 82].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/c613831f74943cd7a56399755b0acd20cc685ae6b5644f5cb57b328c60bb63ef.jpg", "reference_inputs/images/b4d21a7e3f68d56255291f70c6384a51e8da3d9f3012b231848b23a57471c549.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000019", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [364, 278, 417, 299].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/389a42f43ad667172d6855c4182d3f7f74d9b73cea54b3b04b85012cce008507.jpg", "reference_inputs/images/fa1285aa27505af151765e72343905c87820febea37d8825c4068020ee9ec089.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000020", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [522, 274, 534, 288].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/068981882776608705076f4c4c3e998b66611be3d5507c1608daae3a21fad37c.jpg", "reference_inputs/images/414068548b23bcfefea8c1f2c2a3e487b06a62f542faebda1598cf584ee6e015.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000021", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [436, 250, 456, 292].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/2f293f0fcedb50ce4f4aa0d6cc8a6301ac451c58df4f200c613671bb4fd1e5e1.jpg", "reference_inputs/images/9e191a9c833a6d0191d5f681a95a3241003ac2ef139ea07eb111a9690c44cf99.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000022", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [490, 227, 698, 366].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/41e3b47493e7cef3d7bdf3f8545edf826aef809002ba8b49baaeb5b7323e94ac.jpg", "reference_inputs/images/fa1285aa27505af151765e72343905c87820febea37d8825c4068020ee9ec089.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000023", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [530, 22, 551, 66].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/1067e98aa88001b98a828b597e2008619fe0196c560610681b2ecb026556ba1d.jpg", "reference_inputs/images/e8f86a144ff07fab1b6ed4595caeffbf95d4782f342c62d61c804499bb0065bf.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000025", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [490, 244, 680, 366].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/1c86cc5951e53a469d0dfa1d49811abb2200a75b723703e0b03968751a5c81cc.jpg", "reference_inputs/images/5d2f1767d27bd298c77346a506fc1f34a7c27bed6120a4c3a93840e6e303f7cb.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000026", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [496, 264, 577, 321].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/ab4684364fe6a38c2026ce521fb39e2527721bca2271eb4e42b0311ffe907e38.jpg", "reference_inputs/images/1741de326d767084dd4e5fd0fc6e3e2bad8b0f6b9c81fdd632056be66516870c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000029", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [507, 216, 527, 258].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/6b9bb2c06b1dd41d2645fd246fe9fc586009519317c0a6d83047346bf6819718.jpg", "reference_inputs/images/f807901de3dee9c2d7c7f64e832960a87933faac7c63f54bc8101d31810c5d2d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000030", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [32, 167, 330, 324].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/135d5add10ec73a8f8dccd1efdf99084b9cf438829994e55973c598984e0abbc.jpg", "reference_inputs/images/0834c11265656ac724ffc1aa245ff03a0d2c5a91325a8469463638eae5fb2714.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000031", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000031", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [506, 317, 527, 359].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/d1c99ae970acc5d983e9425d72c5654176ba542ae95b32114c33ffcf0ce51a4c.jpg", "reference_inputs/images/4186c556f100d93e53855131b95c8bead2b9379f008b39b0b666c3427e84b2f4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000032", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [444, 282, 465, 291].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/42d29e7610dbd9d7ed53070484ca26768cc0f73c1d9e32dcd9365b6a0432c197.jpg", "reference_inputs/images/a37a9af0d31fc71ad365498f5ae3a0da85b44156d0da5f7360ffe7e57f7910f8.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000034", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000034", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [334, 265, 435, 320].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/1f4e5bfc1dc07dd2fc6ab971db43666d4bee8f303a3194ecb836bf23599c8aa9.jpg", "reference_inputs/images/9825ac2cf22175ea4b1eccdc5ca6a7fbc946d54fd5eba1875ac2cbbcef46b1e9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000035", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000035", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [352, 269, 358, 291].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/b84ea74379a0dc35ed4efa117f79ffd3eeacf5644d4810946141066cea215ba7.jpg", "reference_inputs/images/620fd84c54741d059022f3cef18c57fef7e1a6f5d4530e02050417f9182e68ba.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000036", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000036", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [477, 268, 601, 327].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/d805d4a2c5ab72663e1777357194e745b5bdfd434697b3d4f6a88ba3b05e17b3.jpg", "reference_inputs/images/5cc168c4f41abb4da2660f9d26603948f951d0fc700b54023fa10e550294abe6.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000038", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000038", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [322, 277, 428, 302].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/4a9235ef98cd0be9d0efdcdc3660811c02b4055a0d190afdc8cf695abe4f7bee.jpg", "reference_inputs/images/7b151ccb6720a410136b56eaa274076a3a78d2dc385f54f1531fc6f92aaf3fae.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000039", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000039", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [589, 265, 598, 288].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/14c63e7d5653e75ed98311d7c01089858201ac6abe0a8abf88b1d724630e404a.jpg", "reference_inputs/images/6bc98e360acfa81d5ec96bb415f84c681ae4966689e600648ffd1d915a4203af.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000045", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000045", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [190, 254, 321, 325].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/ee79ccb6790b2978c939ac1dffc6c1fb7f5d18399cc47dc5d3f82e9224410408.jpg", "reference_inputs/images/2aaa410ebc44be4d931c2709933aa25f61866a2f6d6337e3d1299933a8734128.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000046", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000046", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [599, 308, 625, 335].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/0c12d48b25c8bc4d5b0ac7ed14b32318c9bab11bb184b1cdbb5f363e3f5c7fe2.jpg", "reference_inputs/images/2f79ce4bfba6673ebcc4059065d875c12c7004a0bcb4ab1863efafaff622d998.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000047", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000047", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [1002, 94, 1020, 102].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/0f3540ec0501eecd3c0306c95cfb87caf749f53963491588b46f3aaafba62bc7.jpg", "reference_inputs/images/1d5f9dec7756bf1faa672e44115577f7b1d0a59014a9a88e1d52979f20ad2a82.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000051", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000051", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [958, 381, 974, 387].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/1067e98aa88001b98a828b597e2008619fe0196c560610681b2ecb026556ba1d.jpg", "reference_inputs/images/1be0d45fc5eb88503ea3915f0b8fe008a69f91db0a9f5307220b605630c8732b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000053", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000053", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [299, 222, 414, 305].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/5a50298408b5a44380b325f06a63c9e6f92a85c81613f17465ad2b8e5b4836a3.jpg", "reference_inputs/images/90d6c35940c03a21763073bbadcc743ee401c41b7327f8103194e6255a97f686.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000054", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000054", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [434, 261, 456, 305].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/93df5d76f25eb5b5a9e6ee01d8da4dc93df5407d5b90f8ac9ed2264eaa9b2d22.jpg", "reference_inputs/images/4d1429f8a559ffa76ff6dd6181113d642a7d6ccac748309b027172a63950138b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000055", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000055", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [778, 129, 830, 162].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/4f572f335c2a8fe2dd6d0e3ee4134e5947d4f62a9a958e2eb9d9812427b49b1c.jpg", "reference_inputs/images/35a9a25911a9f5b520854214e6ff4b226537fa7e88bee6e19c591f7b4b62e2ac.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000056", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000056", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [927, 246, 946, 302].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/182266c2f040fdee6aaaa0e2c96f76eb54a0cd36c40e98e462f0c0ffbd783088.jpg", "reference_inputs/images/a37a9af0d31fc71ad365498f5ae3a0da85b44156d0da5f7360ffe7e57f7910f8.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000058", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000058", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [853, 227, 859, 246].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/a0ccf0f5e696281230600960e2603ce51a994ea07386f34ba1410ded71be1b4c.jpg", "reference_inputs/images/ec9ee0259f373dca2413d97a227e137f2436e9424f801502cc802d3aedfa473e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000060", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000060", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [506, 101, 528, 145].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/5c0c04b288523e43c62b87496107f521b57697e3c84b671f772cdf6d5a80dbc2.jpg", "reference_inputs/images/22e82f3f5da3ca8e0a41507973a1054fc81a0436bd364afb7118546fe3ec6147.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000061", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000061", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [197, 142, 242, 165].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/93edce0f073f265e71493e76d5e9bea18cd0070de557a1dc8f9748a53742b95e.jpg", "reference_inputs/images/0835003194a0b3507fa1fa9fd6fc3b9c6824642b011ec2703d1fb3540b958ca0.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000063", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000063", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [13, 563, 33, 574].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/3e8d25f175bcbdde41ce4c6af50485ea88ee7c204b7bddd88503341235b27e2a.jpg", "reference_inputs/images/9f8bd7afecd76acd880fa4f00ef12b38a4ea1e2d76f817d9fdbce48abab400e1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000064", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000064", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [520, 234, 565, 312].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/8156fbafe831b86bac65b3a21863b8beeddbe10ef22d7a56993d4274368b123e.jpg", "reference_inputs/images/4471a64a77ea01d2f778c46a4df407fec00e0ad3aa7a02283dc971079df7e68c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000065", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000065", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [542, 154, 564, 198].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/ededbee325399e98678b611d225826acf13d11765882ee4f176ffa431c6d154d.jpg", "reference_inputs/images/64aea7386a93b8d411c31c3c476f955f4bb9d442dea86752dfdbfd3a90eec182.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000068", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000068", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [512, 242, 618, 325].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/6cb61d22cf6c69f8660fb55b13f5d0e640f407992bd7a3629e6ec83ae5883cd5.jpg", "reference_inputs/images/e6ac568847fed700544b66b14c167167d1eee1a1f47aa9f545ae15d69b8fa152.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000071", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000071", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [398, 214, 440, 347].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/7bb72823dd1ff767bcfa0af18ffdf886134eb00fa75108eafdef971bd5ab925b.jpg", "reference_inputs/images/ab6d1bf6377517890052514fce9c2ff6f0574b4f6e6da70f73f7618d98f736b2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000072", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000072", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [729, 270, 780, 298].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/1806b00c7cc3be64da79dd17fc4d0fc017661d03ae2f1c7ab217e2a4bc5b3b0b.jpg", "reference_inputs/images/f1fc913aecc7e9120f053c8663075dea12b284c69569533294f986d4a2d2893c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000075", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000075", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [807, 249, 835, 302].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/a863c0425374d378735269f9a18b0e786addc41c4cfe1ec95b61f4e2c1feb0bd.jpg", "reference_inputs/images/d694827fa58b38fa5cc8c8a7d90659d6e5ce32234bd64db1dda9bec7d2c65c90.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000076", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000076", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [617, 458, 632, 465].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/8e830185c444506afc6719687ec1c17624b01db7c88652e33518d29d0f62951b.jpg", "reference_inputs/images/0babe2b74ff3bef3cdac98263fe4524186cb1fd3c2bd927891a0e7a8de627389.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000077", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000077", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [509, 252, 527, 292].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/e2d3d48257038542e95080942944bed57b0dd925a0658170946298c77bad1501.jpg", "reference_inputs/images/ce6b08db9a41daef36af00206c758b526312c6cf41bef6347f4d72c299e9eb2b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000079", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000079", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [30, 284, 125, 304].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/010870be1ff3452306b89e08d9db3ca31baec1660cdaf57e6ddea918045c7365.jpg", "reference_inputs/images/3e8d25f175bcbdde41ce4c6af50485ea88ee7c204b7bddd88503341235b27e2a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000080", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000080", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [434, 522, 454, 567].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/8d302d30b5c17ddef8f9d1f4c3a26b763ede6edfa8aa11970505ac67ad8afc28.jpg", "reference_inputs/images/1c045cc9d56ac5597dd52c6857718f20cd7716c7605e6a518351ebcde2bc3b2d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000083", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000083", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [43, 247, 71, 305].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/efbffbf5cc31f2b35479861871fbf8dc5c049c8a9a6f5acc5915be6086160d5f.jpg", "reference_inputs/images/6b9bb2c06b1dd41d2645fd246fe9fc586009519317c0a6d83047346bf6819718.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000084", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000084", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [379, 36, 390, 45].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/dd13bd300738eee426faa5a3b0287c593a83a0111f3fa8ede7064a49048fad82.jpg", "reference_inputs/images/59889975d1c2138d1cdf6b09ce66836cf17bb32052b25eec4eda07580915cf06.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000087", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000087", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [65, 210, 144, 314].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/ce18c7291cac5b524f1a529a27f2a334f06d216fc6d67f3a8779cac876076b76.jpg", "reference_inputs/images/6bc98e360acfa81d5ec96bb415f84c681ae4966689e600648ffd1d915a4203af.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000090", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000090", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [509, 210, 526, 250].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/7b151ccb6720a410136b56eaa274076a3a78d2dc385f54f1531fc6f92aaf3fae.jpg", "reference_inputs/images/fe8e110d6bb38794d1bebd55d8ac04077ab3b8a2e90fe6f56803526d46685019.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000091", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000091", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [154, 258, 303, 318].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/31c6abcf7eb4aa46f118f145da3d0ff14ff0c62da2378150b791ad7d61f0b10b.jpg", "reference_inputs/images/03fbec1ea895c3803cbd4ad09f3e7cfdd748ee9c9cd4d6e795535db4b9634076.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000093", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000093", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [133, 271, 209, 306].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/6e3dc1f4e62b47d16d62d043bed5a418481ebc45f4896c9dc7e3614a0781bb42.jpg", "reference_inputs/images/878e4482608f331a650f587188930bd1c430697009c1fbe1fcf0812411ccb569.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000094", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000094", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [512, 44, 532, 87].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/31e2580c1002d36760604c48f72f4c237f21fb3f6cc8a92e05357debb8a80bcc.jpg", "reference_inputs/images/0712923cb9ef7bc88dbc73ca5969dd9ffecc54a415e0f4d2e3a28e04b7f8a7a7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/city06_ground_aerial_v2_50m/000103", "task_name": "occlusion_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000103", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [435, 250, 458, 294].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/978982e89ee44383f5f276568b615973f4481d2be3a247f8cec0f555617eadff.jpg", "reference_inputs/images/e74ac6cd045c319af1517ac285faf98c10db9e03aab35bba1c6aae874a49105a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000002", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [98, 292, 106, 309].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/e52e18a8bf400eae4170aacd6f355bc610c96518c41f87dbac4d2690c75ed766.jpg", "reference_inputs/images/affd2dcdaf8eb527c0164b784fcca03599307d983a8ea5f904659e5ebdc351f7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000003", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [739, 328, 890, 538].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/e5a42d0f22c9ff0bed56dd93014c0350a47427e964250e651c6556b2b9084c14.jpg", "reference_inputs/images/cf35a3896ae139d8a9c9b9fa3914778f5faf06fbcc22a89b42402d42461879ae.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000004", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [802, 405, 833, 414].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/46ca6a46d8e89f5f6d2288e2ff737e8f34222d083d61b4f925c0d2b9633da384.jpg", "reference_inputs/images/f96a9e8a485297f23f48e8c3a52f901786fcb17edf58d48e4476b02dc9058311.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000005", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [529, 302, 565, 322].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/462d3dff450dfd5c907f20d1d934b7066d1a5c8b3121c61e4aef9b8e13af0603.jpg", "reference_inputs/images/3af649bc1ded32eb4c23a0be47a0841d61cf1666558c6fd8369d70e79125362d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000007", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [734, 308, 814, 340].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/fb83154e0a62656e6b5c69979f54dd7b956389860903523b5d68cc091995ba8d.jpg", "reference_inputs/images/fb34b53b679d1237735335f190243505fad3e06e30ee7b8151f738312973ddd5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000008", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [214, 294, 247, 418].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/2eca17fafe3337c779f21f0ae44babda73246fb323ef84c7e14f9337a5c9e108.jpg", "reference_inputs/images/2a2b24ce924573f78bae10b41de1a97a9d7bbfda58f3f13b6de81b92595c916f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000009", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [489, 287, 540, 501].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/b56bce3718b1e94730e1443a60846f444be7972c9132c668a8ac6ae72b60bc38.jpg", "reference_inputs/images/724372a2fd2a537ed5a192a41177e90751b7e096a5fa9ca0de83f754ba02271f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000010", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [633, 293, 643, 331].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/56fdec6518e6fd442d5bf50715cf842df5618b6e7125f80e58d99b5f95fb2643.jpg", "reference_inputs/images/5e4c2c0e430bd207efd500ecdc990d997bcaf0e0d27b3be7700bf9e63ded022e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000011", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [326, 300, 348, 378].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/f520bf01a608080308ba25141907177e132202479de12c7a69b1f364b16d913c.jpg", "reference_inputs/images/3814847d25158f5ccaa4511249134621b0b048607f59da5f14c624ca40ff168a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000013", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [598, 292, 620, 382].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/d07d66cd43b69aa1bb55d55663d0fe46824488dba231bc2b3f4b6183ee95fa86.jpg", "reference_inputs/images/8d73b72cfd9d23656a556327bbce1ad9e8202251671c9b5d92f488eed380d689.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000015", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [322, 301, 334, 343].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/5f9384bac607ee37e2b8a1e132423adf08623b83de9eaf6be2f0a31ece2f4bd8.jpg", "reference_inputs/images/7b64bdab408c95a6643d57103399d9c188327aad39abda79c85b69484ea97c84.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000016", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [392, 308, 493, 334].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/9454bfbf127a6ef586c8cbd21a88516a04c24bf9b4a4d2d70da2bdcf1d16e104.jpg", "reference_inputs/images/9b77a6fdcbcd164e42d030afdba126226416d464d2bb4b234b73eec459bc4446.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000017", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [23, 306, 100, 326].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/3de4896984724c4e37f4913075b118a4c0f7e543277ace9529eb9cd22780b8ad.jpg", "reference_inputs/images/2af14fbe1a42008fa850bbe6da4cb1076d10c39e20f9ac899fb560071177a9cf.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000018", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [501, 303, 526, 363].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/80b069dc881c4c3c39effc35afeb161452aa0ed35ad026fcc0ae00a47c4660ba.jpg", "reference_inputs/images/90b08f8d8c665cd8665504a44efb5cdfd66bba94700888ae109f6ac82f3ab1fd.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000019", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [830, 289, 846, 335].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/7f306da1e2ba7c279ae85f566e60f85e5d4bd1d852c81d35cd2dcf293a69609b.jpg", "reference_inputs/images/f0ba7d7e7df8b29ab831e513363cc3fe6a14ab9c3fbc17865adcf9798007292b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000020", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [770, 322, 826, 357].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/a33085a882986d9846c956ecfe12c05ab84729424d326f2ae71fa576d3b82a23.jpg", "reference_inputs/images/f459bf256a51aa8a9babcda85af13e41bdfe56751b2ca2d2cae1fc8e8bcc5e15.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000021", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [298, 292, 314, 354].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/c89ebf3b60597642c0c2d49de503c610fcf040523e67508e902f147dff6160cc.jpg", "reference_inputs/images/6a6016d2229b73349092e87b4bb01534da951024528805cce457af4c643e780b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000023", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [135, 191, 177, 208].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/98588555a9656d7cfc9f63266d2fcf2b6eb3064482278c14ddac2c450e16d2c2.jpg", "reference_inputs/images/c32b20010a9f85d7f75228c727ce166db866aa2ad664200bc62120323d137ad8.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000024", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000024", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [815, 384, 847, 402].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/c2217bb8e6aa009d3e3ca16374eb01c1883d2cc9c0bda56a8911c621037229e5.jpg", "reference_inputs/images/3d09ae4f750be8d742c79345cce31f91ef6007fac5bdd8f3d3c9caa054a123e2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000026", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [902, 458, 954, 487].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/913b6e587a94082eaf302612259df56c4680a9c9280c99c8866de0038c2d9730.jpg", "reference_inputs/images/5a90ce31dd9fc1a279cce695eff2a453d4915123a6d3eb7c993b12061bab48f1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000027", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [454, 329, 539, 399].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/852baac5526a6621171f2c610f1452b58769ff01607d0541ea0b62c5c826c2fe.jpg", "reference_inputs/images/9350f6e7776f5d668a325e2aa8b02c45ca142ce9e54713d403ee50178fde93ef.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000029", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [502, 347, 577, 462].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/31aecb76edc4859a56ef7861758fd7369720bf45ec14e2df2b9d49d05015cacb.jpg", "reference_inputs/images/a914e07f078048760de48a07e3dddabee235c9f6c19469cc64daa90eca063838.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000031", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000031", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [731, 290, 742, 333].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/88657e5d510e3c88213f9b4d0c5b23fec29d65b08c341351f225d7bd5bc19fef.jpg", "reference_inputs/images/134a0a514b7cf32f0d517b4b824905fa65c80c3f3ed6427a5280b73ea3ca3e45.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000032", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [478, 295, 502, 345].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/9abc00b4e40f897765fbf70bfcb30be772200c61adedab2028d69cc7c3131385.jpg", "reference_inputs/images/affd2dcdaf8eb527c0164b784fcca03599307d983a8ea5f904659e5ebdc351f7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000034", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000034", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [880, 294, 934, 458].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/0ecc4b6ddfb97425150b96a3e4653e6ba0564ee34609c0b01afd58b194f93afe.jpg", "reference_inputs/images/7b9a15fc20b7d171680b09bbfcd007ee5591003dc770b19a61842feef5fce354.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000037", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000037", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [554, 248, 568, 267].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/82de66c33aa7a0590d806c77253770c7eab69ef7bfc0111dd4211431220f31f4.jpg", "reference_inputs/images/9e940adea7c32759f0c95212bcc83902fac8ef5f8cc4fef74e0d153479f1ec3c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000039", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000039", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [553, 334, 583, 467].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/ae001cd42d76e6f87147dd08d6efa1f9ee182e555926abab696efe15c5f62a73.jpg", "reference_inputs/images/d9984a287ff5447a4720a69174962625e9d679ce734ad6a6ccd16e9ae8fa2a56.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000041", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000041", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [898, 287, 915, 315].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/0030a9a2f17d0981e575323186a1ae548bd2b4df00044ed812786725771da950.jpg", "reference_inputs/images/c9f7c7f6f0afa98fff06c666a9b198933b04106cbece6ddf36d410a8c11d1674.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000042", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000042", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [892, 290, 916, 337].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/0e21b79909dbcad1e395e42f1bc899edc1a61134bc233babed9b8aeb536d73de.jpg", "reference_inputs/images/671c05560af838f8e7ce8cfb9c707b838fff0c85439a5c7e4b8e942e3e784aab.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000044", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000044", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [640, 479, 695, 550].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/4edc6aafa7d35d67daa9f41e6f1f66a25806c272325f634343c0e542173fe5f7.jpg", "reference_inputs/images/f732fef1023b152fb2d8586c31d138f9774e49c0669823a575ceb8d45f09768a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000046", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000046", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [830, 306, 928, 327].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/520a20bec35d284cf13d94a8a4282fe05ae41fb70a996c3593515cca23ed45a4.jpg", "reference_inputs/images/0eb8b1386ccad364a44e4e822dedf75bdbb0864ce37e507dfda4d19dbd039aba.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000049", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000049", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [986, 511, 994, 533].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/cf35a3896ae139d8a9c9b9fa3914778f5faf06fbcc22a89b42402d42461879ae.jpg", "reference_inputs/images/2670bbadc21796f521b16caa83169069111b31b4f027b062a050ea42e6e26a80.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000052", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000052", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [359, 295, 398, 305].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/ad8ea47bda37bc03ac00602c793c9d8c73a1bb0ef98e4a48fd8570146b77e8d9.jpg", "reference_inputs/images/f6ee8fcebd766adf15d2a51cb919078bd990a44b54169fcf0b1995cb5e9e9706.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000053", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000053", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [618, 308, 660, 382].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/2b2d6b22125d98413e964783960ccf933791e9a38d0a3efda03636806edc9200.jpg", "reference_inputs/images/b3179d4995386d7424b127065c3906520a6979f7d9e651da211736cdda95402e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000054", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000054", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [40, 287, 55, 325].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/c950cb11a50cf4d28f9ed3c51658284e921239f740a09eeda1a7e56759f58f1a.jpg", "reference_inputs/images/3af649bc1ded32eb4c23a0be47a0841d61cf1666558c6fd8369d70e79125362d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000056", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000056", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [740, 296, 826, 328].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/696db644fc0e650348f4be016e8e5fd9c181eee2de39d8b0f2f71ba4a8fde652.jpg", "reference_inputs/images/2d9c0d21686be8efdd9a86bf727dc62b7458f9c9b66e2b26748d02e03988a651.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000057", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000057", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [95, 298, 146, 313].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/bc60445ea52425ada75540534387f46245b9f63fd8a70fd45360dfd726a0f883.jpg", "reference_inputs/images/dd01ae03a3c96e543bbcc23433cb23a95262a873ac133ed71a8b9ef474ac7a8a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000058", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000058", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [655, 518, 662, 534].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/14ea2bc6950721a50cbe799744bfeb8e25310267ee07f2c07dcf584490493184.jpg", "reference_inputs/images/a2fcf96c11003aca2199a84eab83f07da750e1a4f2748ea667e5047ea488f4dc.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000059", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000059", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [719, 294, 771, 319].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/c343894f686c9a6f9b4d32f34a3a2c8aabb3398fdb1bac343c44670e40d08d32.jpg", "reference_inputs/images/3fdf6aa1683d61f1361aa6904621ad2986d520365cb8109ac469f6c2d8fc6fad.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000060", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000060", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [915, 293, 926, 322].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/f934db43925ef1b3d6a6ca1afb809539d7ec7ad2202d78a326940d39e7aaf25b.jpg", "reference_inputs/images/532a7bb18fb47c2c7a6afad29eee2e22fe5703751da3632aa370b49a67654041.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000061", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000061", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [599, 289, 629, 317].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/702dd7ba3d6815564455245f3199e57483d587890f24e75a7c679226bfe87823.jpg", "reference_inputs/images/39f792258ee5da7ad110e4a0028f5a09df7eaabf2742743885a78cb2856015d7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000064", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000064", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [430, 293, 446, 335].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/bfb7ac2106c55460045812ada076cb35f39bff466d95a2fb071d9f9b1424445e.jpg", "reference_inputs/images/0d509ed961c3e48994ed68b9ce03b3e0fcdbc19dee2ba6fec72648baf7df37e2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000069", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000069", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [110, 297, 158, 446].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/ef1c9e15657d46d314b74cd595a8bf5a6aa2686d5fedec13ec8318fc99ba7f6c.jpg", "reference_inputs/images/5bc11b7f2e38ac137619a4800f8fe2d0896a00695683ac0192ea35c976912819.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000070", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000070", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [939, 353, 990, 371].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/5e4c2c0e430bd207efd500ecdc990d997bcaf0e0d27b3be7700bf9e63ded022e.jpg", "reference_inputs/images/61971f754619b632336ae31ac3f8a605ddfa8ef8f46e1d2f7529f8a352928ade.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000072", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000072", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [451, 298, 480, 353].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/07d94e51ff3107e6f7e0e5c5ff9dae64bf2f5936871f6cb4b82c4a1adeee663b.jpg", "reference_inputs/images/00de2bf5235b40f382f6df163362e3b5d54ea51747f7f4f5ff7c89b03b95e023.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000073", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000073", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [919, 538, 977, 567].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/f459bf256a51aa8a9babcda85af13e41bdfe56751b2ca2d2cae1fc8e8bcc5e15.jpg", "reference_inputs/images/5343fb3594b5630345448d98043996f02515530e2eddc6c3b7565939086db183.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000074", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000074", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [714, 302, 791, 336].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/a3dc7e0844f017fe540d05ab504b06cad110cdbf6266a30230eafb3da0f36020.jpg", "reference_inputs/images/d5147459caeae67ce54f9d4151e6266c145f10f639a335c96f481cd5daf5010b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000075", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000075", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [592, 287, 605, 312].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/5bb32c662a8a225c89953f6ec3262c596abe3d1bea16f3c828053bcc85c1d525.jpg", "reference_inputs/images/7b64bdab408c95a6643d57103399d9c188327aad39abda79c85b69484ea97c84.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000076", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000076", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [838, 513, 849, 533].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/4eb701386ccf91f87d8153ba6786d69689da1ca369180d1c62781fc6f8b641d2.jpg", "reference_inputs/images/d3734fe6806cf1bea75aedd8fed2e857ca548f7986cb59e09507251fc79338c3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000079", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000079", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [379, 291, 390, 329].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/8ed404085ac069432ca6885fa46e77d31c1a9cee698a07ec666853280f8b6fd0.jpg", "reference_inputs/images/dc12abe33bc14f80ec2fcb9ff08c59844092337b36d418833dcffb48ac003407.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000080", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000080", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [419, 288, 432, 344].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/4d509047d55a78ffadaa4755f6f49f4d32580208ec0bc6968e26344bd67ee83f.jpg", "reference_inputs/images/9f86aa4244ba0a6f10005b92de656d75d1fff6c3bde4f6e41da74851b25dbef3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000081", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000081", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [559, 290, 572, 353].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/1ae237aad839b445a89cc0e8f1fdecf1e406c23485f9ed36443a2be1a3304a8c.jpg", "reference_inputs/images/faf5c975d2ec136c30cf96efcb52dbaabe98daa7b1c8995bf160f483c0b5d4bc.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000082", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000082", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [210, 116, 238, 131].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/31aecb76edc4859a56ef7861758fd7369720bf45ec14e2df2b9d49d05015cacb.jpg", "reference_inputs/images/58452ce8cc4433cb68e7c2cbfc4815e0b370966b43d0c5a42d31e4a0be1a8c61.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000089", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000089", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [171, 350, 284, 419].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/01886ce0440f5aca0e712fc42566fbbda12378686622992519e2b014b4fe3e5b.jpg", "reference_inputs/images/954627ffc0e6664ba0b67f2f0c2cde3e00cdf6d91e6b9b7c888a08095e6756f3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000092", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000092", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [285, 312, 466, 382].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/5275ce1621856d5c511eee6aa53f4b0032a9973d2584e679a486449739a29142.jpg", "reference_inputs/images/2a6d422c5d7113659c40116522d55551f55e3fce579ed0dcd074766c3c123762.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000093", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000093", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [122, 301, 164, 407].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/2c4bc66211ccfe5085183b35111b64fa11a0bb4469dc650241a0294db4576f68.jpg", "reference_inputs/images/b3179d4995386d7424b127065c3906520a6979f7d9e651da211736cdda95402e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000095", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000095", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [522, 461, 534, 475].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/e4aac62165cb28afcbd439a83cee130a5bc480fe9cb15115ea14cabc005d8d78.jpg", "reference_inputs/images/52f9f64e14c06f5f784d017fd8cb72625c111c2ded139cc3bbe2085137c9c586.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000097", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000097", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [910, 481, 966, 502].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/913b6e587a94082eaf302612259df56c4680a9c9280c99c8866de0038c2d9730.jpg", "reference_inputs/images/0293f6ba91e95c198996209632d01757f9985ce6daff86303ddc00421f1cc7d7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000100", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000100", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [541, 311, 598, 350].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/b537d27ae67f2abe24e4fc75cba67f2e56273e13ed72686e305fec5a9130eb96.jpg", "reference_inputs/images/560992cab61444042e93f5a26ac131152313b5d9286f58b1651b6779076daa30.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000101", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000101", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [942, 289, 966, 348].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Not in FOV\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/c472c7a537652977f5214ec7087946e5be959ccce589b4a953e2b9ccd03f25d0.jpg", "reference_inputs/images/9f86aa4244ba0a6f10005b92de656d75d1fff6c3bde4f6e41da74851b25dbef3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000104", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000104", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [848, 293, 883, 402].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/edd28a4ee38936fbbbc099529de69562c3e3ac0cda22e938c43d725ce908325d.jpg", "reference_inputs/images/c302fe92cb7865ef9dcc46662b04b7134af45c3351c5b7dfa2733cf9183ba02e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000105", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000105", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [304, 298, 435, 351].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/15dcdc1bd7ab34f879af8b53c5dd8b7d81344bf08a1850f7621fc81f058846b2.jpg", "reference_inputs/images/8d623e668ea8bf1fb46683001b988f91917220a05ed970bf04e08492689ac461.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000107", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000107", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [16, 428, 67, 448].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Visible\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/40d0ce2fd4a56089ea8837e202fe35b54cc6d6a9d77031cfb4521ab135148a32.jpg", "reference_inputs/images/f00be5f5431d5f216408980ab71683bd8405197facf8cec2f1e73396ba223dd5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000114", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000114", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [735, 289, 747, 335].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/30dc9f7172be21eeed43a6f74e7e237b006c03c0a6567a19951aea978f1d307c.jpg", "reference_inputs/images/17b62a74f74c1ea060246e106301a4bc1105d82173d30d693bf2f9d1dd4a700a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000116", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000116", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [636, 309, 662, 379].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/99996be75264845d962cdba0a1ed1039ac5656471bd459aca071179e1ee05317.jpg", "reference_inputs/images/1a8c244f9fd25465901629aefad10819dec40735eac73ba20e3c6519f9c110c0.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000118", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000118", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [487, 294, 502, 354].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Visible\nB: Occluded\nC: Not in FOV\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/a613f0e862743d96f788c7b41c2e4abb6c48def73ada6194e7a8a319fa81b08e.jpg", "reference_inputs/images/c302fe92cb7865ef9dcc46662b04b7134af45c3351c5b7dfa2733cf9183ba02e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "A", "answer_index": 0}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000119", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000119", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [71, 290, 95, 355].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Occluded\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/599608ea459846df9e025b5797ac0b15ee235342a68d56e510beaa37e256a067.jpg", "reference_inputs/images/8d73b72cfd9d23656a556327bbce1ad9e8202251671c9b5d92f488eed380d689.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "C", "answer_index": 2}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000120", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000120", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the aerial. Image 2 is the ground.\n\nYou are Ground Agent, observing the scene from the ground view. Your aerial teammate marked one target in the aerial view. Determine the target's visibility status in your own ground view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [512, 503, 520, 523].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Occluded\nB: Not in FOV\nC: Visible\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/d9984a287ff5447a4720a69174962625e9d679ce734ad6a6ccd16e9ae8fa2a56.jpg", "reference_inputs/images/40cf4580edadcb9d01aed3a17b06aad52789138e141da5f4c764ac23632deaba.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "occlusion_reasoning/town1_ground_aerial_v2/000123", "task_name": "occlusion_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/occlusion_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000123", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Binary Occlusion Reasoning\n\nImage 1 is the ground. Image 2 is the aerial.\n\nYou are Aerial Agent, observing the scene from the aerial view. Your ground teammate marked one target in the ground view. Determine the target's visibility status in your own aerial view. Choose whether the target is visible, occluded, or not in your field of view.\nThe marked target in the source view is the same real-world instance you must reason about in the target view.\nChoose one answer from the three options.\n\nThis is a binary visibility-status question with 3 options.\n\nThe source target in Image 1 is marked by this XYXY bounding box in pixel coordinates: [757, 292, 781, 374].\nThe bbox format is [x1, y1, x2, y2], not [x, y, width, height].\n\nOptions:\nA: Not in FOV\nB: Visible\nC: Occluded\n\nAnswer with only one option letter from A, B, C.", "images": ["reference_inputs/images/cb0db9da20c0b9cc4c4332bedcaa62bc9fc99c8659420cd44020cb244ade6d10.jpg", "reference_inputs/images/99ff36e6b012edd7fa9d4c7b1bc65614497e52512fc40b1cf1cc29158b33f267.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/city6_ground_aerial_v1_1/000004", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v1_1/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [493, 416, 578, 494]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 5\nB: 2\nC: 3\nD: 4\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/dde9a171c5ce413e052fdcb9ad8e69f2ca2b4ebe1568e67fed6e8bb0005ac1c0.jpg", "reference_inputs/images/aba475ecd2717e1e73b4621f83b92fdb5b044e8987a09cafb5d89fab7e5eb75c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "region_counting/city6_ground_aerial_v1_1/000008", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v1_1/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [453, 462, 549, 575]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 4\nB: 5\nC: 2\nD: 3\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/67cd59705fbad1bba1dab6b11ac15406564a7f4635851399ef2494871b98f489.jpg", "reference_inputs/images/a2499412fbe3e5142ab2e0e1dfa65da9686ce5ab331de8efdc64bf167b548c4d.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "region_counting/city6_ground_aerial_v1_1/000009", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v1_1/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [480, 384, 579, 468]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 3\nB: 2\nC: 4\nD: 5\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7fa73ee2b6f672acb006721fa6be199cad19218b261e00aa41cd71773e07adfd.jpg", "reference_inputs/images/30e315ebf0047a04588cc2376b8a60326a1b2c2290a40cb0cf14c2d5bea9cfea.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "region_counting/city6_ground_aerial_v1_1/000012", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v1_1/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [461, 410, 558, 470]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 5\nB: 4\nC: 2\nD: 3\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4a0bf52b300e7db6f764f15817a306d16da0b56bc0918ce08ab46c90c5625ff5.jpg", "reference_inputs/images/a2499412fbe3e5142ab2e0e1dfa65da9686ce5ab331de8efdc64bf167b548c4d.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "region_counting/city6_ground_aerial_v1_1/000023", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v1_1/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [161, 111, 858, 462]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 2\nB: 3\nC: 4\nD: 5\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7e059b82e20bf68ec38285668b537402106658ccebcdb655421ac3b5a5493303.jpg", "reference_inputs/images/72a43d61bab4cbe5f9ed8e61b0f85541b2cbee75abbf2bedc161ebd4afcf6e19.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "region_counting/city6_ground_aerial_v1_1/000025", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v1_1/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [250, 26, 850, 548]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 6\nB: 3\nC: 5\nD: 4\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1b41e1ab34f51223b23ba0625a947a3dc45534143f14faebd2560a4356a1fbe1.jpg", "reference_inputs/images/896a4e8c3796c071d393c0debfd1a7b26b3a87a5236121920fec6a3d235bb8a2.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "region_counting/city6_ground_aerial_v1_2/000001", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [481, 396, 585, 562]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 6\nB: 5\nC: 4\nD: 3\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b71fa2908f68c90f66df84e5a4da9f3d214a227629b9e5e6af4f2e807ec4cc47.jpg", "reference_inputs/images/f107fcdac3aa78cdb5eb81a827ad1517eee94ef79da4b8c373cd56a08d643f7e.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "region_counting/city6_ground_aerial_v1_2/000004", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [222, 10, 373, 446]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 5\nB: 2\nC: 3\nD: 4\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/be3e9ad0f7ae194d14494dff9e961fa8f1c574bd5d4cb7b708ef47319b7b8ec6.jpg", "reference_inputs/images/a59fe69eeccd1666a1b45c2f6b22038744071d2b684a89532f538e027bc66098.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "region_counting/city6_ground_aerial_v1_2/000007", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [382, 0, 569, 527]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 5\nB: 6\nC: 3\nD: 4\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/57105fd9c057cd87c857d3a60679b36a5637c30b1da61636f27d7d5ce87065d8.jpg", "reference_inputs/images/c415fa89e595511eb8e2096a022359394ee5b5250edac94a113e4b0f3e1d75d7.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "region_counting/city6_ground_aerial_v1_2/000013", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [417, 433, 583, 541]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 2\nB: 3\nC: 4\nD: 5\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/03c8cae19d6a0e424f5ab435462d5d84c2654462963a2ebd557276463e22e4d6.jpg", "reference_inputs/images/b18e1a5de34cf4e8435099dc7e97cdcb0821c9cb029b92e386961f42b35497ce.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/city6_ground_aerial_v1_2/000018", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [159, 166, 875, 514]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 3\nB: 5\nC: 4\nD: 2\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ec36b4d8496d395c946d3b80361598540c682cecae53a46d68fb3e804c62869f.jpg", "reference_inputs/images/d6007811c370ff43b023a1aa2af2d27d2b60f393f6a6109f7efd5f984811d889.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "region_counting/city6_ground_aerial_v1_2/000021", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [455, 427, 558, 556]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 5\nB: 4\nC: 3\nD: 2\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/de55f2b8ed6627f74bd54b07463b05b924d8e0f1dc95e3f4899e3474a942c5e4.jpg", "reference_inputs/images/2707218e054b7e9a9392237f86f1033629c39894453b66e6fe794d240089cc41.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/city6_ground_aerial_v1_2/000022", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [495, 413, 615, 530]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 2\nB: 5\nC: 4\nD: 3\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ab1bbcaa1e20f56069537f71f0d7982d2560f0bbd252411b9b4ee46570b317f7.jpg", "reference_inputs/images/34dfdb1ff0f3248e6fda07c08e763dec2eeb9d48e453ec2d13a6ee0fcb0be6ee.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "region_counting/city6_ground_aerial_v1_2/000025", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [463, 420, 618, 575]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 4\nB: 6\nC: 3\nD: 5\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1bd145e74ae70aa6512e13a7b74e456509e618fd5162846da62183d2e3f725cb.jpg", "reference_inputs/images/a087e09d8a74c707f80dbf2e6d18c9b7916873e6fe54cd2b4bbc57e5a20a1836.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "region_counting/city6_ground_aerial_v1_2/000030", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [282, 251, 802, 506]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 5\nB: 3\nC: 6\nD: 4\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b7c583067b5cfdd8ec19415c0b86b30045af7212d927d2ae7ee266ba43614ef0.jpg", "reference_inputs/images/f8a8cc7887545341767169340d41dda82c912e09dc5d764d736f496b6978e491.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "region_counting/city6_ground_aerial_v1_2/000031", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000031", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [445, 398, 570, 531]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 2\nB: 5\nC: 3\nD: 4\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/78ae049959eebb5eb52330dd5f0f8f1f10a16d598ea62af420860420d999f019.jpg", "reference_inputs/images/f71d9101d349a6391827fd1d80baf81df7a2a4cbec5723fd9df838864f2d2ff5.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "region_counting/city6_ground_aerial_v1_2/000032", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [445, 387, 586, 494]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 5\nB: 4\nC: 3\nD: 2\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4f3caa1f362e3eef6d3ab945aa1798016c545875f112aa3e48aa18d838339315.jpg", "reference_inputs/images/3dcd6def636bc743c1e803e8818216fe14c11ecb816622e5e8994ccb5cb0387c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/city6_ground_aerial_v1_2/000033", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000033", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [193, 130, 749, 543]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 3\nB: 5\nC: 4\nD: 2\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cdbc1e307c88c5047315e5b37cbc0f0d9fff09f1983fe1bd2f198bb2b05717eb.jpg", "reference_inputs/images/ae0d24a11eeddcda278c1a85335b5b66eeaddbff95dfe30d92560914d88d2c6e.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "region_counting/city6_ground_aerial_v1_2/000034", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000034", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [467, 420, 559, 558]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 4\nB: 2\nC: 5\nD: 3\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1999bc8f42e6d71fd980dfbb66fc834554c785c5b725dad1812d0a472c7f58d0.jpg", "reference_inputs/images/30b95ee90c03d9a61180654d3ca210163e3310a79e267f2ffa61e6e774ee4b5c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "region_counting/city6_ground_aerial_v2_20m/000004", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [156, 496, 726, 575]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 5\nB: 7\nC: 4\nD: 6\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a4cf592a72e7c039e89f2ce68a4f098467de07ca020316c2c156372cb9a1cfc5.jpg", "reference_inputs/images/9762b261fdab1ae82f5704a59390f7a7a5f789c680e4cf37c3c11ba8d2d2fc0f.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "region_counting/city6_ground_aerial_v2_20m/000017", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [66, 91, 162, 570]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 4\nB: 5\nC: 3\nD: 6\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5227f1514bc26367b9a685962255b280ec8c82f911ad7282b2912b36f5b3dc4c.jpg", "reference_inputs/images/b4124de2733458d6b02f2a8d6824132a5a497241fa556c58ee1b573e171b45bc.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "region_counting/city6_ground_aerial_v2_20m/000018", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [86, 253, 230, 566]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 4\nB: 5\nC: 3\nD: 2\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6af64a6f155a53fd7e9405fd0240f3d88f46f26297b3a634601d63726b1edce1.jpg", "reference_inputs/images/d3a0110b6a67af346162179241e854b0f3917a8527d17d0238e525236b325fc3.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000001", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [209, 142, 723, 489]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 2\nB: 4\nC: 5\nD: 3\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/65ab20dcbba598480d730324f3c9ceffe765f5fccfc0bf381aeb02abec648f0f.jpg", "reference_inputs/images/9c6b3f29f4d9176113b4baac3a280ea580c2ac94190eae813932e337951513d2.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000004", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [57, 415, 330, 575]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 2\nB: 3\nC: 4\nD: 5\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7ef8435af79d5039b6ea4551e5399f6861d16b59baa994c4a98a2e33c187dd61.jpg", "reference_inputs/images/a95ddbdca69a13640b9d9595bb317ae802f40659a5f39d4e97c1ceba66babbbb.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000009", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [217, 125, 284, 302]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 5\nB: 4\nC: 6\nD: 3\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3c008430886d13cc06b3af1ca92b44323c9fdc224aa91af5bf1a822dea3e92b5.jpg", "reference_inputs/images/9bd1be9135b490482662595471ecb13389cc21f0185e9afd81e9b1cf2152433c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000010", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [435, 452, 530, 575]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 4\nB: 5\nC: 2\nD: 3\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3bde83c90aa2675fda8182f60cf5101e9a061fcede4f93aed9f93b53dea81928.jpg", "reference_inputs/images/52169b7a0b342f2d63699bcc716256cbbe071a240289254b98442a1483faace4.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000011", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [650, 276, 750, 526]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 5\nB: 3\nC: 4\nD: 2\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a1d1537e3be8850fe6f3702a692d2159724b057ca4725db2e331af7a19f6bdbf.jpg", "reference_inputs/images/fc2f455a3a134b67bdb86f8967c4e888682308e77817c4a1c0ede8e7b7ffb7cb.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000013", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [36, 53, 366, 575]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 7\nB: 6\nC: 5\nD: 8\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5503db5f5ae4225d2c6f3add36d2f86fa8ee0bd3d0e0cce90ad74ba26b5a2990.jpg", "reference_inputs/images/ee43fb8141aa4576e3ebdf22ef02f663d6e11cac7f7e49ef76189736638c9d2a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000014", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [172, 390, 420, 487]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 3\nB: 5\nC: 2\nD: 4\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5c528b787bb0bb05fa3e469f6e93e9a80d6e2cc70504ec83808d14638d39d228.jpg", "reference_inputs/images/b0af3bb36620b6a360c476abd2224bdf5db62721179541bcc51903f4cb00b64f.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000016", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [246, 81, 730, 476]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 6\nB: 4\nC: 3\nD: 5\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b54de1266305b9b20f5b628edb6cc3be8bf303f5eaaa4ba4235d674b186ebab7.jpg", "reference_inputs/images/4183a0b6de334d0e851eef8ba64a5005f51b443781882f4bdd69e0315019eef6.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000018", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [658, 0, 862, 137]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 3\nB: 4\nC: 5\nD: 2\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d23def54bc38c6919f30f1cf05d51e32af4daf8d2413d5881837e1e5da81ad0b.jpg", "reference_inputs/images/eb97a3246d862464d2b51b720f73d209b8d3a46bb24e54d9e5a2b762b5dfa240.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000020", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [662, 205, 789, 426]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 5\nB: 2\nC: 4\nD: 3\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7312f49d0b403f90f880885c10c3f564ca6b1d2fa5ce2b847dca574c103831d4.jpg", "reference_inputs/images/2506433c82e3e48ac5f1db08b9f7e3f477f766a7bfcdea59cc70da55f5e29026.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000021", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [230, 49, 327, 255]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 2\nB: 4\nC: 3\nD: 5\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/223cbee8041d6f1a55407a61b86a321c281765434511cd82b74d02a28bad5c15.jpg", "reference_inputs/images/5c64e995683b1de21c283c671349dac82351d1b3a64375df3ccdb6bffaedb802.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000023", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [477, 428, 553, 480]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 4\nB: 5\nC: 3\nD: 2\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a13155a96d5482793fe30ee3b3c9daecde73d3211e2642cc63bb2348dcd87762.jpg", "reference_inputs/images/16a8e2767ddc2533679e06b8f287131d7eb896b588ca165bca382bb4df8e8d3c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000025", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [36, 345, 366, 562]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 4\nB: 5\nC: 3\nD: 2\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5503db5f5ae4225d2c6f3add36d2f86fa8ee0bd3d0e0cce90ad74ba26b5a2990.jpg", "reference_inputs/images/6de9c3585cf47e11898d5f77222fa22a23cf12d738d4e6bd2252109ea1039268.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000026", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [444, 444, 555, 528]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 4\nB: 2\nC: 5\nD: 3\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d244fd3db364087e2ec3d7b95f0825a29b0801934b47170eb9f4cba1fe921459.jpg", "reference_inputs/images/b719686ac3ce75d89888dbdd46c328c0141d8802f7ef17467032f5f9c3aea3c0.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000027", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [658, 54, 750, 392]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 2\nB: 5\nC: 3\nD: 4\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e02de4f49edc2fa3c357bdc47dfb2209aa693b1f328f22c5d25b242bed0a2c94.jpg", "reference_inputs/images/d29cac67f917b41bf253feba461d4cd1c2a80234455b6dfce3a8745ec2c770da.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000028", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [460, 414, 558, 523]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 6\nB: 4\nC: 3\nD: 5\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cfa18aef7c5f1fd61ebebff7dafc64d1918ac5e9d9bb3839c48b3dd33dba961b.jpg", "reference_inputs/images/34d027dbab11397f568f524e93a1358bb23b7502da274b0c04a1dba07548b837.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000029", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [650, 64, 750, 317]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 3\nB: 2\nC: 4\nD: 5\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/acba93faab2b2b94bd8f121833dcdc0686ddae44c5819cbb8c7ea4fc08b05082.jpg", "reference_inputs/images/0434ee60b659113cb97e064670e8b02e9955eeff7f24ccedb8f970345927aa67.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000031", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000031", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [493, 452, 565, 559]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 2\nB: 4\nC: 5\nD: 3\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f6f1c52deeda02696e616062d3700b17ffeac1e2d963d9ec5a0d768c6fccd7b7.jpg", "reference_inputs/images/4361054b280cd9d73394edc6e371afc818b989aed9f9327cd8d6243c79133116.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000033", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000033", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [634, 0, 890, 434]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 4\nB: 5\nC: 6\nD: 3\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/54fe527ada2d1060e9abe51e500c229043388ec7321da0fcfb22ea61353b10b0.jpg", "reference_inputs/images/794cfbcbe7aebf0fd07de502de177b39301e6da824fb87e6bb1e4aa716ca8342.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000034", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000034", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [142, 110, 779, 538]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 3\nB: 4\nC: 6\nD: 5\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1a0ea3e0c43e1f8333c7b5ac26d86b0d4127ca6debff07ec7d82fe47ce562749.jpg", "reference_inputs/images/c6777d37805bc1e28ba259a064593108c7c27c287b70aceb3538e5e222e83c53.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "region_counting/city6_ground_aerial_v2_30m/000036", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000036", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [659, 174, 790, 420]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 2\nB: 4\nC: 3\nD: 5\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/425ae2a6e520d6325a6ef5ca44e168b8600759bb84f76528dbe67ab98b445533.jpg", "reference_inputs/images/3b84b5d82519d128b30a19c0cdb9e0532e599f2540e6a0fb617747af66d36169.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "region_counting/city6_ground_aerial_v2_50m/000005", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [419, 0, 542, 438]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 5\nB: 3\nC: 6\nD: 4\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5279243a189dac5df5abe787abc68a73bb0ecaf5b22deee4090ef8daf12ab644.jpg", "reference_inputs/images/876e52177c54d2f1a987ee7d6b60aef790c909d21871bccc19c29ad50f3249e4.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "region_counting/city6_ground_aerial_v2_50m/000008", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [418, 186, 541, 519]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 6\nB: 3\nC: 5\nD: 4\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d78eac01e2212aeff0e4dc193dad35c2de808cf46bddb2f94cf15c2e06f99339.jpg", "reference_inputs/images/772ae829aab097e813edc986df6337aa53a4dfb46188b630cff56caf74228b46.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "region_counting/city6_ground_aerial_v2_50m/000011", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [459, 233, 573, 501]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 3\nB: 2\nC: 4\nD: 5\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/978982e89ee44383f5f276568b615973f4481d2be3a247f8cec0f555617eadff.jpg", "reference_inputs/images/e74ac6cd045c319af1517ac285faf98c10db9e03aab35bba1c6aae874a49105a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "region_counting/city6_ground_aerial_v2_50m/000020", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [488, 138, 611, 575]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 3\nB: 4\nC: 2\nD: 5\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9e6c6109eafac788101cfd3b1b7088160bb23bf14895659d84c525d5f69ff798.jpg", "reference_inputs/images/fd1a741907a15ac959be34c50c789abef4ee72af9a0929d747d349ffbe6f6efb.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/city6_ground_aerial_v2_50m/000022", "task_name": "region_counting", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [422, 108, 542, 378]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 3\nB: 2\nC: 4\nD: 5\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d1c99ae970acc5d983e9425d72c5654176ba542ae95b32114c33ffcf0ce51a4c.jpg", "reference_inputs/images/1f14b718b819a632f51cf4ea6d5486cc08d1760dcb059aa48113fb4a35adc9bd.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "region_counting/town1_ground_aerial_v2/000006", "task_name": "region_counting", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/town1_ground_aerial_v2/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [493, 89, 603, 573]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 4\nB: 3\nC: 2\nD: 5\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/98588555a9656d7cfc9f63266d2fcf2b6eb3064482278c14ddac2c450e16d2c2.jpg", "reference_inputs/images/c32b20010a9f85d7f75228c727ce166db866aa2ad664200bc62120323d137ad8.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/town1_ground_aerial_v2/000008", "task_name": "region_counting", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/town1_ground_aerial_v2/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [170, 357, 399, 517]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 4\nB: 5\nC: 2\nD: 3\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/04ae5b63297ef9faa39e0c51d2a61dcbc1fd82e72a57a0281f9362ae3f704f66.jpg", "reference_inputs/images/e6a930cc13c9435d48975c505cbff4ed5d6ffbc32bde3d87755e13d702c854ea.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "region_counting/town1_ground_aerial_v2/000010", "task_name": "region_counting", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/town1_ground_aerial_v2/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [424, 497, 608, 575]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 4\nB: 3\nC: 2\nD: 5\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5decbba8daa7c5df5a7fff7c056b20766744bd23a17ecce6a98368f20a0ea9b9.jpg", "reference_inputs/images/b85a6b29f399b960fda0517b94c5f99319b190ff4a324961395b1f408498e61b.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/town1_ground_aerial_v2/000011", "task_name": "region_counting", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/town1_ground_aerial_v2/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [367, 472, 604, 575]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 4\nB: 2\nC: 5\nD: 3\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/45161aa0655e931e7924983dd2f4d431231d378aad482002b7e8acb1c3c8ed8a.jpg", "reference_inputs/images/dce70d33d8f9602df8bca856a63b807f674d48cc5394e78f6422eb83339193aa.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "region_counting/town1_ground_aerial_v2/000016", "task_name": "region_counting", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/town1_ground_aerial_v2/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [284, 398, 376, 509]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 2\nB: 3\nC: 4\nD: 5\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2d9c0d21686be8efdd9a86bf727dc62b7458f9c9b66e2b26748d02e03988a651.jpg", "reference_inputs/images/696db644fc0e650348f4be016e8e5fd9c181eee2de39d8b0f2f71ba4a8fde652.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/town1_ground_aerial_v2/000022", "task_name": "region_counting", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/town1_ground_aerial_v2/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [588, 425, 751, 575]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 2\nB: 5\nC: 3\nD: 4\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/01986be9f391684280a129ee31aa9ba31926cad35759a3b95020608ed7dfb573.jpg", "reference_inputs/images/15efa56c21d2bc6e235bb1a3ccf4483e7a48fdd00089930beb82826e589d1667.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "region_counting/town1_ground_aerial_v2/000025", "task_name": "region_counting", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/town1_ground_aerial_v2/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [489, 456, 790, 573]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 3\nB: 2\nC: 5\nD: 4\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/dabcddb2c5f31308ba63bf2401e624fa4873357724a5c22c4e346a03c97f202d.jpg", "reference_inputs/images/c871a2b6bab048eb4995aff931bf7060b4d846a70080dfeef714fc030b01d5d7.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "region_counting/town1_ground_aerial_v2/000026", "task_name": "region_counting", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/town1_ground_aerial_v2/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [332, 483, 759, 575]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 2\nB: 4\nC: 5\nD: 3\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/07ca824b99b314bd94d57eda78a95cebe85ce30bc8e7d8698d800d83e704a7c3.jpg", "reference_inputs/images/1075720d899aa88c5aa14cb7eeb20344e150199624b9bde5c85be648db3d26cb.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/town1_ground_aerial_v2/000027", "task_name": "region_counting", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/town1_ground_aerial_v2/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [74, 138, 706, 430]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 2\nB: 4\nC: 5\nD: 3\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/41821988d9f59431e844993ef642182ec744c51695610e657e3aa61b90f80b96.jpg", "reference_inputs/images/5b5e53f763930b5c8ada4a79985b9f6114cb0b759d93796acd02cb2dc9a44557.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "region_counting/town1_ground_aerial_v2/000028", "task_name": "region_counting", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/town1_ground_aerial_v2/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [478, 70, 910, 304]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 2\nB: 4\nC: 3\nD: 5\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/38b14bfd106225a2e22f062250eb7d51bddb58c2ae67d4a7ab3f274540ea8a91.jpg", "reference_inputs/images/57ae96560ad6020664e5a95327c1992dcf7337d3980ecf843a71447b788d5bfd.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/town1_ground_aerial_v2/000030", "task_name": "region_counting", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/town1_ground_aerial_v2/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [550, 466, 843, 530]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 4\nB: 2\nC: 5\nD: 3\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1b772583e746028a242e2d1e0176a99834e2d1175de12585fea52f2211d89320.jpg", "reference_inputs/images/38526620ed5a3af876796ff5433e7cce2e3ad95683e15a623a09116642efa690.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "region_counting/town1_ground_aerial_v2/000033", "task_name": "region_counting", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/town1_ground_aerial_v2/all_questions.json", "question_id": "000033", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [211, 335, 665, 575]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 2\nB: 3\nC: 5\nD: 4\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/40d0ce2fd4a56089ea8837e202fe35b54cc6d6a9d77031cfb4521ab135148a32.jpg", "reference_inputs/images/f00be5f5431d5f216408980ab71683bd8405197facf8cec2f1e73396ba223dd5.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/town1_ground_aerial_v2/000035", "task_name": "region_counting", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/town1_ground_aerial_v2/all_questions.json", "question_id": "000035", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [377, 426, 531, 575]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 3\nB: 2\nC: 4\nD: 5\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5bc11b7f2e38ac137619a4800f8fe2d0896a00695683ac0192ea35c976912819.jpg", "reference_inputs/images/ef1c9e15657d46d314b74cd595a8bf5a6aa2686d5fedec13ec8318fc99ba7f6c.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "region_counting/town1_ground_aerial_v2/000037", "task_name": "region_counting", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/town1_ground_aerial_v2/all_questions.json", "question_id": "000037", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many vehicles are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [504, 392, 636, 502]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 2\nB: 3\nC: 5\nD: 4\n\nTask: choose the option that gives the correct count of vehicles inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3af649bc1ded32eb4c23a0be47a0841d61cf1666558c6fd8369d70e79125362d.jpg", "reference_inputs/images/c950cb11a50cf4d28f9ed3c51658284e921239f740a09eeda1a7e56759f58f1a.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "region_counting/town1_ground_aerial_v2/000038", "task_name": "region_counting", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/town1_ground_aerial_v2/all_questions.json", "question_id": "000038", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [634, 25, 735, 575]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 4\nB: 5\nC: 3\nD: 6\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b93ab76ddb5a25b217c5c906a07d5efb3a35760afac1f9e5f0eaea7e0b6e0003.jpg", "reference_inputs/images/30142c91b4a377fd8bb01708b286c070258fd5c4a84e7771a9bec6c4b1def4c8.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "region_counting/town1_ground_aerial_v2/000040", "task_name": "region_counting", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/region_counting/town1_ground_aerial_v2/all_questions.json", "question_id": "000040", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Region Counting.\nYou will receive two images from the same scene.\nImage 1 is an aerial image and Image 2 is a ground image.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides a ground view of the same scene.\nOnly your aerial view defines the region of interest. Count how many pedestrians are inside that aerial region.\nUse the ground teammate's view only as a supporting view for fine details and cross-view disambiguation.\n\nRegion bbox rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- Count only instances of the requested category that belong to the aerial region.\n- The ground image is a supporting view only; it does not define an additional counting region.\n\nRegion bbox in Image 1 (AERIAL): [544, 242, 993, 545]\nImage 2 (GROUND) has no region bbox and is support-only.\n\nAnswer options:\nA: 4\nB: 2\nC: 5\nD: 3\n\nTask: choose the option that gives the correct count of pedestrians inside your marked aerial region.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4f1d592425f43ddd44ef33551b00627e3bfa2b20112e736d3202b2b1e08e70c9.jpg", "reference_inputs/images/a04ee1aa01447531e8a5ff396360422102cee8a5d040a61a205db0f05731279d.jpg"], "image_roles": ["aerial", "ground"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_1/000001", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_1/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Left quadrant\nB: Back-Left quadrant\nC: Front-Right quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d1208eff96d92a7fd6bccba5b94173bda540487b2d6c9abb2aed9a029c8b04b1.jpg", "reference_inputs/images/5c5d9ce3fecaed967be187543702de327d080f79264b8f4aa5863b2e867d92ae.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_1/000002", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_1/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Back-Left quadrant\nC: Front-Right quadrant\nD: Front-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/083e39a92311d60c58abc7af2d981a90f3151b1fbf1f73fc50b6ce82fe932261.jpg", "reference_inputs/images/5fe4858638359d1fce812039e657dd7d3ccd5de4c28212446506631be8835da4.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_1/000003", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_1/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Front-Right quadrant\nC: Front-Left quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e02043d23ad65725b764174bf547909727ec026ae4a6b803bc0f51f12659b55c.jpg", "reference_inputs/images/06af7c55fade99bd9800af5b542b2c1e94ef5178e0775f19126b4d1964577403.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_1/000005", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_1/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Front-Left quadrant\nC: Front-Right quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/dfcedf4a9305bcdebe114b2c51108a7a6ac1c652f31e7ed419201400eaf161d5.jpg", "reference_inputs/images/fd64f97f5f7d5a29f5e1abce252c06c610b3319860545ef36e4eb366382a8a20.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_1/000006", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_1/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Back-Left quadrant\nC: Front-Right quadrant\nD: Front-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fae9f44fb95c8f7efc9858fb2f75a4ef5045d9a2e7d6fc119236613879110385.jpg", "reference_inputs/images/7021046165264fe95dbc872a235c58d88a9ff32a687e682a273e38caea1d95a6.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_1/000007", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_1/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Back-Left quadrant\nC: Front-Left quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ed8460dd1ff5417f21c97a3879f20e617fcff3dd1d5f011d895aa5c47563e863.jpg", "reference_inputs/images/2bd1cfe8da01b38c94221429e9656bcc5957c6416565b2eab4436e23792320f1.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_1/000008", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_1/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Back-Right quadrant\nC: Front-Right quadrant\nD: Front-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c1382785bd3a34b4cec9f471dd757fef626471a5a030aa412b95802550f8d532.jpg", "reference_inputs/images/b664a6d92fb18e135af0c7b6f4664d241ca1ae9147c72a28f73919f3e639098c.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_1/000015", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_1/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Back-Right quadrant\nC: Back-Left quadrant\nD: Front-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/12a3d5f3299313231ef3dc6175bf524929b0e56b46baede905cfce974e7948fe.jpg", "reference_inputs/images/d2efa8238db62e0302495679308122f84cac70f0dc3a8aa39149ba070f5e634b.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_1/000016", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_1/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Front-Right quadrant\nC: Front-Left quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8bab8798b11b880a8ed8f23e71ffc10e08406810d4a290748e09ad09cb7ff0a2.jpg", "reference_inputs/images/eb8c4dc491391870d622cb06d72f4cee6f9b6ef1b978e617dc48fca20762e929.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_2/000002", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_2/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Back-Right quadrant\nC: Front-Left quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b1603f49a30bd5b1eded6334cbf46e37120b1506aaaec29442927bd674591e69.jpg", "reference_inputs/images/8fd6eb21dcce29017213df4f1473404cf7790cc8dec769bfba752b74d532dfd3.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_2/000003", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_2/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Left quadrant\nB: Front-Right quadrant\nC: Back-Right quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b8e6958ec81c5241f7593e4f084fe99705415f213fa454e7b77e8844e089112b.jpg", "reference_inputs/images/283d0d9f1fe34c8a53ae4a3a4c7ae7200feb75e42df8333f70e47e6e4fdc325c.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_2/000004", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_2/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Front-Left quadrant\nC: Back-Left quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/315e2d7ed4d1f6f7a1f2fc6dd6a30d0a415ab23ec6f38b6763ee92cb492a8e74.jpg", "reference_inputs/images/19a6e992375d16d862bb085bb6130245fda01b9ec73f93ceb50e6dfb125a97e6.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_2/000005", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_2/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Front-Right quadrant\nC: Front-Left quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ec00873859a0160ba51764b448a4624bf37f740f149224a939e21e35380bd2e1.jpg", "reference_inputs/images/51f4dd616196e6f0e5b26db787d18a5bc9770ed27a0a643d86e7d105fd31f113.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_2/000007", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_2/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Left quadrant\nB: Back-Right quadrant\nC: Front-Right quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/41b6568551cd348fbc64aa42f67113dacaa91fc4a6d9ccdc31a4a8d48990e52e.jpg", "reference_inputs/images/d5b5544a2dd06545d1d557f9d196df448d0551f7316f62dcea8eed151c48e548.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_2/000010", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_2/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Back-Left quadrant\nC: Back-Right quadrant\nD: Front-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c2f7afe355caf4b391e26c5de9ce11de5158fa24bb50a7ae061e694281dba7f0.jpg", "reference_inputs/images/8d9549edae29caf52df3f521af3f9727abf57c4e779b5df759f92d4585a2c1f8.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_2/000011", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_2/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Back-Right quadrant\nC: Front-Left quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/110a3fe6bb726292dfd3d7b26ddbb6bdc91b4c553acbb5a9781122e8223ebbe2.jpg", "reference_inputs/images/19494561a7d9b979a0f5bb64abdd07a3ba7a73976a72e52e88241c7cd0d7f232.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_2/000012", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_2/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Left quadrant\nB: Back-Right quadrant\nC: Front-Right quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/88dc33864b541ddc6ea3fb4ec63c9d6ea22a16381146d08b7f6f5585b21111d9.jpg", "reference_inputs/images/045eda90d638e1a113b48f37d7ea682b5396a45a6489ff794db69a93cd32cbdc.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_2/000013", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_2/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Front-Left quadrant\nC: Back-Right quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9bfc3480727431557dc195aba924168c088cde93df95cfb7985d2693ab1e966b.jpg", "reference_inputs/images/b59b28fee787f5d48fae8555d23a10805e2a409a7b57fb65242a39619bff4977.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_2/000014", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_2/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Back-Right quadrant\nC: Front-Left quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fee7f5e9d0f1da57f2d0c7de34ee8ab7d03a06ace3c2eed48a8e3b40b4a57642.jpg", "reference_inputs/images/380aa0858fc581d535c07ebe933c752c68a90f0d0742715fc0ad2a9db33b62d4.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_2/000015", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_2/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Back-Left quadrant\nC: Front-Left quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/14e30008acf7df46593fc02231bae5300ca56049f483163a99cd01e6cea82532.jpg", "reference_inputs/images/0419fa0fbd205fc87f4c4688ab8d6b1d4c62444f3d40e75a1ac9c59106942716.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_2/000016", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_2/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Front-Left quadrant\nC: Back-Left quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2970d2e2b9a931b838932c39f3510c0cd465670b43c6b233db659c61ad34646c.jpg", "reference_inputs/images/1457808955f943f6446d544b858661e4f00f447e58306a56fdfd29235180729f.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_2/000017", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_2/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Front-Left quadrant\nC: Front-Right quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fb0b9ac5b78316924f8bca5303a851859418d2626d988586b812ceac5e38ec67.jpg", "reference_inputs/images/1457808955f943f6446d544b858661e4f00f447e58306a56fdfd29235180729f.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_2/000018", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_2/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Front-Right quadrant\nC: Front-Left quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/aa6489728460099af47af6afff42c3894d73005d2db0bb3d1680955ebfe1390a.jpg", "reference_inputs/images/e07cdc1179d3b54efeb5b78e5c894a9deda9af6624f8fd01d92477dfcc36e38c.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v1_2/000019", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v1_2/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Front-Right quadrant\nC: Back-Left quadrant\nD: Front-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8db7f03ca37a04c2a02a21817b0d36b708770a1fcaf3c003e6b4e05ae15024b3.jpg", "reference_inputs/images/115ec48cc3877f150e281ac666d8691aed8efbfe1a6c987383742a47fe30c5e3.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_20m/000001", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Back-Left quadrant\nC: Front-Left quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c54d445fb5e69a25def2f0786580f0763b9f0f5d2de55b184c79c61b404f6c1c.jpg", "reference_inputs/images/0b04d2c048d9ec3220836158a321c7d88654a550351b69a7abad236dadc214fa.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_20m/000004", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Left quadrant\nB: Back-Left quadrant\nC: Back-Right quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a139b18a0930e5790b5b31b3e7700838330e2618c4baff3c41395a962f8daabb.jpg", "reference_inputs/images/5da26493d4c611fa8d649289168ac24a082faf6d52edaf57674aa801449571b4.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_20m/000005", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Front-Left quadrant\nC: Front-Right quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/94d4ae3880d921e2f1eab3153ffa9a6fe961d71d8cdbdc06b7ddd8bb6ee67a7d.jpg", "reference_inputs/images/d12375d5e82700b64fab8fe32bb0439c4988961f226f86a61debff37c8aae019.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_20m/000006", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Front-Right quadrant\nC: Back-Left quadrant\nD: Front-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f6f7dfae7acf0bf8d89bb83502bfcf1aadd2ec88d65a58fc8abd51f81f6f6c81.jpg", "reference_inputs/images/1d9bb9aaa8c8be55cb7b7063097cf9ea5a010b9ee47773b5649721b27c4b123e.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_20m/000007", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Front-Left quadrant\nC: Front-Right quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0f943d81caba41056da75cb6bea0f8698a19e3a1813962597bc0e1446c8319e3.jpg", "reference_inputs/images/daa4685c900c7485f0b3274935690a92a3a0e3d43a08f6d210f5627453dcfb6d.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_20m/000008", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Back-Left quadrant\nC: Front-Right quadrant\nD: Front-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/07231de9b22c3cbfbe8075b98376d20ed20de251cda1dcb875ac218dbad58435.jpg", "reference_inputs/images/326e309493b74656ab397a5c7fe4a7d782f7c238e5d6a88d3395323169ed9146.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_20m/000009", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Left quadrant\nB: Back-Left quadrant\nC: Front-Right quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/37785d8da58dbeab5ad30f20987c6606a8d765aa162c84b695f22e8745863b8b.jpg", "reference_inputs/images/16652bd8a0749e2b8903fd22753cff94fa7d4fe455aeb70555cc2ffe81a55955.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_20m/000010", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Front-Left quadrant\nC: Back-Right quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f3304d8f6c30eba680b8003db8951429c0a08af11730147b1d38f431a8a96da4.jpg", "reference_inputs/images/6f0f1e07119f53abe8f08c28c43962d83b3debe85de2c64a6e89e21ebb95a02c.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_20m/000011", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Front-Left quadrant\nC: Back-Left quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/33332c57beff12b8e96f081e37e79ee8ec31d7c4f05ee036a50517fa45de83cb.jpg", "reference_inputs/images/b3574f508469643224817c0968375f2dbc9b8fb5695e469f277f5de5349a47a5.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_30m/000001", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Left quadrant\nB: Back-Left quadrant\nC: Back-Right quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b4aeabd301c5f0d67e60cbba9d229fc31560671eef0a7530427bef5e9fbf057a.jpg", "reference_inputs/images/690c4ff142b67551d17093b15fc4457e32eec0e742c732f44d880ff63a680016.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_30m/000002", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Left quadrant\nB: Front-Right quadrant\nC: Back-Left quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a912937c2bbc973486857363c2d00ec02c731d2ba00134008da9fa6b7c614af1.jpg", "reference_inputs/images/912001ea44deb2e90793bbecd6ab0407711e1c2fc6b14c92a7fff3b001b742e5.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_30m/000004", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Back-Right quadrant\nC: Front-Left quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fbe108ba4579f03b08727eeccadc3a522d76afdfa5f2531274a23402fe21f052.jpg", "reference_inputs/images/3f6b8fcca774c8e8aea257100830169913e8608cc0bcabc19239f3bc3545c962.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_30m/000005", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Back-Left quadrant\nC: Front-Right quadrant\nD: Front-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f10fb2e0d88e96fb338a1e165737e9ea83a6574149226476d2879e83d649efb6.jpg", "reference_inputs/images/a48a18e7653d72be7bef71d38f63063825f2392195ab9c9f2863e8d057499145.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_30m/000006", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Left quadrant\nB: Back-Right quadrant\nC: Front-Right quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a99089a90f035a0cf0fa4e05a2c2f5f93f8f3f3b89f3511a1059ce48ba4c3f85.jpg", "reference_inputs/images/29fc872b749e944ac734776f6892fe8e4600cf609cbd0a71f1ff2dda5f3dc4e4.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_30m/000007", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Left quadrant\nB: Front-Right quadrant\nC: Back-Right quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9330cc6d6a3e3ecb4e8d92b126cda21e3be3b60f171c1a2b51839e2e69f8daab.jpg", "reference_inputs/images/0b04d2c048d9ec3220836158a321c7d88654a550351b69a7abad236dadc214fa.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_30m/000008", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Back-Right quadrant\nC: Front-Left quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/03df6e1c982636742536da0f7c383c2d3a61b4df8c8d42db9327d61372254e10.jpg", "reference_inputs/images/c5b7c80344326eadf236c095b74caab02fec7f079c683b63408b21f0a4fb239d.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_30m/000009", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Front-Right quadrant\nC: Front-Left quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/41272714c461c64948b95c34de253fc336faf2e87c94db7c4fd67f3f8310187f.jpg", "reference_inputs/images/bf6e1f1858c88fb977f0dbc32936d9c13cdd676f51a1e07b64daf08ce71bb42b.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_30m/000010", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Back-Right quadrant\nC: Front-Right quadrant\nD: Front-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ee41c89ff2535667868fd53b4470854018bd7a47604631637b1a47bde327eb34.jpg", "reference_inputs/images/16652bd8a0749e2b8903fd22753cff94fa7d4fe455aeb70555cc2ffe81a55955.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_30m/000011", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Front-Right quadrant\nC: Front-Left quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/270fa9464b163556dc9aa067570bc16cf9433517ed4e73dcf29334333df7bee8.jpg", "reference_inputs/images/29fc872b749e944ac734776f6892fe8e4600cf609cbd0a71f1ff2dda5f3dc4e4.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_30m/000012", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Back-Right quadrant\nC: Front-Right quadrant\nD: Front-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/123b16680ca4edf9a33930035e9277d581676b9ff8dcec696fa7b58dd1908695.jpg", "reference_inputs/images/690c4ff142b67551d17093b15fc4457e32eec0e742c732f44d880ff63a680016.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_30m/000013", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Front-Right quadrant\nC: Front-Left quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7c3a0730ab00fd3d1f07480f82d0f2dc67c7faca9b5254d862dc8d637fadd99d.jpg", "reference_inputs/images/16652bd8a0749e2b8903fd22753cff94fa7d4fe455aeb70555cc2ffe81a55955.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_30m/000014", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Left quadrant\nB: Back-Right quadrant\nC: Back-Left quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9828f6bc92cdae868b25124e5f4081fb2d583d9e4b282caab3ad1fab458d4e1b.jpg", "reference_inputs/images/09369189902886ec50eef896e38315dca10ecad2ab106483ab70da4efe8e2992.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_30m/000015", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Front-Left quadrant\nC: Back-Left quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/38c62d125b0a8fb8e40253d75b8bb210a5cad7b17077d0b36ea32743e1272175.jpg", "reference_inputs/images/d12375d5e82700b64fab8fe32bb0439c4988961f226f86a61debff37c8aae019.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_30m/000016", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Left quadrant\nB: Back-Left quadrant\nC: Back-Right quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1a59c413be262310ff5d15b65c52075feadb3a0ae730ae1c3bd4fbe142e23ded.jpg", "reference_inputs/images/0d248986a749fc13bbbbd7f16f564db05fc709e1a398f8b7f179f2517a96214c.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_30m/000017", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Left quadrant\nB: Back-Right quadrant\nC: Front-Right quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4d89a929368c52937b2de84eb242f01ae3780c1b790e0c71584dffad39de9dc2.jpg", "reference_inputs/images/8608d7fa0052e50a249d9eecea803abfe1084392302cdee1d25b4e9a4628e6c0.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_30m/000018", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Front-Left quadrant\nC: Front-Right quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/979c390503f8aff58aeb976f809cea3d39b2cc9e001c415a37f9c7011a979b66.jpg", "reference_inputs/images/6f0f1e07119f53abe8f08c28c43962d83b3debe85de2c64a6e89e21ebb95a02c.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_50m/000001", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Back-Right quadrant\nC: Front-Left quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4f978865c4e267fd1c053d8c0510ef383c84895018a58773dee358ae55dc1819.jpg", "reference_inputs/images/be2b0ac59b8c52f200d97194ede610aa1ec0fd0d2c3d5e6dbdd550689f004515.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_50m/000002", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Front-Left quadrant\nC: Back-Left quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f102f31a8a16aac2b3b5ee9349f10343a6e6b60b7205af6d2a51c22bd17fa017.jpg", "reference_inputs/images/a48a18e7653d72be7bef71d38f63063825f2392195ab9c9f2863e8d057499145.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_50m/000003", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Front-Right quadrant\nC: Front-Left quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/80328c294651e1f194fc273b372b0c27b09ba067dfb545b3bb42c01774b64ef4.jpg", "reference_inputs/images/6f0f1e07119f53abe8f08c28c43962d83b3debe85de2c64a6e89e21ebb95a02c.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_50m/000004", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Back-Left quadrant\nC: Front-Left quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c55a3a99b02d33b6ac2aab379cf9512d06573d25f1e7cf563994550604627033.jpg", "reference_inputs/images/1d5f9dec7756bf1faa672e44115577f7b1d0a59014a9a88e1d52979f20ad2a82.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_50m/000005", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Back-Right quadrant\nC: Front-Left quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b0cb88fad061da22585030cbd89c5177a9daee378c3fca13526dbdba1361b784.jpg", "reference_inputs/images/6c5c03a35e0c4e5b9e2856dcedf7645980702336ab76709ea801dcffd18ca8b6.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_50m/000006", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Front-Right quadrant\nC: Back-Left quadrant\nD: Front-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f5e6eeecb068d3f6daed89a67d62d3c82df260d065bff871392e5c4e72cd34f1.jpg", "reference_inputs/images/93f8c8c3cf249ce2b1f1c01c90e4fd3362d4265bc084ad006f6dc4e0f549605f.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_50m/000007", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Back-Right quadrant\nC: Front-Left quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8ae82511f963c74af712f2d260f1a49a1300ec30b06cabcf02c802c983f272fc.jpg", "reference_inputs/images/0835003194a0b3507fa1fa9fd6fc3b9c6824642b011ec2703d1fb3540b958ca0.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_50m/000008", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Front-Left quadrant\nC: Back-Right quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/695f31f6256253c5b473376e64ddd51ae0cf4dbf2bbe77553e9e8e52f992e754.jpg", "reference_inputs/images/c5b7c80344326eadf236c095b74caab02fec7f079c683b63408b21f0a4fb239d.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_50m/000009", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Back-Left quadrant\nC: Front-Left quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ab9b799b03bf0718dd0ec1fd77845f3673183da68faa90ad4a1af406f37803d3.jpg", "reference_inputs/images/5b6ac385e3afda8add29752aadc7568dcfefbad224d4b7851478badc13cf2acf.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_50m/000010", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Left quadrant\nB: Back-Left quadrant\nC: Back-Right quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/90cedb64de2f2ae81a0a5ff39469787009b0752efc2329de7a31ee311f879621.jpg", "reference_inputs/images/14d00662632d88a16a41db70141f674d478f7fe0f9e54aca5f487a826653bc06.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_50m/000011", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Front-Right quadrant\nC: Front-Left quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b44343ec2acd390d47383baf669b3f0b3f7c183baeb4f08a2ec1e48353e36758.jpg", "reference_inputs/images/74e818dd8ccefda5cdf4d1ea88f33d95a27373669fe205b437f32b3db295ef87.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_50m/000012", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Front-Left quadrant\nC: Back-Left quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b4d2eb3d3905bd50b6234893d2e6a88ea02c3bee42d72abc7bbe8183d162194c.jpg", "reference_inputs/images/c54c29820fccc930f14033600affd66e73aaf2bf3b2687b9314b0a6066c1e607.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_50m/000013", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Front-Left quadrant\nC: Back-Right quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cca1201208052f721a6fc45bd18189e3f61bbe9cd2220b28b887ccfe69d63c7b.jpg", "reference_inputs/images/c5b7c80344326eadf236c095b74caab02fec7f079c683b63408b21f0a4fb239d.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_50m/000014", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Front-Left quadrant\nC: Back-Left quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/98c095a643aa85e292def4d94386cec12cc0a31eb14fbbeb792f1ff8be98db7f.jpg", "reference_inputs/images/3113e6aed27291d0d5e31774ea7fa14fe07171180a04d9447efad9dbd3662ab3.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/city06_ground_aerial_v2_50m/000015", "task_name": "relative_orientation_determination", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Back-Right quadrant\nC: Front-Left quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0e2648bdfe4af03f9e673cda603f6f88a1c15b42cfa7ee3a4377bb4ff8914bd8.jpg", "reference_inputs/images/d12375d5e82700b64fab8fe32bb0439c4988961f226f86a61debff37c8aae019.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000001", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Back-Right quadrant\nC: Front-Left quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/584ec3d7908ea762ad1dbc85abd748cf9c0590c712de59d50060a256e23c9c90.jpg", "reference_inputs/images/f4d708703f37d41d902dbc76f9f05ff0026ab1f3fecd01e1bfa0c2a0f17b5f2e.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000002", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Front-Right quadrant\nC: Back-Left quadrant\nD: Front-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/40b599020b3ae1ea5bd607835f5bdce9e4e8b9acea4a77c40594f8de70a54a47.jpg", "reference_inputs/images/90ce09656a6a167da5f7fab284aee25ea2185e3afc014d9222e307ef4bb41141.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000003", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Front-Left quadrant\nC: Front-Right quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/145c653f95d8dc4bf7d92707ed873b1535b868da88e27367234a892adf211445.jpg", "reference_inputs/images/2eca17fafe3337c779f21f0ae44babda73246fb323ef84c7e14f9337a5c9e108.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000004", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Back-Left quadrant\nC: Front-Left quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7770f6d8171d83d58b5080a8da0226da0f94df2f63a214beaf6eb04b31184606.jpg", "reference_inputs/images/8a27f6f27475bd1f64e42fd6b54dfb6a99b3e314601ab91db35ee291f67bcc1f.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000005", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Front-Right quadrant\nC: Back-Left quadrant\nD: Front-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5fecc50cbf1474e1d5fa645c1487ba47767cf5139af73b288b65f3b63220bcf1.jpg", "reference_inputs/images/165fbcae599cf7a3a388132616855423a6c4a51f382b94071614dded67ca684d.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000006", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Left quadrant\nB: Back-Right quadrant\nC: Front-Right quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9b957e5691c42b20971ec00dfb3a897ec36cb2bc3ecec0bc03f0682dd77c9cce.jpg", "reference_inputs/images/8e8c6870887d16e22edd66002d5c430f1199eb16a4d03e811e55837409f737a7.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000007", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Front-Left quadrant\nC: Back-Right quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/12c473f48adea1c252985951dadc80ecb2ff5eef76564d64997f4c4016e196b2.jpg", "reference_inputs/images/8e8c6870887d16e22edd66002d5c430f1199eb16a4d03e811e55837409f737a7.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000008", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Front-Left quadrant\nC: Back-Right quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/58a814ac4dad306af52928dda756605c566fd10375762ce26dc229da95c2dc94.jpg", "reference_inputs/images/3e64f8f2b4008e4ea227a893b7986c74745672835ad8a09898a2f36778877b0c.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000009", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Left quadrant\nB: Back-Left quadrant\nC: Back-Right quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/234d17ce25c803b4a1fe34833f8c3d36cdfc63c9f1247cd5062705aa5eb18ce6.jpg", "reference_inputs/images/614bea61ac9a60f84dd3186ff090ac89c59c5f19c19daa82db3fc2053a6cd2cb.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000010", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Left quadrant\nB: Front-Right quadrant\nC: Back-Right quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/69133cce3ec680b40721f9c8c905ae1a007c0f6397a7ce1d8a3c10bcd329d5e9.jpg", "reference_inputs/images/a75e375d0eb6d4151b86fe3e505b0c8d1603b45502213337387d22a1a023ad95.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000011", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Back-Right quadrant\nC: Front-Right quadrant\nD: Front-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/527a2737d0f6f1ec32402f1d3385342b729d1cc018677d7db2b50b3025317b85.jpg", "reference_inputs/images/0b1a37a656d3fb6c2b0b1eedc6c8bb0527fef397451d75847ff913fc5e5dcd54.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000012", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Front-Left quadrant\nC: Back-Left quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/99ca249a24ae3ecaa8fbd28081b26086c6bccf9d9b52f4fcb2a9ac1f25833b50.jpg", "reference_inputs/images/e8babd8ce36edca7bdab176a21640e321a9f203f01b37bc12897140c0e80f5aa.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000013", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Back-Left quadrant\nC: Front-Left quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/84e6f68737db5029764ff6f616abbf23de2df12d95afa938871f2dd757971838.jpg", "reference_inputs/images/e6a930cc13c9435d48975c505cbff4ed5d6ffbc32bde3d87755e13d702c854ea.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000014", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Back-Left quadrant\nC: Front-Left quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f051102fbf0d94e573784db2f65335add50453a7b7c660f73dc6e6c7a020ad0a.jpg", "reference_inputs/images/696db644fc0e650348f4be016e8e5fd9c181eee2de39d8b0f2f71ba4a8fde652.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000015", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Left quadrant\nB: Front-Right quadrant\nC: Back-Left quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9bf4dcea088a9e3e347fb63efe801a77ee84e43f4f021c0eef1aa4cc99c64ac8.jpg", "reference_inputs/images/2c57f3e448d39cce4b09504a4ae1620a90cfaa78e110b78269db0fbd1d03ddb7.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000018", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Back-Left quadrant\nC: Back-Right quadrant\nD: Front-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d26100ebb0f93747b427edda57717dfd3da447ab40f3021ccc6cd4f7a92aece7.jpg", "reference_inputs/images/ae001cd42d76e6f87147dd08d6efa1f9ee182e555926abab696efe15c5f62a73.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000019", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Left quadrant\nB: Back-Right quadrant\nC: Front-Right quadrant\nD: Back-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/75a553fa686517a5743971fff15f461c1d520e59a83ccd39ce28d69eb97e3f8a.jpg", "reference_inputs/images/604b0bba6fab3a8e838fe265fe169c465a082d5863abfe365e17550612bb0f6d.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000020", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Back-Right quadrant\nC: Back-Left quadrant\nD: Front-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/13e9c37b35c87b6cb7d52561c413be8ef772eb570076d53eec18223054dacb7e.jpg", "reference_inputs/images/2b9ebe127de558219779e74edbe2262829c28ac14fa229591b8245e5b1109e5c.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000021", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Front-Right quadrant\nC: Back-Right quadrant\nD: Front-Left quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bf8d4e4953ce2860e08cb0e6a129d8ca9aa535b345df5227dc0192b12141b89e.jpg", "reference_inputs/images/d4ce21fb471bc9ebf9528a2e13052e2a567ca745976ee2b1323e8a89a9407f72.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000022", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Front-Left quadrant\nC: Back-Left quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/876c47d36d8170077ac58dff84a9270974f193e74e3a0933aafe8ce451c2019d.jpg", "reference_inputs/images/66bfe999816257280380ed7e222809ae0590b3b808eed6a01afee544df594ec3.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000023", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Front-Left quadrant\nC: Back-Left quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1c3c5d3d9665a2f9727cb9dde833f75d59c69e9d6303cfee1224c98ffb6c52d2.jpg", "reference_inputs/images/56fdec6518e6fd442d5bf50715cf842df5618b6e7125f80e58d99b5f95fb2643.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000024", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000024", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Front-Right quadrant\nB: Back-Left quadrant\nC: Front-Left quadrant\nD: Back-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/52e2252188b9f626a02ceb9fe89d5c261988d7a470c79cfabd648d57f8858b64.jpg", "reference_inputs/images/a08ba365f1665b5e5328292bcae275b989203d18db9bfb02c198e3b340b49c5f.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000025", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Right quadrant\nB: Back-Left quadrant\nC: Front-Left quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/796367e3cd1af1690ef176ba0bf4ff51b9ed44797c42c23765beeed3e862f2af.jpg", "reference_inputs/images/3e64f8f2b4008e4ea227a893b7986c74745672835ad8a09898a2f36778877b0c.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000026", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Front-Left quadrant\nC: Back-Right quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a12102b1ff2167605667085bfecc02f8e96909e05a6369437f1a3d2d029bf8dd.jpg", "reference_inputs/images/f91f0b62ae01fb1ee46bf353137e81238e6485d743493d53a8a731cbd22caeb7.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "relative_orientation_determination/Town1_ground_aerial_v2/000028", "task_name": "relative_orientation_determination", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/relative_orientation_determination/Town1_ground_aerial_v2/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Relative Orientation Determination.\nYou will receive exactly two images from the same scene.\nImage 1 is the AERIAL image (aerial view). It marks the ground observer's position with label G.\nImage 2 is the GROUND image (ground view).\n\nCoordinate convention:\n- Use the ground agent's egocentric horizontal frame as the reference.\n- FRONT-RIGHT means between straight ahead and the agent's right-hand side.\n- FRONT-LEFT means between straight ahead and the agent's left-hand side.\n- BACK-LEFT means between directly behind and the agent's left-hand side.\n- BACK-RIGHT means between directly behind and the agent's right-hand side.\n- Each option covers one 90-degree quadrant, and the four options are mutually exclusive.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate provides the aerial view, where your ground position is visibly marked with G. Determine which egocentric diagonal sector the aerial view falls into relative to your current facing direction. Use your own ground-facing heading as the reference: FRONT-RIGHT means between straight ahead and your right-hand side, FRONT-LEFT means between straight ahead and your left-hand side, BACK-LEFT means between behind you and your left-hand side, and BACK-RIGHT means between behind you and your right-hand side. Each option covers one 90-degree quadrant. Answer with only one option letter from A, B, C, D.\n\nAnswer options:\nA: Back-Left quadrant\nB: Back-Right quadrant\nC: Front-Left quadrant\nD: Front-Right quadrant\n\nTask: use Image 1 together with the marked point G and the current ground view in Image 2 to determine in which relative direction sector the aerial view is located from the ground agent's current viewpoint.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b345f14e387696f7f5004a9b8258d787bb5883d6b0fe840b569d6eee7107f2dc.jpg", "reference_inputs/images/9e940adea7c32759f0c95212bcc83902fac8ef5f8cc4fef74e0d153479f1ec3c.jpg"], "image_roles": ["source_with_position_marker", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_20m/000001", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G2\nB: G5\nC: G6\nD: G4\nE: G3\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0baaf3ecbf3387effb96b1ea4e0930ecd7670c5a2a576c179aee7f0ae8dc1408.jpg", "reference_inputs/images/3e0a4ac2e88225b48e8c398702ff8943834354639cb6fce673aafc0cc14fd71e.jpg", "reference_inputs/images/ab664e3a1c1d9879f2ca1d50406f4e89dff8c209dd7abe815855c9c49daa0388.jpg", "reference_inputs/images/772ae829aab097e813edc986df6337aa53a4dfb46188b630cff56caf74228b46.jpg", "reference_inputs/images/c5b600c09a6d37cd46165ccf13dda6e2f27b7d463018639951817900fbdef907.jpg", "reference_inputs/images/165cc78f38c0f249eb17768c4ed0e3d7f315364b213c61a5d41731ef443b90ce.jpg", "reference_inputs/images/98e184ef25113e99d3acb81d2b999b2dc5d3a8d18056ae3ea6167479bc81beba.jpg", "reference_inputs/images/136f851f4ce6fede76cc50fdc3fc6048d01d9a5ec12763732777c13fbb606be9.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "F", "answer_index": 5}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_20m/000004", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G3\nB: G6\nC: G4\nD: G2\nE: G7\nF: G5\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/78db7c947d8e568717113dae586f632b0a69ff8e5ae8808b22b8fc9f81675f40.jpg", "reference_inputs/images/67edc9f39ba4f2605ef24a73488e1d0298525c61eabf99e0c8b5b8008875bf37.jpg", "reference_inputs/images/006e4b97e40fb9fb2f043974ca480bf6e4fbbcfac6f82f579c3ab98b45c6c1f4.jpg", "reference_inputs/images/136f851f4ce6fede76cc50fdc3fc6048d01d9a5ec12763732777c13fbb606be9.jpg", "reference_inputs/images/fd1a741907a15ac959be34c50c789abef4ee72af9a0929d747d349ffbe6f6efb.jpg", "reference_inputs/images/98e184ef25113e99d3acb81d2b999b2dc5d3a8d18056ae3ea6167479bc81beba.jpg", "reference_inputs/images/0712923cb9ef7bc88dbc73ca5969dd9ffecc54a415e0f4d2e3a28e04b7f8a7a7.jpg", "reference_inputs/images/61999a50526ca2dc081cd3dadbaaa02c11884c86f17f6719f4b4a68f8a6f67f8.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "F", "answer_index": 5}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_20m/000006", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G5\nB: G3\nC: G4\nD: G2\nE: G6\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/aebe4c70ee5dd8afd846180b73cd48c8390d049e87e837b5a96e973a9e9d1aa7.jpg", "reference_inputs/images/165cc78f38c0f249eb17768c4ed0e3d7f315364b213c61a5d41731ef443b90ce.jpg", "reference_inputs/images/98e184ef25113e99d3acb81d2b999b2dc5d3a8d18056ae3ea6167479bc81beba.jpg", "reference_inputs/images/3dcc3b6ab37085f8ee320458b61a7cf7f2eab3f312e6f84ddc2d277d5a81c266.jpg", "reference_inputs/images/affe958434ac9600d59e4e439ed10d5a41e9ed75ffaf1bb9084b9fe73c2ace8c.jpg", "reference_inputs/images/f48461b9bd7b9a45c46d40f6e514e1c95595739c2f5b3a83e72c20e56d0569b0.jpg", "reference_inputs/images/ab664e3a1c1d9879f2ca1d50406f4e89dff8c209dd7abe815855c9c49daa0388.jpg", "reference_inputs/images/a9890d2fc6825dfa001f6cd4b2a99c35f9b6812d292bc7dd7e49d59b19403a0f.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "E", "answer_index": 4}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_20m/000007", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G5\nB: G2\nC: G3\nD: G4\nE: G6\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2ac5dfcc71fa773e93e93905ce64f4e8ae446c1c3a23d9b20029b4baf6fb4789.jpg", "reference_inputs/images/a13aec45195756770182b27499d472f47b160ff507dee6ba042827d5c22fdc1c.jpg", "reference_inputs/images/0a637c542cd55efcd7db57209cb296cd5b3b8122a25a1081e21a75d2467f1eec.jpg", "reference_inputs/images/a37278d6f61a09212963b84767982a4ff7065abe6dae528d356c579d594d5b87.jpg", "reference_inputs/images/daa4685c900c7485f0b3274935690a92a3a0e3d43a08f6d210f5627453dcfb6d.jpg", "reference_inputs/images/81046b24d8b0d3e578bdfbf2026c08137a814732d76eb9b298a0b6178fe200d5.jpg", "reference_inputs/images/f53561ce652da8ac3b7e5f0f6c6de9c351851774d487ef2584080a2549e351c1.jpg", "reference_inputs/images/89c89dabbd00616996b427ac6caa0c2a050066838db1202e7f3e682f2365a2a7.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "A", "answer_index": 0}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_20m/000009", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G6\nB: G4\nC: G5\nD: G3\nE: G7\nF: G2\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/660bb86ede4c0d7544b7b733a0c94cfede5228147124150e5f071cd447054636.jpg", "reference_inputs/images/c0d2a3e08cf989cab00b16bae02eabdb9d196be80eb7ba851f154ec5d4a68f11.jpg", "reference_inputs/images/ed2189f53b3f6d1c05bceb524c0005a7508a1ff0431f3713d530ae623c5b2642.jpg", "reference_inputs/images/34d027dbab11397f568f524e93a1358bb23b7502da274b0c04a1dba07548b837.jpg", "reference_inputs/images/a4ff943138300e14df5dde6b5b4d48e1550533ac77ffe671ed28e02e96353024.jpg", "reference_inputs/images/0183cfd2866b057e06464418777bcf14316141f2c7c9027b08baf55273d48d83.jpg", "reference_inputs/images/37fdf1cfa1885a50dc1e3a45f4b4196e30f9e8011ea8b00b2aa25e97696ee22d.jpg", "reference_inputs/images/bb5fb5ea9e5337cd8439330bc758a7b429296eb13e3a974b49d53f06abdc07c5.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "B", "answer_index": 1}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_20m/000010", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G3\nB: G6\nC: G7\nD: G4\nE: G5\nF: G2\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b426e5307910180039abb0cb3e4d24ce9134cc3199644eed9314e9c7f8ecc614.jpg", "reference_inputs/images/f2449548a2fb5afee173a59713a49688fa35c32ab2d5843906f2f9ba6baa0112.jpg", "reference_inputs/images/5940e2d6e51e07fc71ca2c19e8d33a5a3ee00569378163517bc52e26913430ab.jpg", "reference_inputs/images/2a4ced4ecbda3f0ad9bfde9034be5e0a49ec9bab9c712884affd3ad08ecb94fb.jpg", "reference_inputs/images/1920d160bf8c207cc2cfe03c2f54780add45804d245793d2333db80413b5314a.jpg", "reference_inputs/images/326d0df537ddef5c1539b9ddbad502e4b450bb1d6abf8df538cadcd6be629d0a.jpg", "reference_inputs/images/21f2c320c1e570119f3c51a4276d427d1289f16ef62a1d359f7a1f4beaf17ca3.jpg", "reference_inputs/images/9e2e596d7969171bb308d5bcb2eac235b9f3cff7b6367317b01383788740ce13.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "F", "answer_index": 5}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_20m/000011", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G3\nB: G4\nC: G7\nD: G2\nE: G6\nF: G5\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d996f4427282f7ffa78a7e85e2cdd2962820a6d96aa3fbdfe26cdaf777b381cf.jpg", "reference_inputs/images/ad26a1f82d612f9596c6e104842d9b22eb76e19fdd9cd98173f1cffb07c5df33.jpg", "reference_inputs/images/02ee26196f052db1fd6094389ca8a203753d4dfdd0196d5aad127f63ffb772e3.jpg", "reference_inputs/images/606dd5fea012af2756b373ee21b6a9f4a7ff6eca37a4eb356fb07f1cb3476f3e.jpg", "reference_inputs/images/5fd84a272bd50d7760735ef75150f5f8cfb3012b7bd945c024e8f2b370da7d2c.jpg", "reference_inputs/images/b3e61a6479831944e0b4dbdb7dcc29239bd5d34ca3e7660f572cd61e6b90bf2d.jpg", "reference_inputs/images/ea0e3e9b4c409129e963d807558d1d1ac6dc3b2990952ffdf51820bee3d05583.jpg", "reference_inputs/images/90fe445bc2bbef05bb93f323c40841595e00c8f66f329dcfb2a85a5bd684128d.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "C", "answer_index": 2}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_20m/000013", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G4\nB: G5\nC: G2\nD: G6\nE: G7\nF: G3\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2cbf409de770c1b16bacf2ec33c37117c02b8d9879094001d8232e44c97ae48c.jpg", "reference_inputs/images/bb1234fc75db8b2004f6e5dad03553677922756abe73a2bc76ac3db2608e9339.jpg", "reference_inputs/images/b01bcd72038777584cc4a4cc511b5af23e6bd0cf19996a6ed670e78ddfb4a5e1.jpg", "reference_inputs/images/f1cd24ff6d8c28314e77e586a6057bbb97f01a83e0555e13b0911369b7e423eb.jpg", "reference_inputs/images/861fd07722f02e7775960ebcbb39ddd697e58e8a2aa057c62f68cc8e41bceecf.jpg", "reference_inputs/images/dbdbafd1873f8374fc2da3b5dc8e39767781e722be8ed25fdcf2da41eac4a224.jpg", "reference_inputs/images/5a50298408b5a44380b325f06a63c9e6f92a85c81613f17465ad2b8e5b4836a3.jpg", "reference_inputs/images/e74ac6cd045c319af1517ac285faf98c10db9e03aab35bba1c6aae874a49105a.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "E", "answer_index": 4}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_20m/000019", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G2\nB: G5\nC: G3\nD: G4\nE: G7\nF: G6\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3733b401bda7f702b8fecaf78a49d9efce6547aeaa39645d24684179f0d059d9.jpg", "reference_inputs/images/bcdc1e69c0a36fdd6f81b6dc138bac8948640e1f401b454a6715a6941bd5a4a5.jpg", "reference_inputs/images/c0d2a3e08cf989cab00b16bae02eabdb9d196be80eb7ba851f154ec5d4a68f11.jpg", "reference_inputs/images/ed2189f53b3f6d1c05bceb524c0005a7508a1ff0431f3713d530ae623c5b2642.jpg", "reference_inputs/images/34d027dbab11397f568f524e93a1358bb23b7502da274b0c04a1dba07548b837.jpg", "reference_inputs/images/d58c9ae08eef29500a5bf2870e2b136cf317de0992e07d5be2a8fc5ab35b65a2.jpg", "reference_inputs/images/0183cfd2866b057e06464418777bcf14316141f2c7c9027b08baf55273d48d83.jpg", "reference_inputs/images/37fdf1cfa1885a50dc1e3a45f4b4196e30f9e8011ea8b00b2aa25e97696ee22d.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "B", "answer_index": 1}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_20m/000026", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G7\nB: G2\nC: G6\nD: G3\nE: G5\nF: G4\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9e543a5df4ac8e7232901b549642bd61d682ee6a85e225241b1e0fc6847e03b5.jpg", "reference_inputs/images/75bb9f3ac1921d7fed1cce6bc9bbfa9956c3988c4fcf366d0c757f6c48b81399.jpg", "reference_inputs/images/50a8652b77bf5cbed4b0cd0bb8bab9494f07e35a1602248d716e193cd683d51a.jpg", "reference_inputs/images/00a29879f7698025a5ffb5abd742219b52f130a6c1dbfc47820d2544470eddf8.jpg", "reference_inputs/images/f3d0a4d31d16c204ff9a76a17bcb947c6cb4b7bbb5cf7f59f6e1d7ad3080d41a.jpg", "reference_inputs/images/6a7f9e051b900b178f7d7c6eefc92cdd3bb4cf1baf907fdc77313333c8ba0243.jpg", "reference_inputs/images/121ca0a666b27ee44d92209d4b2bcecda33df73f1f6332b6dff01271e6619a5e.jpg", "reference_inputs/images/81046b24d8b0d3e578bdfbf2026c08137a814732d76eb9b298a0b6178fe200d5.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "E", "answer_index": 4}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_20m/000029", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G5\nB: G6\nC: G4\nD: G2\nE: G3\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f2622b920ee132fb165ac31e2606d19c55971ea5db4796ce2f470608aaba1c6c.jpg", "reference_inputs/images/e8f86a144ff07fab1b6ed4595caeffbf95d4782f342c62d61c804499bb0065bf.jpg", "reference_inputs/images/199b8fbe4a152e9d5ee803958dac2f67a88f2cf0659fdff516308206c27adca2.jpg", "reference_inputs/images/7f94190665dd70f7d3a511d556daa5af6ea4f449ea5a64eaf2788cf6931abf62.jpg", "reference_inputs/images/2c4d8a2bd0304a00d3fcdc28715ad25aedb7c02628a93f4a466fc02ebe7ae0fe.jpg", "reference_inputs/images/64ecb14e96931fa12d63d7d52e97e6551d9ab8794bec873526ad42f1cf64fd39.jpg", "reference_inputs/images/31c6abcf7eb4aa46f118f145da3d0ff14ff0c62da2378150b791ad7d61f0b10b.jpg", "reference_inputs/images/423edee79517840c130357a8a1a08ab8664be9cec297a266fc992caea0152d92.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "B", "answer_index": 1}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_20m/000032", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G5\nB: G4\nC: G7\nD: G3\nE: G6\nF: G2\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c0cac91b3e4357b1ae2b7e0efc5e16a41e713a04bcd7bf5c9ea33e5f1a154579.jpg", "reference_inputs/images/502fe2b7bf5d1fe5c44e96178c57cdc8f6cda7fad2cf013c6853ede35a442838.jpg", "reference_inputs/images/07a0ad132afd99207a266320b96b576f8cf8b1518a143c3d17471bc368268383.jpg", "reference_inputs/images/9c4e551e68a3a6d6aa8c608c169cf03df9762fc2789534aeba9a6a5bf69ba888.jpg", "reference_inputs/images/64aea7386a93b8d411c31c3c476f955f4bb9d442dea86752dfdbfd3a90eec182.jpg", "reference_inputs/images/0babe2b74ff3bef3cdac98263fe4524186cb1fd3c2bd927891a0e7a8de627389.jpg", "reference_inputs/images/bc9ecf1a01f58e317cba1408870ca8cd9224a3287a4cfcfb65985c9815fc847b.jpg", "reference_inputs/images/44166659bf064b7b7f17ebe1cd27d0ffcc7acc728691ab96f5c75bf2df465995.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "B", "answer_index": 1}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000001", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G3\nB: G5\nC: G6\nD: G7\nE: G4\nF: G2\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b9aed418127730ac6dcce25014d245fe44574a85eaf01225858a47c233268bba.jpg", "reference_inputs/images/98e184ef25113e99d3acb81d2b999b2dc5d3a8d18056ae3ea6167479bc81beba.jpg", "reference_inputs/images/3dcc3b6ab37085f8ee320458b61a7cf7f2eab3f312e6f84ddc2d277d5a81c266.jpg", "reference_inputs/images/772ae829aab097e813edc986df6337aa53a4dfb46188b630cff56caf74228b46.jpg", "reference_inputs/images/affe958434ac9600d59e4e439ed10d5a41e9ed75ffaf1bb9084b9fe73c2ace8c.jpg", "reference_inputs/images/f48461b9bd7b9a45c46d40f6e514e1c95595739c2f5b3a83e72c20e56d0569b0.jpg", "reference_inputs/images/a9890d2fc6825dfa001f6cd4b2a99c35f9b6812d292bc7dd7e49d59b19403a0f.jpg", "reference_inputs/images/96f1bc32503e2dfa6e0f9c94b70312349e21eff30ac691b9e4f2e9f64cda9a1c.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "A", "answer_index": 0}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000003", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G2\nB: G6\nC: G3\nD: G4\nE: G5\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d0a991b667605daf5a5e004afed80e249d8def08534b8fb79a496148f3ac751e.jpg", "reference_inputs/images/7851f9a2784ea4549a5d2fa4c73ca071e9fc113790d62d9784b7186ed916dc5c.jpg", "reference_inputs/images/aa8f82f245f77dcb2f3486e154c356a050e275afdecf4171fa9fcf3dfb0f9c30.jpg", "reference_inputs/images/ddd5cc7103e6c241522c5a84922eb8dbb23cc83dd72ec92bc371f4d0b684f4ce.jpg", "reference_inputs/images/d51ea870530d60397e31de4daf06a29f1f5c6919ef5f1ec66f0e45ab6e30faa2.jpg", "reference_inputs/images/eb97a3246d862464d2b51b720f73d209b8d3a46bb24e54d9e5a2b762b5dfa240.jpg", "reference_inputs/images/9668b0e9de3420dc53f6742f0eaa1d9277e1041769fad82048ac9ad8b7f0e19a.jpg", "reference_inputs/images/57c3f70f35fc88b642076e72a90f518e3004de43237b3e468d97c6500ead1958.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "A", "answer_index": 0}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000004", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G2\nB: G5\nC: G4\nD: G6\nE: G7\nF: G3\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c09b21f70b972a8b4d94b68a5f30c53c3e5a50ce39e6df49973fb628fee25482.jpg", "reference_inputs/images/f4a7098cbe94ae907eafc0f5997b8a099436c210eb839ecea4f7bac29d27b2a5.jpg", "reference_inputs/images/a4ff943138300e14df5dde6b5b4d48e1550533ac77ffe671ed28e02e96353024.jpg", "reference_inputs/images/bcdc1e69c0a36fdd6f81b6dc138bac8948640e1f401b454a6715a6941bd5a4a5.jpg", "reference_inputs/images/c0d2a3e08cf989cab00b16bae02eabdb9d196be80eb7ba851f154ec5d4a68f11.jpg", "reference_inputs/images/ed2189f53b3f6d1c05bceb524c0005a7508a1ff0431f3713d530ae623c5b2642.jpg", "reference_inputs/images/34d027dbab11397f568f524e93a1358bb23b7502da274b0c04a1dba07548b837.jpg", "reference_inputs/images/0183cfd2866b057e06464418777bcf14316141f2c7c9027b08baf55273d48d83.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "A", "answer_index": 0}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000006", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G4\nB: G7\nC: G5\nD: G3\nE: G2\nF: G6\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0b8e4a727401c7aceb05c316d31a1f81f3c1b13a095e7d61595b909ff85a20a8.jpg", "reference_inputs/images/d29cac67f917b41bf253feba461d4cd1c2a80234455b6dfce3a8745ec2c770da.jpg", "reference_inputs/images/9f7eef9b329e7fd56acc8642c3e759719dbe90c61363422788930013ce9ee346.jpg", "reference_inputs/images/c0316a436e666a9a4e034067935dde38dd44cd7999b44800d78aff318cd543b9.jpg", "reference_inputs/images/4ea6834c85dd9da4d0e393b1fbd27477d525a255f2c04252747fb6893e6cca2b.jpg", "reference_inputs/images/0434ee60b659113cb97e064670e8b02e9955eeff7f24ccedb8f970345927aa67.jpg", "reference_inputs/images/576d8be8f3dbf01571e62e0948abed9b4f3a1622dc4e9a7e3af95ef90daf400e.jpg", "reference_inputs/images/9f89d5cdd755643210dfe8c38a9f83fd2ae46616e34f41ba1e56266b6bf6ac4d.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "E", "answer_index": 4}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000007", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G6\nB: G7\nC: G5\nD: G2\nE: G4\nF: G3\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/63544db7828b5fd8a2f82a6e0ff437a0744d955c86901e8ebc75a0a2366813c6.jpg", "reference_inputs/images/031f6f2231483341d76c0c8386636ae6741e856842165940c8c7df0fe15ec9e4.jpg", "reference_inputs/images/b4c54b7efb7c62abee50332956aaca8e96265358d30cb894395dc79f2eff5b28.jpg", "reference_inputs/images/ad26a1f82d612f9596c6e104842d9b22eb76e19fdd9cd98173f1cffb07c5df33.jpg", "reference_inputs/images/20aed003acb0895d9de2516f687f06e635b79ae30172ba9b8f93159e2c2fea6f.jpg", "reference_inputs/images/02ee26196f052db1fd6094389ca8a203753d4dfdd0196d5aad127f63ffb772e3.jpg", "reference_inputs/images/606dd5fea012af2756b373ee21b6a9f4a7ff6eca37a4eb356fb07f1cb3476f3e.jpg", "reference_inputs/images/5fd84a272bd50d7760735ef75150f5f8cfb3012b7bd945c024e8f2b370da7d2c.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "E", "answer_index": 4}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000010", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the LEFT-facing camera, so the ground images look toward the teammate's left side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G7\nB: G2\nC: G5\nD: G3\nE: G6\nF: G4\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/82f9a925f2bb85af895e89c13099eca8ab9fe3ea00f2fe1f6fd234627211aac9.jpg", "reference_inputs/images/c39e0e236839a3d83a8b6350bc0efe76a6fbc67aa9b6e4c2fc368e7de0a592cb.jpg", "reference_inputs/images/388b410086023ba52a82e8be7cc76f44d7c7a20c19a0c5cf882a1a665d6dfb2c.jpg", "reference_inputs/images/1e39378ac5766e78a1ddbeac3d2435169317ebdbb78727242c19741fb2c1f858.jpg", "reference_inputs/images/3eaadbc0c25a3b5713f45480a423aeb3ce28d83db7d841b55cc57d979135108f.jpg", "reference_inputs/images/1f2a00a0e56882a06b6b2ad1adfb16344a112cd9c178bc762cf53cbf501e6f03.jpg", "reference_inputs/images/17f870c3e80b4ea658ab9fad190bfd9efb07496fc221e0102bd1988432097401.jpg", "reference_inputs/images/884aef5dde587364e8a5aaea841d252a134051e9dc83fb3b8169cdfcc8320216.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "B", "answer_index": 1}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000011", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G5\nB: G2\nC: G7\nD: G6\nE: G3\nF: G4\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e3e56677e944eedc7e92c2f762c56e485f94a51b71c210e86992afc13e1c0f00.jpg", "reference_inputs/images/c6ac3c4fb2e5c73c58f26744d3b42a68e4504e8ec1b0f59cd9733ed6e295c7be.jpg", "reference_inputs/images/b0382be12d2c85e51697193d9193414c662b366062cf8c11b124a8efee5b4bf9.jpg", "reference_inputs/images/314d01bbfb4d254b346e7f6bc4fb62da813832c57ce8b979e13a5d0aae67e2bd.jpg", "reference_inputs/images/f2449548a2fb5afee173a59713a49688fa35c32ab2d5843906f2f9ba6baa0112.jpg", "reference_inputs/images/2a4ced4ecbda3f0ad9bfde9034be5e0a49ec9bab9c712884affd3ad08ecb94fb.jpg", "reference_inputs/images/1920d160bf8c207cc2cfe03c2f54780add45804d245793d2333db80413b5314a.jpg", "reference_inputs/images/326d0df537ddef5c1539b9ddbad502e4b450bb1d6abf8df538cadcd6be629d0a.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "B", "answer_index": 1}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000013", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the LEFT-facing camera, so the ground images look toward the teammate's left side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G4\nB: G3\nC: G6\nD: G5\nE: G2\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/087f1c24430ba4f5af36f9154a91f7a735f5996dedc4d1226e70679bfee1e4ca.jpg", "reference_inputs/images/a15fd033bcc55d3488847c7803672077677b2202ca743366107ecc874c18fad1.jpg", "reference_inputs/images/49f27fb1f76eac5e0f52afca875ebe856b4e9f60dcc74105da3782689cdaf5cf.jpg", "reference_inputs/images/7314b8f1b877457153c9f79fd61f048a1d3a6dda0ef1c25315ce50ed1f1c9712.jpg", "reference_inputs/images/55a985bde0abb3e5ad08f9b837111c168ca3381e9df6347ad27a6ef7782536a7.jpg", "reference_inputs/images/a6bbb87b287f5401a21be4253d6a9ba6d1b1fb7dd5a4446eec287e86133e81c1.jpg", "reference_inputs/images/4c59aef792f7b4cb068ed9ec42b7d976b072de55be5002948870efa2648c1b5f.jpg", "reference_inputs/images/010870be1ff3452306b89e08d9db3ca31baec1660cdaf57e6ddea918045c7365.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "B", "answer_index": 1}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000014", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the LEFT-facing camera, so the ground images look toward the teammate's left side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G6\nB: G5\nC: G3\nD: G7\nE: G2\nF: G4\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4409c7af1b18b6d703726ee2e8b79e2c1aae7ce0929b84005e1fbd35a84f170b.jpg", "reference_inputs/images/511dc7d282931166d0a31a405453bd26ae804a4dc5214093bb6a1982726f200e.jpg", "reference_inputs/images/5050b13e3207376f5eaac13dd4c2b17aa7de047c93b4adb173daf9a227c6bf3b.jpg", "reference_inputs/images/6de9c3585cf47e11898d5f77222fa22a23cf12d738d4e6bd2252109ea1039268.jpg", "reference_inputs/images/526d042da138444008692c69fea74589f8f58bda0125d9daeb37be0c85723551.jpg", "reference_inputs/images/82e49c092a52ba14ad4d6e963fee05e3b260013b8eddacfdcbdcbc606ff915ef.jpg", "reference_inputs/images/6fbf9e0872d4e16add6d278a01df5aa9f4b19360391d2d9887a86189ddce823b.jpg", "reference_inputs/images/ee43fb8141aa4576e3ebdf22ef02f663d6e11cac7f7e49ef76189736638c9d2a.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "F", "answer_index": 5}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000018", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the LEFT-facing camera, so the ground images look toward the teammate's left side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G2\nB: G6\nC: G7\nD: G4\nE: G5\nF: G3\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/24a85544167c46cd6ca38de6207846b791b110f0a922d13987121e58516e7d08.jpg", "reference_inputs/images/38b794c38af6958c62c3bf8c175f24dd96ded835eb611aef8431b16ece19e7c4.jpg", "reference_inputs/images/7a378ac14d01b7357c4a444b2c39e1b18ddac9b65d3d67bea2e8dca3ed826433.jpg", "reference_inputs/images/d1a2e0d8a7b1234553fb0f2b6b49098b855cc90c859e55d2b0ab0d19a2abd575.jpg", "reference_inputs/images/adc9f84d780eac8c352b4449a1a123832c0703b5e33227e663e90dc2368f5759.jpg", "reference_inputs/images/a60763d0d1b0ce193d9f444a7d18b463b8dd9af80e65adad392a15122348c50c.jpg", "reference_inputs/images/e304d7a9e22f0336047f9a4c01a819c39ffd0e777ada57b545c1312a956f2a58.jpg", "reference_inputs/images/bcd88a7dd65998984f1c98affe205dffa7973e75197bada87c1e55205fecc4bd.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "D", "answer_index": 3}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000019", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G5\nB: G6\nC: G2\nD: G3\nE: G4\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/abbe3d5d42abf29c3ee596b3de57e2b213cd7d714c699f30c11fc844afc61f98.jpg", "reference_inputs/images/9c4e551e68a3a6d6aa8c608c169cf03df9762fc2789534aeba9a6a5bf69ba888.jpg", "reference_inputs/images/0babe2b74ff3bef3cdac98263fe4524186cb1fd3c2bd927891a0e7a8de627389.jpg", "reference_inputs/images/e8f86a144ff07fab1b6ed4595caeffbf95d4782f342c62d61c804499bb0065bf.jpg", "reference_inputs/images/bc9ecf1a01f58e317cba1408870ca8cd9224a3287a4cfcfb65985c9815fc847b.jpg", "reference_inputs/images/44166659bf064b7b7f17ebe1cd27d0ffcc7acc728691ab96f5c75bf2df465995.jpg", "reference_inputs/images/ab4684364fe6a38c2026ce521fb39e2527721bca2271eb4e42b0311ffe907e38.jpg", "reference_inputs/images/64aea7386a93b8d411c31c3c476f955f4bb9d442dea86752dfdbfd3a90eec182.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "D", "answer_index": 3}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000021", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the LEFT-facing camera, so the ground images look toward the teammate's left side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G4\nB: G7\nC: G6\nD: G5\nE: G2\nF: G3\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d380cdf97e75bc342dc25779efb68fcbb34e40e01ec055874da5e65dc2daae11.jpg", "reference_inputs/images/3c1eed17076a7921cc0af79964c28dec9371a24e9179fdf879f58928ac06acc9.jpg", "reference_inputs/images/53851ffaf7a1cd23d50ed89c31c148ef608b08af5c51d1c73461e56cac54c969.jpg", "reference_inputs/images/ca8487ff33e0062cfeb68ed777e96a88e0c9431a893d7a94f47e9934bcac648e.jpg", "reference_inputs/images/59e4aa2a08930d64de52105fa7f3349672d21f92bf3adba38145d6c228691901.jpg", "reference_inputs/images/b0af3bb36620b6a360c476abd2224bdf5db62721179541bcc51903f4cb00b64f.jpg", "reference_inputs/images/bc2e0edb705d8f037d3486450a818548bd165f98f1d8657483a838a902624bb6.jpg", "reference_inputs/images/7bb72823dd1ff767bcfa0af18ffdf886134eb00fa75108eafdef971bd5ab925b.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "C", "answer_index": 2}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000024", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000024", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the LEFT-facing camera, so the ground images look toward the teammate's left side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G5\nB: G4\nC: G6\nD: G2\nE: G3\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/387fdacaccb7f502d0e0ab6daf897e1c7aed54393d37260a9a956d16da666010.jpg", "reference_inputs/images/adc9f84d780eac8c352b4449a1a123832c0703b5e33227e663e90dc2368f5759.jpg", "reference_inputs/images/39588efdc10aaf0df7d51c05278e318714f181e7498860913fc0400e413554af.jpg", "reference_inputs/images/6d416a27ccd8e05a56094ad0d2585cf7065b4df2698ad5b976a0c76dcc4e9bd8.jpg", "reference_inputs/images/38b794c38af6958c62c3bf8c175f24dd96ded835eb611aef8431b16ece19e7c4.jpg", "reference_inputs/images/7a378ac14d01b7357c4a444b2c39e1b18ddac9b65d3d67bea2e8dca3ed826433.jpg", "reference_inputs/images/d1a2e0d8a7b1234553fb0f2b6b49098b855cc90c859e55d2b0ab0d19a2abd575.jpg", "reference_inputs/images/a60763d0d1b0ce193d9f444a7d18b463b8dd9af80e65adad392a15122348c50c.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "E", "answer_index": 4}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000026", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the LEFT-facing camera, so the ground images look toward the teammate's left side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G4\nB: G5\nC: G7\nD: G2\nE: G6\nF: G3\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/91a132ad3ff779052b1f7063b2d4a28eaa233a34820858f7ea7cd6549be08d40.jpg", "reference_inputs/images/1c35f6d72b01ec6af2caa178bff92d13f3a4d9b4ec57b5e271a75bff2446db9a.jpg", "reference_inputs/images/aa07265962ce9129c110ee741915093469cd3f25c12624b6ce009dc9b39e385e.jpg", "reference_inputs/images/ff6c1dc8e7cddbfa32a9b40096be7e72aaf3a09ae2f0b40094b712370828563c.jpg", "reference_inputs/images/6f1921492b831a61cabf2999fbd5323f30262e211f6d8f33afff514562e0ae71.jpg", "reference_inputs/images/a15fd033bcc55d3488847c7803672077677b2202ca743366107ecc874c18fad1.jpg", "reference_inputs/images/137841317eedd34bac33fb0fa9528cf8515675bc81c503205b05f1026a3b2852.jpg", "reference_inputs/images/7314b8f1b877457153c9f79fd61f048a1d3a6dda0ef1c25315ce50ed1f1c9712.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "B", "answer_index": 1}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000029", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G5\nB: G4\nC: G6\nD: G7\nE: G2\nF: G3\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3d6490d40e5fba24c6232e7e115107cf475f1f9e9e1ba390f3029c50e45d4542.jpg", "reference_inputs/images/c5b600c09a6d37cd46165ccf13dda6e2f27b7d463018639951817900fbdef907.jpg", "reference_inputs/images/165cc78f38c0f249eb17768c4ed0e3d7f315364b213c61a5d41731ef443b90ce.jpg", "reference_inputs/images/3e0a4ac2e88225b48e8c398702ff8943834354639cb6fce673aafc0cc14fd71e.jpg", "reference_inputs/images/98e184ef25113e99d3acb81d2b999b2dc5d3a8d18056ae3ea6167479bc81beba.jpg", "reference_inputs/images/3dcc3b6ab37085f8ee320458b61a7cf7f2eab3f312e6f84ddc2d277d5a81c266.jpg", "reference_inputs/images/affe958434ac9600d59e4e439ed10d5a41e9ed75ffaf1bb9084b9fe73c2ace8c.jpg", "reference_inputs/images/f48461b9bd7b9a45c46d40f6e514e1c95595739c2f5b3a83e72c20e56d0569b0.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "F", "answer_index": 5}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000030", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the LEFT-facing camera, so the ground images look toward the teammate's left side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G3\nB: G5\nC: G2\nD: G6\nE: G4\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/03f192acd1322458f3e1d21b2686e6e3669b4d2b253e20531aaf2a690695fb1d.jpg", "reference_inputs/images/05653c6340d7df55442d03ed3d8119b93e8374d103577b227acac495e750863c.jpg", "reference_inputs/images/8d95f0b46b9a39bb273e3130d0a98cce355bab03a4458d815689821b0b28e7fa.jpg", "reference_inputs/images/45da49fdb274749d63e970bd37daf93437d2d5567cbf72ecf6ffa776e5544b3f.jpg", "reference_inputs/images/55aca58644f2001dd1fad5ed84021bd3b993bd877b87301b5ec3265ac716a94a.jpg", "reference_inputs/images/441d719f5a79e0e38f05c78dba4b0ec2d4909df8eb3d07b9ea9c8c88e86d1b1e.jpg", "reference_inputs/images/e76dbe907c6837b3d44cf0439fb413a89882221ab447303c0434c7702c4ba57b.jpg", "reference_inputs/images/8c76db10b89a1cf3ac66190f7fe41fc36d0d46416cce3ba018f1edc118ee675b.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "B", "answer_index": 1}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000031", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000031", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the LEFT-facing camera, so the ground images look toward the teammate's left side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G6\nB: G3\nC: G7\nD: G5\nE: G4\nF: G2\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/00632c3059146e0d01190e011617a51970779602ab498d65b68645f6a41a3e80.jpg", "reference_inputs/images/8c5f73d844824d15dda918f79413219de0ec916fc701703ec066bad68bcb72e8.jpg", "reference_inputs/images/0f0ea43f2b6fd8241a5377bbd7cd403bd37949f2e803d8fe2a088645c622dd3b.jpg", "reference_inputs/images/df7ec07cab13331ddda6099c5512cdb132ede891efe381bc1d7e1d41e4053a0d.jpg", "reference_inputs/images/c29d2ae0b8e4d26fa8e1494af668e571e75459fc2c61d548d1fdfe35fbebe69b.jpg", "reference_inputs/images/3a78694109d6f4e2aa6c3201e224184d3c1c7c4736b8adac3b14a0233f1f3ea6.jpg", "reference_inputs/images/1fcb23b032daeca65a1a4f829f0e94b427e95c2ad50d71810101b247352557ba.jpg", "reference_inputs/images/d3a0110b6a67af346162179241e854b0f3917a8527d17d0238e525236b325fc3.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "B", "answer_index": 1}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000034", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000034", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the LEFT-facing camera, so the ground images look toward the teammate's left side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G7\nB: G2\nC: G3\nD: G6\nE: G4\nF: G5\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/82f9a925f2bb85af895e89c13099eca8ab9fe3ea00f2fe1f6fd234627211aac9.jpg", "reference_inputs/images/c39e0e236839a3d83a8b6350bc0efe76a6fbc67aa9b6e4c2fc368e7de0a592cb.jpg", "reference_inputs/images/1e39378ac5766e78a1ddbeac3d2435169317ebdbb78727242c19741fb2c1f858.jpg", "reference_inputs/images/3eaadbc0c25a3b5713f45480a423aeb3ce28d83db7d841b55cc57d979135108f.jpg", "reference_inputs/images/1f2a00a0e56882a06b6b2ad1adfb16344a112cd9c178bc762cf53cbf501e6f03.jpg", "reference_inputs/images/17f870c3e80b4ea658ab9fad190bfd9efb07496fc221e0102bd1988432097401.jpg", "reference_inputs/images/884aef5dde587364e8a5aaea841d252a134051e9dc83fb3b8169cdfcc8320216.jpg", "reference_inputs/images/388b410086023ba52a82e8be7cc76f44d7c7a20c19a0c5cf882a1a665d6dfb2c.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "A", "answer_index": 0}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000036", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000036", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G6\nB: G5\nC: G3\nD: G4\nE: G7\nF: G2\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/81146e7c0afe0cc0f248304094fb17fb5ec0e52500622584dac1dc8ed091f8aa.jpg", "reference_inputs/images/ec8a2bc363414258252d8d5764ed286bf6287be8c9005fb019cd57f6c39f9294.jpg", "reference_inputs/images/9dabade18b21c6aa67c58127d7a894f94f432a931037529764f20ab13e6c9cdc.jpg", "reference_inputs/images/5f7d374ea1f609856cc0a326dbce75f0c14822bcb32ea1b54fe31767631b1b79.jpg", "reference_inputs/images/794cfbcbe7aebf0fd07de502de177b39301e6da824fb87e6bb1e4aa716ca8342.jpg", "reference_inputs/images/31eb9c1df4ddb78aec732f2efe66add5c8283f6802890055c69006d2f802da0d.jpg", "reference_inputs/images/16b98ca976b6de186d99c34f240152f162895e25eef46ec2f005aaf3034c6c1a.jpg", "reference_inputs/images/7b76f625b0e713b81c3b2fd9005caf85291cd0af2b37d0152ae222104a3abd57.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "B", "answer_index": 1}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000039", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000039", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G3\nB: G4\nC: G5\nD: G7\nE: G2\nF: G6\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/91da7e1d6ed94f58c0609254d3ea7589ef32c5973302203558e88a79174dc590.jpg", "reference_inputs/images/16652bd8a0749e2b8903fd22753cff94fa7d4fe455aeb70555cc2ffe81a55955.jpg", "reference_inputs/images/5b6ac385e3afda8add29752aadc7568dcfefbad224d4b7851478badc13cf2acf.jpg", "reference_inputs/images/6f0f1e07119f53abe8f08c28c43962d83b3debe85de2c64a6e89e21ebb95a02c.jpg", "reference_inputs/images/a48a18e7653d72be7bef71d38f63063825f2392195ab9c9f2863e8d057499145.jpg", "reference_inputs/images/37f3b24621705c5ca5ca35f0206472c1147c1f78bba9a690554c9eed51564071.jpg", "reference_inputs/images/f17f0642177cfd4dabd289ea57ad315d5726afa8866c914631a2578bdb2e5167.jpg", "reference_inputs/images/0954fbaf2620c6f5b47a8b30ce0b61391a5fa9e332bb66020514f148e81de546.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "F", "answer_index": 5}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000040", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000040", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the LEFT-facing camera, so the ground images look toward the teammate's left side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G4\nB: G2\nC: G3\nD: G6\nE: G7\nF: G5\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/823a9cb8c2fb770e1d46a50e57124cd03f01669109d7325ff0c09188588dc4ab.jpg", "reference_inputs/images/7ab9ff8e8464d2297c2789cd229fc3ea0b8c16defc2c0a0f58f8a335511201db.jpg", "reference_inputs/images/3e7836c5acb2dcdcaa8ae2df4de0058e570a9f85222862e2d16221447f8c8257.jpg", "reference_inputs/images/3c1eed17076a7921cc0af79964c28dec9371a24e9179fdf879f58928ac06acc9.jpg", "reference_inputs/images/53851ffaf7a1cd23d50ed89c31c148ef608b08af5c51d1c73461e56cac54c969.jpg", "reference_inputs/images/ca8487ff33e0062cfeb68ed777e96a88e0c9431a893d7a94f47e9934bcac648e.jpg", "reference_inputs/images/59e4aa2a08930d64de52105fa7f3349672d21f92bf3adba38145d6c228691901.jpg", "reference_inputs/images/b0af3bb36620b6a360c476abd2224bdf5db62721179541bcc51903f4cb00b64f.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "B", "answer_index": 1}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000044", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000044", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the LEFT-facing camera, so the ground images look toward the teammate's left side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G2\nB: G3\nC: G6\nD: G4\nE: G7\nF: G5\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9039f0fb911c805f08b947759eb972c0f0dfa1ffa6e0af32c4d309addec9ad21.jpg", "reference_inputs/images/ff6c1dc8e7cddbfa32a9b40096be7e72aaf3a09ae2f0b40094b712370828563c.jpg", "reference_inputs/images/6f1921492b831a61cabf2999fbd5323f30262e211f6d8f33afff514562e0ae71.jpg", "reference_inputs/images/137841317eedd34bac33fb0fa9528cf8515675bc81c503205b05f1026a3b2852.jpg", "reference_inputs/images/55a985bde0abb3e5ad08f9b837111c168ca3381e9df6347ad27a6ef7782536a7.jpg", "reference_inputs/images/7314b8f1b877457153c9f79fd61f048a1d3a6dda0ef1c25315ce50ed1f1c9712.jpg", "reference_inputs/images/d760e180b6d0fe7d5f5740e5ff281233dff08dc94c3c056a4d088a7a0203454d.jpg", "reference_inputs/images/a15fd033bcc55d3488847c7803672077677b2202ca743366107ecc874c18fad1.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "D", "answer_index": 3}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_30m/000045", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000045", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the LEFT-facing camera, so the ground images look toward the teammate's left side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G3\nB: G6\nC: G7\nD: G5\nE: G2\nF: G4\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/387fdacaccb7f502d0e0ab6daf897e1c7aed54393d37260a9a956d16da666010.jpg", "reference_inputs/images/adc9f84d780eac8c352b4449a1a123832c0703b5e33227e663e90dc2368f5759.jpg", "reference_inputs/images/39588efdc10aaf0df7d51c05278e318714f181e7498860913fc0400e413554af.jpg", "reference_inputs/images/38b794c38af6958c62c3bf8c175f24dd96ded835eb611aef8431b16ece19e7c4.jpg", "reference_inputs/images/bcd88a7dd65998984f1c98affe205dffa7973e75197bada87c1e55205fecc4bd.jpg", "reference_inputs/images/7a378ac14d01b7357c4a444b2c39e1b18ddac9b65d3d67bea2e8dca3ed826433.jpg", "reference_inputs/images/d1a2e0d8a7b1234553fb0f2b6b49098b855cc90c859e55d2b0ab0d19a2abd575.jpg", "reference_inputs/images/a60763d0d1b0ce193d9f444a7d18b463b8dd9af80e65adad392a15122348c50c.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "F", "answer_index": 5}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_50m/000001", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the LEFT-facing camera, so the ground images look toward the teammate's left side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G7\nB: G6\nC: G5\nD: G2\nE: G4\nF: G3\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0328e3176e3fb24cfa02edf01b9d0eef397c3196fb95a697d97c52a78989a092.jpg", "reference_inputs/images/b829ddf4c587e79a67ffe16fe25873fe19448bfe353db00f3602571b6511c0a3.jpg", "reference_inputs/images/5a3d23ade29fc32a74a40c56381bcee2a1f0227641011a213150ba8ecfd7e2f5.jpg", "reference_inputs/images/a0ccf0f5e696281230600960e2603ce51a994ea07386f34ba1410ded71be1b4c.jpg", "reference_inputs/images/6e0075721026d6fd27359c56d0e9c4227a96e72271da50ba1dddeb99595ec730.jpg", "reference_inputs/images/8e7f568f954f7c0f0bfb8ec47010fe4db1e648b710dace15eaeeb1fbe0629316.jpg", "reference_inputs/images/28a6f0a4998ff42a8c5b646c1daccf908c66bb0013d2bed11aef9a7f0fd6783a.jpg", "reference_inputs/images/6d95db184899f666420ea1bf17eb9cd5de13fef6d9f1366e130c76bf9637816f.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "A", "answer_index": 0}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_50m/000002", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the LEFT-facing camera, so the ground images look toward the teammate's left side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G3\nB: G5\nC: G6\nD: G2\nE: G4\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/63ecc5869d1c71f55d2f331147f3689e38c8e28ae67d925434019482e1314fe2.jpg", "reference_inputs/images/55a985bde0abb3e5ad08f9b837111c168ca3381e9df6347ad27a6ef7782536a7.jpg", "reference_inputs/images/a6bbb87b287f5401a21be4253d6a9ba6d1b1fb7dd5a4446eec287e86133e81c1.jpg", "reference_inputs/images/fb65f7995f334114946a52d3172118daba99133ed34921956d2365e3e709e2a6.jpg", "reference_inputs/images/4c59aef792f7b4cb068ed9ec42b7d976b072de55be5002948870efa2648c1b5f.jpg", "reference_inputs/images/010870be1ff3452306b89e08d9db3ca31baec1660cdaf57e6ddea918045c7365.jpg", "reference_inputs/images/c76018fa09beecb814e02681235d7bb43301726270df27b200cad7ec178d37ca.jpg", "reference_inputs/images/45d2275c42453df8341b5a3938a8a257c4613bfe4da9c4cc1ccb1a773cbd935a.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "A", "answer_index": 0}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_50m/000003", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the LEFT-facing camera, so the ground images look toward the teammate's left side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G6\nB: G7\nC: G3\nD: G5\nE: G4\nF: G2\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/edc8f249637ac7fddf4446972ad1dc339f684b5f3164192994073167f3d659c0.jpg", "reference_inputs/images/49f27fb1f76eac5e0f52afca875ebe856b4e9f60dcc74105da3782689cdaf5cf.jpg", "reference_inputs/images/55a985bde0abb3e5ad08f9b837111c168ca3381e9df6347ad27a6ef7782536a7.jpg", "reference_inputs/images/7314b8f1b877457153c9f79fd61f048a1d3a6dda0ef1c25315ce50ed1f1c9712.jpg", "reference_inputs/images/a6bbb87b287f5401a21be4253d6a9ba6d1b1fb7dd5a4446eec287e86133e81c1.jpg", "reference_inputs/images/4c59aef792f7b4cb068ed9ec42b7d976b072de55be5002948870efa2648c1b5f.jpg", "reference_inputs/images/010870be1ff3452306b89e08d9db3ca31baec1660cdaf57e6ddea918045c7365.jpg", "reference_inputs/images/c76018fa09beecb814e02681235d7bb43301726270df27b200cad7ec178d37ca.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "C", "answer_index": 2}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_50m/000004", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the LEFT-facing camera, so the ground images look toward the teammate's left side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G3\nB: G5\nC: G7\nD: G2\nE: G4\nF: G6\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/16c2519ceaac3aad8d15f086476ad5dbd01116d81d4c2e39de8a1e266878eb2f.jpg", "reference_inputs/images/9b83d6129aa61cdfbf7ff383b9a29ddbba8ea0418bbd5a27c7cb21866300128a.jpg", "reference_inputs/images/b829ddf4c587e79a67ffe16fe25873fe19448bfe353db00f3602571b6511c0a3.jpg", "reference_inputs/images/5a3d23ade29fc32a74a40c56381bcee2a1f0227641011a213150ba8ecfd7e2f5.jpg", "reference_inputs/images/a0ccf0f5e696281230600960e2603ce51a994ea07386f34ba1410ded71be1b4c.jpg", "reference_inputs/images/9d72af50c15709d60b52bcbcf822fc04d179054cfbbb51b138d609d973586ff5.jpg", "reference_inputs/images/6e0075721026d6fd27359c56d0e9c4227a96e72271da50ba1dddeb99595ec730.jpg", "reference_inputs/images/8e7f568f954f7c0f0bfb8ec47010fe4db1e648b710dace15eaeeb1fbe0629316.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "B", "answer_index": 1}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_50m/000005", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G6\nB: G5\nC: G2\nD: G4\nE: G7\nF: G3\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/148ff9a89b41824761ee7ce55648f2b3362a08c92264d364fb06325575870816.jpg", "reference_inputs/images/6e3dc1f4e62b47d16d62d043bed5a418481ebc45f4896c9dc7e3614a0781bb42.jpg", "reference_inputs/images/340ccec5bc13c8a3e32bf853b57b03107e0fbb2ab626287e3d7c66eabfe7e837.jpg", "reference_inputs/images/43b8a14348f565382ab0cc4fbbf8e087a90487c0764d3c3e5bec20884a3cdbb8.jpg", "reference_inputs/images/6d6df61bab2fdecf4a4ade212800131b7757e843304464e1832f01685f9cfb59.jpg", "reference_inputs/images/8d37f800ffbf71905c3386e8f60ca28c2dadd52fc67793a1c4e7b790f4df4fcb.jpg", "reference_inputs/images/d6fbe93a4f5e999ad00e816341ccc35818ab5a6bd7ce3f2b31a5f3a913b357d5.jpg", "reference_inputs/images/92282795f8085fefb8e9ffcc8ce2e6883e363ad40c06d9ed36c6c19d3912a194.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "A", "answer_index": 0}
{"sample_id": "reverse_sequence_reasoning/city06_ground_aerial_v2_50m/000006", "task_name": "reverse_sequence_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G2\nB: G5\nC: G6\nD: G4\nE: G3\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/463631a80757bd885d357e4bc9fd6852094004bec0d2c90270d4a9145a5cf913.jpg", "reference_inputs/images/1be0d45fc5eb88503ea3915f0b8fe008a69f91db0a9f5307220b605630c8732b.jpg", "reference_inputs/images/3e0a4ac2e88225b48e8c398702ff8943834354639cb6fce673aafc0cc14fd71e.jpg", "reference_inputs/images/ab664e3a1c1d9879f2ca1d50406f4e89dff8c209dd7abe815855c9c49daa0388.jpg", "reference_inputs/images/772ae829aab097e813edc986df6337aa53a4dfb46188b630cff56caf74228b46.jpg", "reference_inputs/images/c5b600c09a6d37cd46165ccf13dda6e2f27b7d463018639951817900fbdef907.jpg", "reference_inputs/images/165cc78f38c0f249eb17768c4ed0e3d7f315364b213c61a5d41731ef443b90ce.jpg", "reference_inputs/images/136f851f4ce6fede76cc50fdc3fc6048d01d9a5ec12763732777c13fbb606be9.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "F", "answer_index": 5}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000002", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G3\nB: G4\nC: G5\nD: G6\nE: G7\nF: G2\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/83783d28a6bdc32c9eb8c41fbfb7208a2f6cf292cd8ee27b29dd49eb0314f37b.jpg", "reference_inputs/images/518149accc0f9856453cce22182a5e4a7795254ce8e7f4fa19e0b75df58470d7.jpg", "reference_inputs/images/d364b0e3541250a939545692adbf2b7f5dd4eda4eed7a82b53526764b5620522.jpg", "reference_inputs/images/ee619f07d01e40018c525e41f160c6184809b171a852607efc945dc96e411053.jpg", "reference_inputs/images/87c8c0fa878d99e3199f221c3218b81775105d0a3f4a531a399201e685b8bb72.jpg", "reference_inputs/images/d6e149454f9363282bab06cfbbca15cfbb01c5944ba3526f28baff43fa8ffcfb.jpg", "reference_inputs/images/ed2797fcf780f0d9265d28800e0c45d9f1ba10a1346d31bd5cf51b29afd10787.jpg", "reference_inputs/images/05c8df436da91bf7e804433e570a7e078d64c9b41c40e31b46ac71eb9d7ad0ff.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "D", "answer_index": 3}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000003", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G6\nB: G5\nC: G7\nD: G2\nE: G3\nF: G4\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c7142a20c3f2faec284a42bf776fe0671c0dd1fde6a47f62c6e965707dd1f214.jpg", "reference_inputs/images/9454bfbf127a6ef586c8cbd21a88516a04c24bf9b4a4d2d70da2bdcf1d16e104.jpg", "reference_inputs/images/f51cdc4fdab962d37e5494600db16c9ce68b3079d84ae214af152e6704b497ed.jpg", "reference_inputs/images/4b42b02e39f44393275ac5e9a5490bd7d17f5f4b37b6848eb0fc0c58462fd46a.jpg", "reference_inputs/images/702dd7ba3d6815564455245f3199e57483d587890f24e75a7c679226bfe87823.jpg", "reference_inputs/images/8ed404085ac069432ca6885fa46e77d31c1a9cee698a07ec666853280f8b6fd0.jpg", "reference_inputs/images/153c42acc503b7f49f9717f56313e3f98e87cf0cd70405ca91151993eafe3310.jpg", "reference_inputs/images/b28aa8aaad129d53f280c819325ac5c973dbd9c05b862f43953b1c6b3a2077d7.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "E", "answer_index": 4}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000008", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G6\nB: G7\nC: G5\nD: G2\nE: G4\nF: G3\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/187a36c1af25b85bb917f5ef69024615d9bde230a9036c61aa3635046210b8dd.jpg", "reference_inputs/images/f732fef1023b152fb2d8586c31d138f9774e49c0669823a575ceb8d45f09768a.jpg", "reference_inputs/images/696db644fc0e650348f4be016e8e5fd9c181eee2de39d8b0f2f71ba4a8fde652.jpg", "reference_inputs/images/4b42b02e39f44393275ac5e9a5490bd7d17f5f4b37b6848eb0fc0c58462fd46a.jpg", "reference_inputs/images/9c43dbfb9730487e2314dacc8e225558ab9608abdab849984b6cfb3a9243b4d4.jpg", "reference_inputs/images/8f4940264acc56f4ce551a90fee909e68ecf5da8ccb78f8c6385b3774c7c049d.jpg", "reference_inputs/images/b458fc1c79ce19d141010b566ef27d6d7d4f034b0f01d15dfc3ad0ca65fb4341.jpg", "reference_inputs/images/15efa56c21d2bc6e235bb1a3ccf4483e7a48fdd00089930beb82826e589d1667.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "E", "answer_index": 4}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000009", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G3\nB: G7\nC: G5\nD: G4\nE: G2\nF: G6\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/eb6d142f637354cad8dc989ee80bbe7a372d79bc0ce1c4ac3a83e29f65584c5b.jpg", "reference_inputs/images/efa5658689482f7351dfb2d3e76efc3166d2ac78bdd9cedd302f88a92851f48f.jpg", "reference_inputs/images/b222eeecc0a0b3660e5632c382b7f4bdbff21bc456f06d0165822026551dda65.jpg", "reference_inputs/images/1e8dfe34369b30ada45b2492470778472ef1f107e14282b5d897e55a5f42301c.jpg", "reference_inputs/images/65b7e6f323bd65c5d7669cb2b5b8e28d3141f449e05c138b7d4ca072cfd32073.jpg", "reference_inputs/images/b3659a307c6248b2ff3d779a5a18fdfc2d903b7eec7988550335e97ebd3edd10.jpg", "reference_inputs/images/d63334c465eb308bedf9b92da88defca48fd5a31e60274ea5fa33125cb24724d.jpg", "reference_inputs/images/5a27af426e8071277d41fa0fca00b264ac130475ccaf1dd9557379cfc0702f31.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "B", "answer_index": 1}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000010", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G2\nB: G5\nC: G6\nD: G3\nE: G4\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ec4971cd69e2978b8b0bf8e6864acb1c3d35efea6b93eb11ed463049e608f6a7.jpg", "reference_inputs/images/bb1b3b3955a3f2a6491e5176b2d74a97d2bd6453e48df238d57ef60e4f23c296.jpg", "reference_inputs/images/42597b51969d57de1b9ac76713fac75b49147821fed12c2e40ffd93bef514610.jpg", "reference_inputs/images/0e5ff0ab8d31164aded8f8513ac5bd67763baf668e9c2ee63493117ca582e2ce.jpg", "reference_inputs/images/08ad072f4beec4f109ee7613067d420465a71994083955ca5b6c3f3dce6308e6.jpg", "reference_inputs/images/720a86271cc3052d849c2923e24186ae1089daec73036a13fb5bfee0d564c6e1.jpg", "reference_inputs/images/518149accc0f9856453cce22182a5e4a7795254ce8e7f4fa19e0b75df58470d7.jpg", "reference_inputs/images/d364b0e3541250a939545692adbf2b7f5dd4eda4eed7a82b53526764b5620522.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "D", "answer_index": 3}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000014", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G5\nB: G7\nC: G6\nD: G4\nE: G3\nF: G2\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7c9a8b1c8c4780f39858733cc498c3a63d25634119838ede9874dcba60ffba12.jpg", "reference_inputs/images/a1551f96b3d25ba9ed2f87e863aae600d24c590493c03bc33b735a099462156f.jpg", "reference_inputs/images/419c8e8c0256f002c06ea832dc2d2986c4c4ccb63acb18dc519acc5e537e18a2.jpg", "reference_inputs/images/b7cce4d0c6e9e1f8e3bbdf264066a794ae08968e56c63157c9245e38f9a67787.jpg", "reference_inputs/images/c63e1ff48fec679249f03adce2581fb770f181347ffa8ef2f6f887c9599ab6f5.jpg", "reference_inputs/images/8a88af8aed0a862d09c57f4777a709d8920949e88250c84d2a84d9db1501ea1e.jpg", "reference_inputs/images/84fa91bee78795f7087343d63acb876160d6be00b2cf4dafc8f6d77db679e763.jpg", "reference_inputs/images/6b5df6351b777266ba97aa5ce4ccfc7c33b5fafa8409b33b3fc6151bfd4b616a.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "A", "answer_index": 0}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000015", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G4\nB: G6\nC: G7\nD: G3\nE: G5\nF: G2\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/515e2bd838986896af66ea96ec1dc0a79eaf23994a0597268ce514f9f029c64a.jpg", "reference_inputs/images/518149accc0f9856453cce22182a5e4a7795254ce8e7f4fa19e0b75df58470d7.jpg", "reference_inputs/images/d364b0e3541250a939545692adbf2b7f5dd4eda4eed7a82b53526764b5620522.jpg", "reference_inputs/images/ee619f07d01e40018c525e41f160c6184809b171a852607efc945dc96e411053.jpg", "reference_inputs/images/87c8c0fa878d99e3199f221c3218b81775105d0a3f4a531a399201e685b8bb72.jpg", "reference_inputs/images/84fa91bee78795f7087343d63acb876160d6be00b2cf4dafc8f6d77db679e763.jpg", "reference_inputs/images/d6e149454f9363282bab06cfbbca15cfbb01c5944ba3526f28baff43fa8ffcfb.jpg", "reference_inputs/images/05c8df436da91bf7e804433e570a7e078d64c9b41c40e31b46ac71eb9d7ad0ff.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "E", "answer_index": 4}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000017", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G3\nB: G2\nC: G6\nD: G5\nE: G7\nF: G4\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/827bbea8ed80cce3ac3de30a62390e27f7cd1f415e5c5f73e55652536fbda023.jpg", "reference_inputs/images/ac9a067321586244d3dae504d5e0938ffa4068e23f67bbdf00ea1667bfd1597d.jpg", "reference_inputs/images/b8a086918400b04805c6ab807ac91ad2108f05ac9d2379c90888138768a44312.jpg", "reference_inputs/images/3287c0cde9d06973f68e05609776faa670ad090c0dde10bcec003ba20bdbc5b3.jpg", "reference_inputs/images/5130830fdc667694537466c1b0917d92aa254c90c62b32fe6253654bcdf7e859.jpg", "reference_inputs/images/5150e45c134801ea14d103814c322770db053a9823846f287b8ae18b5a709c32.jpg", "reference_inputs/images/2e21fad0690d5f1b76873564a0c197839d27731f4ddc73083afac1751880b3db.jpg", "reference_inputs/images/dce70d33d8f9602df8bca856a63b807f674d48cc5394e78f6422eb83339193aa.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "D", "answer_index": 3}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000018", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G2\nB: G6\nC: G3\nD: G4\nE: G5\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d18b0bb2df78e5bc8de5f05aafb8d644a6ebdd180cb4b61c7d2fcb7a2dfde27b.jpg", "reference_inputs/images/c89b84a5e3e02c5afcc2931c6736b1e0e33d291e8e3af8e94d7079fd1e3c67dd.jpg", "reference_inputs/images/6dd7fca2b6d17dd9c99effce28ce8aa2c7b8f274bbff06780d240f09f540dca2.jpg", "reference_inputs/images/b11bf35d3712babc0da4f0313f7dd5a36c4101ed617633742de2e8677fb30e41.jpg", "reference_inputs/images/290740d19e85a34a6cade332bbff9dd10a00c6ed1395dbf78e0130c3a1996c38.jpg", "reference_inputs/images/973e2c87516b4091ec950f851f5e749c541b80b23e338723ae4ee782c7301349.jpg", "reference_inputs/images/ef3c218ecaea2ac99c114f64d51124cc74e44a720230ab7f9568e9954f95e8b2.jpg", "reference_inputs/images/5a90ce31dd9fc1a279cce695eff2a453d4915123a6d3eb7c993b12061bab48f1.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "D", "answer_index": 3}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000020", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G6\nB: G2\nC: G4\nD: G7\nE: G3\nF: G5\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/01d6d7d1577df382b227a369d597e07898ff0ec2cf8bfdc6ead1c4573c9240e6.jpg", "reference_inputs/images/d6e149454f9363282bab06cfbbca15cfbb01c5944ba3526f28baff43fa8ffcfb.jpg", "reference_inputs/images/05c8df436da91bf7e804433e570a7e078d64c9b41c40e31b46ac71eb9d7ad0ff.jpg", "reference_inputs/images/1d84a9a0b9eae92903a90403b8e36ed93169b339aaf2332fd02ab0937159618e.jpg", "reference_inputs/images/ed2797fcf780f0d9265d28800e0c45d9f1ba10a1346d31bd5cf51b29afd10787.jpg", "reference_inputs/images/6f39b629e2c109619264a43baa030a98e49a1b530eb85cf2a6b2df0d7336e4fe.jpg", "reference_inputs/images/82352d0ec68cbcc276f89b6e3957823cbca981552fe094fbc81873bd4a89bdfe.jpg", "reference_inputs/images/8a27f6f27475bd1f64e42fd6b54dfb6a99b3e314601ab91db35ee291f67bcc1f.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "F", "answer_index": 5}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000021", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G3\nB: G5\nC: G2\nD: G4\nE: G6\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fb349754668e624cfc6015db1258aac81eb531849f97dd3d43b799b699d40037.jpg", "reference_inputs/images/ee619f07d01e40018c525e41f160c6184809b171a852607efc945dc96e411053.jpg", "reference_inputs/images/87c8c0fa878d99e3199f221c3218b81775105d0a3f4a531a399201e685b8bb72.jpg", "reference_inputs/images/d6e149454f9363282bab06cfbbca15cfbb01c5944ba3526f28baff43fa8ffcfb.jpg", "reference_inputs/images/05c8df436da91bf7e804433e570a7e078d64c9b41c40e31b46ac71eb9d7ad0ff.jpg", "reference_inputs/images/1d84a9a0b9eae92903a90403b8e36ed93169b339aaf2332fd02ab0937159618e.jpg", "reference_inputs/images/6b5df6351b777266ba97aa5ce4ccfc7c33b5fafa8409b33b3fc6151bfd4b616a.jpg", "reference_inputs/images/ed2797fcf780f0d9265d28800e0c45d9f1ba10a1346d31bd5cf51b29afd10787.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "E", "answer_index": 4}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000022", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G4\nB: G7\nC: G5\nD: G6\nE: G2\nF: G3\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4e70ca772a0c490a270856e4159d4e714955ac8625c9694929158077b24c0529.jpg", "reference_inputs/images/ae20cf585532b43c40ad46f616b7781893b01fe8238e42691c424f3f746caf4a.jpg", "reference_inputs/images/d8b90566b8858c821792263f232b94b97cfbbcdd36a917c9d73fc911fd7760da.jpg", "reference_inputs/images/23690fff2c9f600d9272f12cad58dc499c1720d2d1fe0a9117ee949497c20b4a.jpg", "reference_inputs/images/5c813a42655a831c1e9be223a0ce85cc187b01a741ae67b1a073bee29034fea9.jpg", "reference_inputs/images/30260f4557385cfaffdbff2134fbf3283819655eb5e46a4ff757dbb097a74244.jpg", "reference_inputs/images/b9fd7c5ca9dc4373d3f3da922e6b43fb529998bc3bbd700ce86ad3c69a4b8540.jpg", "reference_inputs/images/7b93f66815b4158d5b0fc541b3bab05a50fe7688b77c59e8fdbcad8a16266806.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "F", "answer_index": 5}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000023", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G4\nB: G7\nC: G6\nD: G5\nE: G2\nF: G3\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8712a2fc54f2995d04bbd70d64d4d3fc7671c5d4e7dc77197cc9e4695b7ff11d.jpg", "reference_inputs/images/ae20cf585532b43c40ad46f616b7781893b01fe8238e42691c424f3f746caf4a.jpg", "reference_inputs/images/d8b90566b8858c821792263f232b94b97cfbbcdd36a917c9d73fc911fd7760da.jpg", "reference_inputs/images/5c813a42655a831c1e9be223a0ce85cc187b01a741ae67b1a073bee29034fea9.jpg", "reference_inputs/images/30260f4557385cfaffdbff2134fbf3283819655eb5e46a4ff757dbb097a74244.jpg", "reference_inputs/images/b9fd7c5ca9dc4373d3f3da922e6b43fb529998bc3bbd700ce86ad3c69a4b8540.jpg", "reference_inputs/images/b28e4de0f5b59dbe34d9977ae7924f1d6f127b56999d871fee4297beb2c3d38a.jpg", "reference_inputs/images/7b93f66815b4158d5b0fc541b3bab05a50fe7688b77c59e8fdbcad8a16266806.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "C", "answer_index": 2}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000024", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000024", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G4\nB: G5\nC: G2\nD: G6\nE: G3\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9b66ff21a24efece02d4713df66ac5a6b2ed81588c3e733167e1d5a0e166c9c4.jpg", "reference_inputs/images/5c813a42655a831c1e9be223a0ce85cc187b01a741ae67b1a073bee29034fea9.jpg", "reference_inputs/images/30260f4557385cfaffdbff2134fbf3283819655eb5e46a4ff757dbb097a74244.jpg", "reference_inputs/images/b9fd7c5ca9dc4373d3f3da922e6b43fb529998bc3bbd700ce86ad3c69a4b8540.jpg", "reference_inputs/images/7b93f66815b4158d5b0fc541b3bab05a50fe7688b77c59e8fdbcad8a16266806.jpg", "reference_inputs/images/23690fff2c9f600d9272f12cad58dc499c1720d2d1fe0a9117ee949497c20b4a.jpg", "reference_inputs/images/9bc68839e00db54a19f912d51d29cbf922c2bfd5caa8992f4ca230df633f852f.jpg", "reference_inputs/images/7d758c7d1c79b92a440d7e00a7e32861ba3eac76600935692b879683b2fa4d62.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "B", "answer_index": 1}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000025", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G6\nB: G4\nC: G7\nD: G3\nE: G2\nF: G5\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b3e585b3834fcea98382f7faa062d40c96d52ddb099fef2d12b2ae44f3b8bf9b.jpg", "reference_inputs/images/b7cce4d0c6e9e1f8e3bbdf264066a794ae08968e56c63157c9245e38f9a67787.jpg", "reference_inputs/images/c63e1ff48fec679249f03adce2581fb770f181347ffa8ef2f6f887c9599ab6f5.jpg", "reference_inputs/images/84fa91bee78795f7087343d63acb876160d6be00b2cf4dafc8f6d77db679e763.jpg", "reference_inputs/images/6b5df6351b777266ba97aa5ce4ccfc7c33b5fafa8409b33b3fc6151bfd4b616a.jpg", "reference_inputs/images/bcc608cf737be42508f55b7b8593598ed9ffcad4a311d402638a541def9ff015.jpg", "reference_inputs/images/cbf42e21aa5e52d5b279acc05c3f5cee3c30e56eea52a6a51464d271a4771a84.jpg", "reference_inputs/images/0270e6c49c8feef3c5b6875c7c99e75229547dceb165bee3d8a659b59ba4fe54.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "F", "answer_index": 5}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000026", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G6\nB: G2\nC: G3\nD: G5\nE: G7\nF: G4\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3ffec15505a41da9747415e39a0ccaccb59121f46ad0deba928d75a56ab6357c.jpg", "reference_inputs/images/4ac750855081052d67777616342037ffb929eb82e8c5482dd32135476a09a272.jpg", "reference_inputs/images/b28aa8aaad129d53f280c819325ac5c973dbd9c05b862f43953b1c6b3a2077d7.jpg", "reference_inputs/images/368df42215c11cc923177df9dbd1e6976d129ac20bea2499186d182c89dabf62.jpg", "reference_inputs/images/991ccec01d07275f54ef0c23d1fec3855790da8a1c97e1745754d92509cfb8b9.jpg", "reference_inputs/images/d7773a87bfc65223ef136cddb83885c4a9ac2237e41809e937b5e43cb65c17c2.jpg", "reference_inputs/images/8c822065e6a4a3137d25a6abc18bb3c2c267e21f7d088cad1a3a17dbee7b5c2a.jpg", "reference_inputs/images/9454bfbf127a6ef586c8cbd21a88516a04c24bf9b4a4d2d70da2bdcf1d16e104.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "B", "answer_index": 1}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000027", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G4\nB: G2\nC: G6\nD: G3\nE: G5\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/acdf5df5a562a746e3550b3dee1e54cd4994fd5c52f03160cd4c5e65241e92eb.jpg", "reference_inputs/images/be43c524ae2c0a4f5cba49ce67e3a33f7ca881c5b79e6c1eb61e6f8372beb563.jpg", "reference_inputs/images/cb10762cc79fb403ed3d060779fd177fdbc256d05c9dc19842a408604ffd0524.jpg", "reference_inputs/images/d5086606c7e6c9862835a0ca1ce281e192ed5d712cdffd431b2c73a4c3e1e998.jpg", "reference_inputs/images/9454bfbf127a6ef586c8cbd21a88516a04c24bf9b4a4d2d70da2bdcf1d16e104.jpg", "reference_inputs/images/a64a4eeffc4e041cfee0750c7ca09cfaaa10ab0dd50029a956576eef779531a2.jpg", "reference_inputs/images/209bd4735a24f658dec24cceca9171d144dbbabcfe4542d281051a67651e5bf0.jpg", "reference_inputs/images/4ac750855081052d67777616342037ffb929eb82e8c5482dd32135476a09a272.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "A", "answer_index": 0}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000030", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G5\nB: G6\nC: G4\nD: G2\nE: G3\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1cb9e05d45474809837f9fc0aa3038e93d3b6c4ed58f43434f32e1961f927748.jpg", "reference_inputs/images/17677f743d8a6da8619151c529c0aeabda048dcabd2d71f1a689c00c52dca099.jpg", "reference_inputs/images/5e553b549fe3011ee9d0101f7ac56380df2405e35ffa82fed2dd7d99d59fda7d.jpg", "reference_inputs/images/6294d811879d71885a913179fe579173a851d3f2de039e4ba15a9825209b3a95.jpg", "reference_inputs/images/5a27af426e8071277d41fa0fca00b264ac130475ccaf1dd9557379cfc0702f31.jpg", "reference_inputs/images/f8b96bdfbc62edc88f05d93287b7bcce7634eacffe424588fe3a13787206af43.jpg", "reference_inputs/images/0cc5a1937e0c95eac9263f2c5b07755f969ac41381db0cd1bc25a07dffb78689.jpg", "reference_inputs/images/84a8bbc0fe5ff4efe8a3e6a049b3903e2c2bd276c010cd43b70cfd360d7bf897.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "B", "answer_index": 1}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000032", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G7\nB: G6\nC: G3\nD: G5\nE: G2\nF: G4\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/25f3316c1a4bc622c91c532d60cc58bc8112f9c56e665a0af8090799c36d03a8.jpg", "reference_inputs/images/604b0bba6fab3a8e838fe265fe169c465a082d5863abfe365e17550612bb0f6d.jpg", "reference_inputs/images/ae4d1bcde5555a2743e98a6971bd3ff58fc27085dd56b5a570a70204a4430c3f.jpg", "reference_inputs/images/165fbcae599cf7a3a388132616855423a6c4a51f382b94071614dded67ca684d.jpg", "reference_inputs/images/c52d3ed15c8ad4699b4bc938831c5b06416414b65cdf6d09847efe920d0a1925.jpg", "reference_inputs/images/6aeff7f4a3fe7015e7da04d5c9d75664c3247c67da7bc0db28bb73528a27a828.jpg", "reference_inputs/images/9a69d2281c075c029d9e853ab666d2329b7565fb2a65dbefa849426a50e127a5.jpg", "reference_inputs/images/111827f8170c458bbf3f0f21bfd81924bd4896a320b3e2350563f7896d9f636c.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "A", "answer_index": 0}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000034", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000034", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G6\nB: G2\nC: G7\nD: G3\nE: G4\nF: G5\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9dc7b827a98b5f314e11a12c8f62b0cf4fa199768a0df12452dd66d25c247365.jpg", "reference_inputs/images/778acd66ce6b79ac291b2d9c2a18217f97c9cf2ed7c14781bd1bcf9246f8b465.jpg", "reference_inputs/images/fdf2d889c2dffa46fe69ae36b3d498a5696e384b59b17da14b65c985f03e64a3.jpg", "reference_inputs/images/eb9def764525c9bf1b1f3228bdf3015051c1b352d9a60cef5c6a46bbb4e73500.jpg", "reference_inputs/images/ee5f53244b68d895548d299b14a24319f0482734af72d321c9198ba5b830dcaf.jpg", "reference_inputs/images/462d3dff450dfd5c907f20d1d934b7066d1a5c8b3121c61e4aef9b8e13af0603.jpg", "reference_inputs/images/8f0e184da0c5396462d9803ef969a439f6cc3a75d2dd124ddf0e54e8cb05b6a0.jpg", "reference_inputs/images/fb83154e0a62656e6b5c69979f54dd7b956389860903523b5d68cc091995ba8d.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "B", "answer_index": 1}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000035", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000035", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G4\nB: G7\nC: G5\nD: G2\nE: G3\nF: G6\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/850cb0b446a4e548cf3a3717f936a0de7c12f58dfd5b54912b946b8a003a90ec.jpg", "reference_inputs/images/2bec413c323574edc50994e6be33fd04e6dd8462cd85f8440d3597eba943747a.jpg", "reference_inputs/images/327ed54209a4e0208580f49293da1361bcfb4613054cee6bb41c17e9c5835bb1.jpg", "reference_inputs/images/0cff59f867fe7b86097ed51822007c1913efb6a96246b1483f058af263a31611.jpg", "reference_inputs/images/92d0180982628f8a0f6c22980f4af4c90172b66fa59a007b5a04f60b2a1a9574.jpg", "reference_inputs/images/7a0bf23ec2ab312ed51cb9da293f10c6cf147291121dfec77e840170a2766309.jpg", "reference_inputs/images/29180798b6def462313b800f7cbd24174dba0446ce21c69b7e0ad3d3e8083450.jpg", "reference_inputs/images/0c8a9b920fb752d17854e01eaeabd648352c592e688d3546ef3830f74c8df359.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "D", "answer_index": 3}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000037", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000037", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G2\nB: G3\nC: G7\nD: G5\nE: G4\nF: G6\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1d94208f45ed7a1e82c6a8a663efc535bb830a3955b46e9ccfcf36ba89d0a5e8.jpg", "reference_inputs/images/65b7e6f323bd65c5d7669cb2b5b8e28d3141f449e05c138b7d4ca072cfd32073.jpg", "reference_inputs/images/b3659a307c6248b2ff3d779a5a18fdfc2d903b7eec7988550335e97ebd3edd10.jpg", "reference_inputs/images/d63334c465eb308bedf9b92da88defca48fd5a31e60274ea5fa33125cb24724d.jpg", "reference_inputs/images/ccf2f534069fb5ee7c79971e7991be130cf08e6f3755a61c937f36f653b973c8.jpg", "reference_inputs/images/3e91e639215d5b63a541258e591be23d5a87fbebe2f57fbccf561df8154d2497.jpg", "reference_inputs/images/7934020d0894b6819bb95d004f542b8eb70efd00efd7cdb166a0e2f98a02f6f0.jpg", "reference_inputs/images/356e298d4b7742b6b7c826b061b76cb8db7acd128030789aca33d2c107501fd6.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "C", "answer_index": 2}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000038", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000038", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G5\nB: G6\nC: G2\nD: G7\nE: G3\nF: G4\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/eab6bad6282e4bda42ad7b3ea0c8f5cffeae778dd7485e49abdf46e41247015c.jpg", "reference_inputs/images/c950cb11a50cf4d28f9ed3c51658284e921239f740a09eeda1a7e56759f58f1a.jpg", "reference_inputs/images/f96a9e8a485297f23f48e8c3a52f901786fcb17edf58d48e4476b02dc9058311.jpg", "reference_inputs/images/a28860f9ddaa250432484793a3d3aca1595d7b0ffe5d4f6f3cc5fbd29e8417ad.jpg", "reference_inputs/images/15b8080277590b51a8efffc4f9fe5df0fa8de5ee3f4e6ea68443818304fc8d0a.jpg", "reference_inputs/images/49a8d354e7a01418072c7c697722506dd87cf39e7bd223bd9a07d8e47d694c7a.jpg", "reference_inputs/images/fdf2d889c2dffa46fe69ae36b3d498a5696e384b59b17da14b65c985f03e64a3.jpg", "reference_inputs/images/90ce09656a6a167da5f7fab284aee25ea2185e3afc014d9222e307ef4bb41141.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "E", "answer_index": 4}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000039", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000039", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G7\nB: G3\nC: G4\nD: G6\nE: G2\nF: G5\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7039dac35dca98d0d51e2b0b902db86372320b7c34eb2df6ea88d4cd3152157c.jpg", "reference_inputs/images/614bea61ac9a60f84dd3186ff090ac89c59c5f19c19daa82db3fc2053a6cd2cb.jpg", "reference_inputs/images/18af91d51cab25cd269148c24de3e6c462001bcbe4cdf2ad120569133edaefec.jpg", "reference_inputs/images/d3558a936a65613cbbec820d11b51cc2ac958af87155f8146abf0217b9691467.jpg", "reference_inputs/images/852baac5526a6621171f2c610f1452b58769ff01607d0541ea0b62c5c826c2fe.jpg", "reference_inputs/images/317ab9b52b0f25ddefae5a487bf7d0fe2d07557b135018f3c808b09f864b4bca.jpg", "reference_inputs/images/30142c91b4a377fd8bb01708b286c070258fd5c4a84e7771a9bec6c4b1def4c8.jpg", "reference_inputs/images/6c1faeec1b212385a6473aba22bc10edfcccad8ee4a4d132b5fa6116fc167924.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "A", "answer_index": 0}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000042", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000042", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G7\nB: G6\nC: G5\nD: G2\nE: G4\nF: G3\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5de7d04579ce4ed7a3cad7ff43cc2621cc7277342993edf7803fb99dd647fcc7.jpg", "reference_inputs/images/7e335e03c825200cce7a9e42c6bbfaa0e960ed5feec24233d18da53e6340c06c.jpg", "reference_inputs/images/8495e0a6e823168148c752bc98323fbbb0310bc13ba7e889b83d5fa2332d8989.jpg", "reference_inputs/images/58ada6f3bd0b224329f107337fcf267ddbbfba0aa96a7e436680b6f7e41541d5.jpg", "reference_inputs/images/7cad57175d82d4f201ed3adc2e233f500c5b0f78d029ee33ac9fc179addfe37d.jpg", "reference_inputs/images/ffeabae965ca87a9a76fa6420a55de21d573a30f19c479c2e2f9f7fc91d2e546.jpg", "reference_inputs/images/a1551f96b3d25ba9ed2f87e863aae600d24c590493c03bc33b735a099462156f.jpg", "reference_inputs/images/c63e1ff48fec679249f03adce2581fb770f181347ffa8ef2f6f887c9599ab6f5.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "A", "answer_index": 0}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000043", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000043", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G4\nB: G2\nC: G5\nD: G3\nE: G6\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b402901ec90efff4a8deb177ef8df6f42e400cafd1af843f53f619bdfda8e80a.jpg", "reference_inputs/images/7e07668d7a6b94a053a99a2a815b4722d357061700a25e036f0f4e03b45484e0.jpg", "reference_inputs/images/b85a6b29f399b960fda0517b94c5f99319b190ff4a324961395b1f408498e61b.jpg", "reference_inputs/images/d4ce21fb471bc9ebf9528a2e13052e2a567ca745976ee2b1323e8a89a9407f72.jpg", "reference_inputs/images/e725185d137f557ca071a8e8d091ddcdd4b5a0ec63ab624c26e1d69739cc434e.jpg", "reference_inputs/images/f9b4accc7fca80bd494080fd0fd6409ff3db8da78407b9f4a7be9560dc358392.jpg", "reference_inputs/images/fa38e654e7d24f439480ca85078b43de575e08da51ef24adf0cb83b923b47e43.jpg", "reference_inputs/images/5da55d9a67c0398174a73b7c502ba42ae65f2b8981379fd7b7b6fb3d720d9bf2.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "A", "answer_index": 0}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000044", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000044", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G5\nB: G2\nC: G7\nD: G6\nE: G4\nF: G3\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e03678471ad0a0a882f49f90f49cba746cf3da7d2784a587a09095732448b497.jpg", "reference_inputs/images/90ded380178ee48995f1ae42f5000217afa8d28b717d39d22a2dea889bf2c026.jpg", "reference_inputs/images/209bd4735a24f658dec24cceca9171d144dbbabcfe4542d281051a67651e5bf0.jpg", "reference_inputs/images/e725185d137f557ca071a8e8d091ddcdd4b5a0ec63ab624c26e1d69739cc434e.jpg", "reference_inputs/images/af4b4b12b6990b9197b41e2fce0d52e8a3d2d083ff1a31cf5b3c6ed0e4e29190.jpg", "reference_inputs/images/8d1edcc8b4f0ac34deaa785673c23e6570cfe67f401af03ad7373b196ee902b3.jpg", "reference_inputs/images/be43c524ae2c0a4f5cba49ce67e3a33f7ca881c5b79e6c1eb61e6f8372beb563.jpg", "reference_inputs/images/cb10762cc79fb403ed3d060779fd177fdbc256d05c9dc19842a408604ffd0524.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "B", "answer_index": 1}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000046", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000046", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G6\nB: G2\nC: G3\nD: G7\nE: G5\nF: G4\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ca864ac2cad28e84bf2a45811fad6e53c0780d7e291454c55c1b788cd1679a98.jpg", "reference_inputs/images/852baac5526a6621171f2c610f1452b58769ff01607d0541ea0b62c5c826c2fe.jpg", "reference_inputs/images/317ab9b52b0f25ddefae5a487bf7d0fe2d07557b135018f3c808b09f864b4bca.jpg", "reference_inputs/images/30142c91b4a377fd8bb01708b286c070258fd5c4a84e7771a9bec6c4b1def4c8.jpg", "reference_inputs/images/7c4f0d9c44a8609aa39e413ab375dbd6030efec0819d7b0772e76f733400349a.jpg", "reference_inputs/images/02fa92474473ebe9f1a1fda3ba193c9d869369b4b68692e22ea9a78bfa93953f.jpg", "reference_inputs/images/2969ea4fa30a89a9dec3fd0ec29a8b7a41dd1fbfc3dfc6a0028bc779a0abcff1.jpg", "reference_inputs/images/6c1faeec1b212385a6473aba22bc10edfcccad8ee4a4d132b5fa6116fc167924.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "E", "answer_index": 4}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000049", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000049", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G2\nB: G6\nC: G5\nD: G7\nE: G4\nF: G3\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c325566774b4c35463fe53805f4ee134a6a4109cc5e9590bbf571df2bf2b2ae6.jpg", "reference_inputs/images/8d920b98306faae1dd570d511d985769d0571056126f4c467bc883416cf04161.jpg", "reference_inputs/images/462d3dff450dfd5c907f20d1d934b7066d1a5c8b3121c61e4aef9b8e13af0603.jpg", "reference_inputs/images/cfd7002de6019ef0fb8c2c3841d17a36f430d07616395ad65086aea03ef16598.jpg", "reference_inputs/images/bc60445ea52425ada75540534387f46245b9f63fd8a70fd45360dfd726a0f883.jpg", "reference_inputs/images/c1fb896aae3912e6c5790383435d3e0aba6876c0dfa6ce3ef5a0d93ba3be3375.jpg", "reference_inputs/images/839e7480594c685cc455a57a7f79881f7053bafcfd91d1204a3a1244770b77d3.jpg", "reference_inputs/images/0d1249b0b489795184fdf483552ed60b7f003460f9096becd497208470f58342.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "A", "answer_index": 0}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000051", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000051", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G2\nB: G5\nC: G3\nD: G6\nE: G4\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ee8a5a53f7b4acf998b6a0450b6e4d5c2aba695ff733ea625490c942de53e5c0.jpg", "reference_inputs/images/78fe1bce8137d790e4412107e620d82758be4b98df99fea12ae0bcf5de56d8af.jpg", "reference_inputs/images/cbdd1a6ed8b5b494d10165aa541a404c08ab908f6e9e23b313eeebf05ec1f2b2.jpg", "reference_inputs/images/99844ffd0082a8c59921a0d522eed8757940efe1d74aca3a8fa296e7efc3337a.jpg", "reference_inputs/images/0f90ddd60f46164f884041a99a443c3b40d81a26059b3783928740213b5fc2f6.jpg", "reference_inputs/images/ac103d670d95b052d3b29ca9676e329d6a667ee6aef5f8017117c0d8e1f4031b.jpg", "reference_inputs/images/c2c95b5c6331eb56f5a918511ac79d808619a3ab82ef6bd3274f59e313d75da7.jpg", "reference_inputs/images/bcd85d25ea43bf1d43d6949176713d3e2bd483d6674c0f1f659ac95eda5d6ac0.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "B", "answer_index": 1}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000052", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000052", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G6\nB: G3\nC: G5\nD: G7\nE: G2\nF: G4\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6411a9e71ca8f0ca181f5e5c2ae981967ab69dc28001954511cacac7abc20d14.jpg", "reference_inputs/images/94ad25a01e79c6f2579f8ee6c6169e22e305ebca76a4a24e2a23c64949e9cb50.jpg", "reference_inputs/images/9aad48e0f370f5a3bbc727932acf204f3be58bf1e9515df50d9ed8abbe4eeebd.jpg", "reference_inputs/images/eacaf1a98a9dbc744a26038fdc19ad727532b4a998c7823ded6c5b202ae5e67a.jpg", "reference_inputs/images/bfc1fb5fe7cb11654b62cb5be6043c83f36ee5eba7f398484c2053fbf88b30c7.jpg", "reference_inputs/images/09877c19121d39e605beb372f1c53f06d994bf324b26e6c61a76ac2eaa76c408.jpg", "reference_inputs/images/520a20bec35d284cf13d94a8a4282fe05ae41fb70a996c3593515cca23ed45a4.jpg", "reference_inputs/images/02fa92474473ebe9f1a1fda3ba193c9d869369b4b68692e22ea9a78bfa93953f.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "A", "answer_index": 0}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000053", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000053", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the LEFT-facing camera, so the ground images look toward the teammate's left side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G2\nB: G6\nC: G3\nD: G5\nE: G7\nF: G4\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/594ef05f8c36b19e452b157aa0970518b4e56ef9ca807ed3555821b5e9a8b5a9.jpg", "reference_inputs/images/4cfed7a96247f40958fccc6a64d4c0870f3c3095186978f5cd49706332461ad9.jpg", "reference_inputs/images/6c19efe1630f6eca1709b8bf989300cf0ebc534fdb36b20b60aeb358ef605f59.jpg", "reference_inputs/images/75579460e06e62505caf9be8ee37c42a511634e15b9779869acef744a060d385.jpg", "reference_inputs/images/fdb606cadb3ed9dec989cac8afde68eac34da1d710796fc2e02eac839c823c29.jpg", "reference_inputs/images/b070879f6f83f57cf93c03702a572013de9ffec6c280cf1d8b371afbd67e4b7f.jpg", "reference_inputs/images/a08ba365f1665b5e5328292bcae275b989203d18db9bfb02c198e3b340b49c5f.jpg", "reference_inputs/images/51318467cb385ab61b209033e8b4a2d5fdb5d51777430e17e3a6c507567a175b.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "D", "answer_index": 3}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000057", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000057", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G4\nB: G3\nC: G6\nD: G5\nE: G2\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c505ea11f596c911a077f40458bec1e7fc5e9bcba7d7d3cb09eb626e141c79a8.jpg", "reference_inputs/images/614bea61ac9a60f84dd3186ff090ac89c59c5f19c19daa82db3fc2053a6cd2cb.jpg", "reference_inputs/images/b537d27ae67f2abe24e4fc75cba67f2e56273e13ed72686e305fec5a9130eb96.jpg", "reference_inputs/images/18af91d51cab25cd269148c24de3e6c462001bcbe4cdf2ad120569133edaefec.jpg", "reference_inputs/images/d3558a936a65613cbbec820d11b51cc2ac958af87155f8146abf0217b9691467.jpg", "reference_inputs/images/852baac5526a6621171f2c610f1452b58769ff01607d0541ea0b62c5c826c2fe.jpg", "reference_inputs/images/317ab9b52b0f25ddefae5a487bf7d0fe2d07557b135018f3c808b09f864b4bca.jpg", "reference_inputs/images/30142c91b4a377fd8bb01708b286c070258fd5c4a84e7771a9bec6c4b1def4c8.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "E", "answer_index": 4}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000058", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000058", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G3\nB: G6\nC: G2\nD: G5\nE: G7\nF: G4\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1049830325d87a2245c0ef51adf503796483c1de9f45b09ce6dc12599fd306e3.jpg", "reference_inputs/images/ccf2f534069fb5ee7c79971e7991be130cf08e6f3755a61c937f36f653b973c8.jpg", "reference_inputs/images/3e91e639215d5b63a541258e591be23d5a87fbebe2f57fbccf561df8154d2497.jpg", "reference_inputs/images/7934020d0894b6819bb95d004f542b8eb70efd00efd7cdb166a0e2f98a02f6f0.jpg", "reference_inputs/images/5aea1acb4606d1679dceeb5fbc13b2a0e6feb59803bc104fc692fb11fd244e8a.jpg", "reference_inputs/images/d670cb5449c0af4fb04ac1a94f3ce27ab2beba4c7ead8ef553f6021fcead4ddb.jpg", "reference_inputs/images/0cc5a1937e0c95eac9263f2c5b07755f969ac41381db0cd1bc25a07dffb78689.jpg", "reference_inputs/images/f934db43925ef1b3d6a6ca1afb809539d7ec7ad2202d78a326940d39e7aaf25b.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "F", "answer_index": 5}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000060", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000060", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G6\nB: G2\nC: G3\nD: G5\nE: G4\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/165cbe35b4f6bccddb7be46dbea14827b20afec21ed0992ba2aad75343c6f003.jpg", "reference_inputs/images/6294d811879d71885a913179fe579173a851d3f2de039e4ba15a9825209b3a95.jpg", "reference_inputs/images/5a27af426e8071277d41fa0fca00b264ac130475ccaf1dd9557379cfc0702f31.jpg", "reference_inputs/images/f8b96bdfbc62edc88f05d93287b7bcce7634eacffe424588fe3a13787206af43.jpg", "reference_inputs/images/84a8bbc0fe5ff4efe8a3e6a049b3903e2c2bd276c010cd43b70cfd360d7bf897.jpg", "reference_inputs/images/65b7e6f323bd65c5d7669cb2b5b8e28d3141f449e05c138b7d4ca072cfd32073.jpg", "reference_inputs/images/356e298d4b7742b6b7c826b061b76cb8db7acd128030789aca33d2c107501fd6.jpg", "reference_inputs/images/85e7535193a20ac9c72bf62a7ece488939b7e74c041620a63134e9fff1237c40.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "D", "answer_index": 3}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000062", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000062", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G2\nB: G3\nC: G5\nD: G7\nE: G4\nF: G6\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1e13b9f287f1138b07ee40c4f02303f661da341f11c0a721159c9c034157b63e.jpg", "reference_inputs/images/153c42acc503b7f49f9717f56313e3f98e87cf0cd70405ca91151993eafe3310.jpg", "reference_inputs/images/b28aa8aaad129d53f280c819325ac5c973dbd9c05b862f43953b1c6b3a2077d7.jpg", "reference_inputs/images/71c1d97e3384db44f94c215ca7826cfdc15574bf16d173d6c6f24d4872a16d89.jpg", "reference_inputs/images/f732fef1023b152fb2d8586c31d138f9774e49c0669823a575ceb8d45f09768a.jpg", "reference_inputs/images/b458fc1c79ce19d141010b566ef27d6d7d4f034b0f01d15dfc3ad0ca65fb4341.jpg", "reference_inputs/images/696db644fc0e650348f4be016e8e5fd9c181eee2de39d8b0f2f71ba4a8fde652.jpg", "reference_inputs/images/4b42b02e39f44393275ac5e9a5490bd7d17f5f4b37b6848eb0fc0c58462fd46a.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "C", "answer_index": 2}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000063", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000063", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G4\nB: G7\nC: G3\nD: G5\nE: G2\nF: G6\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d9050ca1e6fbe6357525108021a63ce6197d60a11f20b67e7ffd48c39606a773.jpg", "reference_inputs/images/92d0180982628f8a0f6c22980f4af4c90172b66fa59a007b5a04f60b2a1a9574.jpg", "reference_inputs/images/5c973ec6b63e4c1d2c149201ffa5cd3b1388a2ea9967a81b72eefd098be5952a.jpg", "reference_inputs/images/7a0bf23ec2ab312ed51cb9da293f10c6cf147291121dfec77e840170a2766309.jpg", "reference_inputs/images/29180798b6def462313b800f7cbd24174dba0446ce21c69b7e0ad3d3e8083450.jpg", "reference_inputs/images/0c8a9b920fb752d17854e01eaeabd648352c592e688d3546ef3830f74c8df359.jpg", "reference_inputs/images/cd76bcd55385041f93231333c73afa5c7080a078cd83c505f954450c93282456.jpg", "reference_inputs/images/c65191082fc8b97e9ab0c42da44ca2eb736d70349e3529af14528a6ea6decc4d.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "E", "answer_index": 4}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000067", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000067", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the FRONT-facing camera, so the ground images look roughly along the teammate's motion direction on the green aerial trajectory.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G7\nB: G4\nC: G5\nD: G6\nE: G3\nF: G2\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a0c899c432155e2a67e94360c0db55ea435f6a235de506b549e2ed9d5bb44a6c.jpg", "reference_inputs/images/09877c19121d39e605beb372f1c53f06d994bf324b26e6c61a76ac2eaa76c408.jpg", "reference_inputs/images/7e335e03c825200cce7a9e42c6bbfaa0e960ed5feec24233d18da53e6340c06c.jpg", "reference_inputs/images/02fa92474473ebe9f1a1fda3ba193c9d869369b4b68692e22ea9a78bfa93953f.jpg", "reference_inputs/images/c4b4361d5a1279566dac8a9d342747059a39a8d9a68a07eb976dc2ee90f536c2.jpg", "reference_inputs/images/1e7b3515bbc18e19e874005bb680a85b86b524dbafd95cdbb9a85308de0ac895.jpg", "reference_inputs/images/e7b7f59c7b63eefb38a9ac6bdee663ae7f5f7b8d38b61ca7f1812b245bb861f7.jpg", "reference_inputs/images/0e5ff0ab8d31164aded8f8513ac5bd67763baf668e9c2ee63493117ca582e2ce.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "F", "answer_index": 5}
{"sample_id": "reverse_sequence_reasoning/town1_ground_aerial_v2/000069", "task_name": "reverse_sequence_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/reverse_sequence_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000069", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Reverse Sequence Reasoning.\nYou will receive one aerial image followed by individually provided candidate ground-view images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your aerial view, where the true motion trajectory of your ground teammate is already drawn as a green path with a start marker.\nEach remaining image is exactly one candidate ground-view frame from that teammate, provided separately rather than stitched into a panel.\n- Image 2: candidate G1\n- Image 3: candidate G2\n- Image 4: candidate G3\n- Image 5: candidate G4\n- Image 6: candidate G5\n- Image 7: candidate G6\n- Image 8: candidate G7\n\nGround-camera orientation hint:\n- The ground teammate uses the RIGHT-facing camera, so the ground images look toward the teammate's right side relative to the motion direction on the green aerial trajectory, not straight along the path.\n- The aerial path direction describes how the ground teammate moves; the ground image content is what that mounted camera sees from its own facing direction.\n\nThe first candidate G1 is guaranteed to belong to the teammate's true trajectory sequence.\nThe remaining candidates must be compared against the drawn aerial trajectory and the scene layout.\nExactly one remaining candidate does not belong to the teammate's true trajectory sequence.\nUse the aerial trajectory layout, scene continuity, and viewpoint progression across the ground-view frames to identify the outlier.\n\nAnswer options:\nA: G2\nB: G6\nC: G5\nD: G3\nE: G4\nF: G7\n\nBenchmark mode instruction: respond with only one option letter from A, B, C, D, E, F. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2a2f5638fa857c45fcb70cbbd168cbe5ea6ba70e2837e46c19a98c9537695fc6.jpg", "reference_inputs/images/b222eeecc0a0b3660e5632c382b7f4bdbff21bc456f06d0165822026551dda65.jpg", "reference_inputs/images/b3326e82ebd0776424f0ffd76f9ec5461e4bb8834505c636702c283f24589fac.jpg", "reference_inputs/images/1e8dfe34369b30ada45b2492470778472ef1f107e14282b5d897e55a5f42301c.jpg", "reference_inputs/images/65b7e6f323bd65c5d7669cb2b5b8e28d3141f449e05c138b7d4ca072cfd32073.jpg", "reference_inputs/images/b3659a307c6248b2ff3d779a5a18fdfc2d903b7eec7988550335e97ebd3edd10.jpg", "reference_inputs/images/d63334c465eb308bedf9b92da88defca48fd5a31e60274ea5fa33125cb24724d.jpg", "reference_inputs/images/ccf2f534069fb5ee7c79971e7991be130cf08e6f3755a61c937f36f653b973c8.jpg"], "image_roles": ["marked_aerial_trajectory", "G1", "G2", "G3", "G4", "G5", "G6", "G7"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D", "E", "F"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000001", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [244, 503, 271, 530]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [236, 262, 255, 297]\nB: [528, 274, 602, 309]\nC: [684, 262, 703, 297]\nD: [319, 278, 386, 302]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0afbd9401679e8a5f9f739e851e2d6565ca0847966b9c870ff25e6bd5f0227c3.jpg", "reference_inputs/images/a37278d6f61a09212963b84767982a4ff7065abe6dae528d356c579d594d5b87.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000003", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [246, 453, 277, 500]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [845, 242, 878, 302]\nB: [470, 278, 489, 292]\nC: [537, 250, 579, 294]\nD: [311, 274, 377, 304]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/65bc64984b2063a131f48fe526cf6a7b07da1d78c38097687558bd599f26a609.jpg", "reference_inputs/images/326e309493b74656ab397a5c7fe4a7d782f7c238e5d6a88d3395323169ed9146.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000004", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [537, 250, 579, 294]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [178, 453, 197, 472]\nB: [246, 453, 277, 500]\nC: [218, 411, 231, 424]\nD: [156, 423, 176, 436]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/326e309493b74656ab397a5c7fe4a7d782f7c238e5d6a88d3395323169ed9146.jpg", "reference_inputs/images/65bc64984b2063a131f48fe526cf6a7b07da1d78c38097687558bd599f26a609.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000005", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [173, 237, 215, 308]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [766, 471, 796, 503]\nB: [142, 2, 178, 34]\nC: [680, 412, 704, 430]\nD: [779, 66, 830, 86]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d6fbe93a4f5e999ad00e816341ccc35818ab5a6bd7ce3f2b31a5f3a913b357d5.jpg", "reference_inputs/images/32ec1cfd3248914be2bb9127bbec2417bc640c23321fbce528d484cc3448edf0.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000009", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [650, 529, 663, 551]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [698, 265, 714, 294]\nB: [422, 275, 449, 293]\nC: [816, 247, 834, 301]\nD: [549, 279, 571, 293]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2dc7235177eab99e00758787c60e4230c1c1263feedc79745fdcd2eae1d0ed46.jpg", "reference_inputs/images/9ad8a50cccc69d4d19587748ca75a29e617773e0bff87ecdaec8ddeb22d583f1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000010", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [422, 275, 449, 293]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [542, 463, 560, 480]\nB: [439, 454, 460, 479]\nC: [522, 372, 537, 390]\nD: [506, 526, 532, 562]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9ad8a50cccc69d4d19587748ca75a29e617773e0bff87ecdaec8ddeb22d583f1.jpg", "reference_inputs/images/2dc7235177eab99e00758787c60e4230c1c1263feedc79745fdcd2eae1d0ed46.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000011", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [652, 382, 672, 397]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [670, 249, 684, 301]\nB: [372, 262, 381, 296]\nC: [989, 254, 1004, 300]\nD: [582, 269, 595, 292]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b1a49ca872c3f3278160aee709e43f8ca2626f153aea3e317ad1abac4661ac4d.jpg", "reference_inputs/images/df6c595233f3d9f15667bd1cfb2f3a7dea4b5caa3ea1cb1d6eec6223c48b2f2f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000012", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [458, 246, 478, 306]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [162, 46, 215, 70]\nB: [102, 565, 126, 575]\nC: [792, 474, 820, 499]\nD: [121, 475, 156, 499]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/55aca58644f2001dd1fad5ed84021bd3b993bd877b87301b5ec3265ac716a94a.jpg", "reference_inputs/images/6263899b1ea338624c0248869d9cfc22791beeb4099213fcee8cd99486bbd9a5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000013", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [381, 422, 497, 488]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [994, 236, 1023, 311]\nB: [541, 265, 790, 340]\nC: [167, 243, 232, 298]\nD: [703, 248, 717, 270]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/90f04a6a2238a7fdcc466fd00f9d19ae60103ef575e4a7d3a7eff0223c6367b0.jpg", "reference_inputs/images/a48a18e7653d72be7bef71d38f63063825f2392195ab9c9f2863e8d057499145.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000014", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [327, 553, 345, 575]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [52, 266, 63, 294]\nB: [262, 263, 276, 295]\nC: [482, 258, 493, 296]\nD: [700, 270, 706, 291]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b8792d16dea90b6ef0611eaee2ce16526d9ae336d106646657cea1b42a3c1d7a.jpg", "reference_inputs/images/82cfc3331a51c8535d78f6cbee980b1c6984fdb50a3c30fe64984dbe6581143a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000015", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [482, 258, 493, 296]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [652, 54, 673, 81]\nB: [327, 553, 345, 575]\nC: [942, 43, 992, 66]\nD: [506, 570, 542, 575]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/82cfc3331a51c8535d78f6cbee980b1c6984fdb50a3c30fe64984dbe6581143a.jpg", "reference_inputs/images/b8792d16dea90b6ef0611eaee2ce16526d9ae336d106646657cea1b42a3c1d7a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000016", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [670, 190, 692, 214]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [848, 265, 859, 294]\nB: [831, 264, 849, 294]\nC: [393, 266, 401, 294]\nD: [462, 266, 467, 283]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/37dd7261a32480af73cecd84b29f052454c744831f75ac89b5a7713ba47c9065.jpg", "reference_inputs/images/5962876e2b4cb80abcce4a9888010a061342e63f3d85987a6738be3e4f6b7af9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000017", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [174, 234, 195, 306]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [825, 112, 859, 131]\nB: [861, 565, 880, 575]\nC: [103, 101, 142, 132]\nD: [413, 10, 460, 116]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/feda76232f2405ca0f902f5e80fccb76864fe71c636e873aaad62eb1c082cd77.jpg", "reference_inputs/images/f2f80da1a6b22f4d549f719e57b16815e0a24c91e341527ac187a7ca38be6e9f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000018", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [422, 526, 427, 549]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [47, 194, 98, 326]\nB: [427, 263, 436, 294]\nC: [226, 266, 237, 294]\nD: [682, 257, 694, 298]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6e61ffda44d47d21044a3b977e73565628b58af41ddc02b1f2381d26eaf21934.jpg", "reference_inputs/images/85f9a3f775813d9252af0c381aed9c8cabe1ecb780d3f7ecf7b8bd71a58dfb2c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000019", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [122, 346, 163, 362]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [750, 258, 764, 298]\nB: [0, 239, 22, 305]\nC: [727, 254, 738, 296]\nD: [114, 243, 135, 302]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/27579f644325a8ad7d9b4eb3774ab00c1617860d245edf1122ec84095e57b823.jpg", "reference_inputs/images/6735f98cc46d273f973863ea4bee7b2468bdd114361efff6508e8164bd3036ee.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000020", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [0, 239, 22, 305]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [163, 294, 204, 311]\nB: [122, 346, 163, 362]\nC: [156, 432, 188, 456]\nD: [511, 0, 553, 26]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6735f98cc46d273f973863ea4bee7b2468bdd114361efff6508e8164bd3036ee.jpg", "reference_inputs/images/27579f644325a8ad7d9b4eb3774ab00c1617860d245edf1122ec84095e57b823.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000021", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [754, 112, 780, 137]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [384, 244, 398, 305]\nB: [21, 242, 49, 302]\nC: [213, 261, 222, 297]\nD: [388, 229, 433, 318]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d185a65e2420ed7d26f40dd0df1b158c836a094d8997559febbdf0997ce1e750.jpg", "reference_inputs/images/f4b724f9882422540f8c4dfcc45653e3be4fab7dffab2344dfeb96c6b92544c1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000022", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [524, 422, 539, 434]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [0, 247, 60, 280]\nB: [819, 279, 886, 297]\nC: [0, 276, 85, 298]\nD: [272, 265, 291, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e0381b03e29afa40e6d18828c0f36bc0d89bb5a74d329b1a9516c1a79e4b51a4.jpg", "reference_inputs/images/57e4694e7f958f730076926ae88c4407181e8cbbd683ad2fdc003377784f44e2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000024", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000024", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [465, 505, 490, 542]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [476, 277, 494, 291]\nB: [559, 270, 592, 297]\nC: [501, 271, 522, 309]\nD: [420, 281, 449, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/db9fc74c10b486ab63a0d92483697632bbdab6f70838191d395d4b86b21e4f3c.jpg", "reference_inputs/images/44166659bf064b7b7f17ebe1cd27d0ffcc7acc728691ab96f5c75bf2df465995.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000025", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [476, 277, 494, 291]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [509, 490, 529, 514]\nB: [465, 505, 490, 542]\nC: [554, 560, 585, 575]\nD: [503, 410, 525, 441]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/44166659bf064b7b7f17ebe1cd27d0ffcc7acc728691ab96f5c75bf2df465995.jpg", "reference_inputs/images/db9fc74c10b486ab63a0d92483697632bbdab6f70838191d395d4b86b21e4f3c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000026", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [688, 519, 695, 544]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [27, 225, 74, 314]\nB: [70, 238, 99, 305]\nC: [354, 251, 371, 302]\nD: [432, 273, 438, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a00ea25c32d5a5a8e9818c7c7d02f9e95b2ee0792e92f82461c929232b25ea4e.jpg", "reference_inputs/images/eb97a3246d862464d2b51b720f73d209b8d3a46bb24e54d9e5a2b762b5dfa240.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000027", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [70, 238, 99, 305]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [373, 532, 378, 557]\nB: [58, 560, 66, 575]\nC: [306, 544, 311, 571]\nD: [688, 519, 695, 544]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/eb97a3246d862464d2b51b720f73d209b8d3a46bb24e54d9e5a2b762b5dfa240.jpg", "reference_inputs/images/a00ea25c32d5a5a8e9818c7c7d02f9e95b2ee0792e92f82461c929232b25ea4e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000028", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [458, 422, 474, 433]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [592, 277, 642, 302]\nB: [0, 262, 219, 346]\nC: [472, 245, 594, 338]\nD: [437, 270, 473, 298]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ce8efefe6296b719d57192eb8ddd3beade6a537c2ce5be71960d4a9c54866577.jpg", "reference_inputs/images/360617ebd1ddf8668083527a4aeebbdda1341224f852b2297b1aa4e0edd6dbf3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000029", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [303, 249, 358, 357]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [112, 276, 135, 290]\nB: [192, 279, 217, 294]\nC: [0, 270, 42, 313]\nD: [151, 271, 170, 287]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cd224bd408613911ba79295c4e794772bb1a16a0158f3acba03ed8334def41b4.jpg", "reference_inputs/images/3f6b8fcca774c8e8aea257100830169913e8608cc0bcabc19239f3bc3545c962.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000031", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000031", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [765, 111, 791, 138]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [432, 242, 450, 307]\nB: [72, 230, 122, 314]\nC: [498, 235, 526, 314]\nD: [893, 268, 905, 299]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6938d96204b44fb7457ec43fc9176431bdb8d6e4aac5408010f2a3bfe0d6efcc.jpg", "reference_inputs/images/46a777e3a51d6c39f60dd75a2a9a9f7a3b523703d394b82437cbb2aba0121458.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000035", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000035", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [446, 278, 490, 304]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [516, 428, 532, 439]\nB: [817, 556, 885, 575]\nC: [502, 366, 515, 383]\nD: [424, 446, 459, 501]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b9ea434992c8ccf1bf64d5832baa3689874b9ca1ea387eab26fc72aa1fa2dad6.jpg", "reference_inputs/images/35cd9180f440e70c746afd7565e001662570370fa734d09ad27e609693f282e9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000037", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000037", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [710, 239, 730, 270]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [135, 78, 182, 100]\nB: [804, 410, 846, 427]\nC: [161, 388, 197, 406]\nD: [582, 566, 622, 575]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7c40dda52bb9866996ccd7d539527860ee133d45134812dab45eff791b1e2824.jpg", "reference_inputs/images/6cc5f2a32deb5f6d718ce65a2f84ff9863030ca7431c9373a3d6d1bd209ce13d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000038", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000038", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [391, 237, 433, 313]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [650, 519, 656, 541]\nB: [437, 458, 458, 476]\nC: [294, 539, 301, 566]\nD: [669, 549, 678, 574]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3096dc032ee4f1c1233f05b3ee86dd0b9222065ea734f6c8b6b07f00cf58a156.jpg", "reference_inputs/images/c59859cfddffb1b85b9b6f29d08d0ea93e860d12bcec3d1506330d3b849d4f6d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000039", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000039", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [164, 441, 206, 460]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [951, 234, 962, 246]\nB: [722, 250, 741, 302]\nC: [638, 271, 644, 293]\nD: [182, 266, 194, 300]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/642a532c0e078adda14cd9f0d316a540d319113490fbcf5da0243e65f50e01eb.jpg", "reference_inputs/images/59e4aa2a08930d64de52105fa7f3349672d21f92bf3adba38145d6c228691901.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000040", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000040", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [722, 250, 741, 302]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [842, 488, 889, 514]\nB: [13, 518, 61, 548]\nC: [329, 493, 353, 514]\nD: [164, 441, 206, 460]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/59e4aa2a08930d64de52105fa7f3349672d21f92bf3adba38145d6c228691901.jpg", "reference_inputs/images/642a532c0e078adda14cd9f0d316a540d319113490fbcf5da0243e65f50e01eb.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000042", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000042", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [861, 254, 909, 293]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [454, 432, 471, 443]\nB: [506, 367, 520, 385]\nC: [533, 422, 549, 434]\nD: [487, 405, 499, 419]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a66bc4c9128bd2588127e2d2ce36dee5e8c6a98ae97c206adeaf3f94258f6ee1.jpg", "reference_inputs/images/df59edf618f1768dd5184af7d6da5fda60eb9c250beb66b6d42ac32de9770ce0.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000043", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000043", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [729, 13, 754, 55]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [916, 218, 926, 246]\nB: [781, 238, 833, 258]\nC: [510, 238, 519, 265]\nD: [646, 227, 674, 255]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/12a5c1693e965c644508b6abab0f5fd0314a4bf1eadf2ef3d5612281e96bcd61.jpg", "reference_inputs/images/a0ccf0f5e696281230600960e2603ce51a994ea07386f34ba1410ded71be1b4c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000044", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000044", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [507, 434, 524, 446]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [913, 232, 945, 307]\nB: [438, 282, 460, 292]\nC: [0, 30, 268, 333]\nD: [484, 269, 593, 323]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9e3c55173c9e66d07f914d2a2284cea3ba284452cb7b3dc166d8cad737c5426c.jpg", "reference_inputs/images/ce6b08db9a41daef36af00206c758b526312c6cf41bef6347f4d72c299e9eb2b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000045", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000045", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [484, 269, 593, 323]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [334, 522, 350, 543]\nB: [447, 421, 474, 462]\nC: [537, 506, 580, 574]\nD: [507, 434, 524, 446]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ce6b08db9a41daef36af00206c758b526312c6cf41bef6347f4d72c299e9eb2b.jpg", "reference_inputs/images/9e3c55173c9e66d07f914d2a2284cea3ba284452cb7b3dc166d8cad737c5426c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000046", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000046", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [92, 330, 136, 348]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [158, 254, 173, 298]\nB: [680, 269, 686, 293]\nC: [689, 266, 710, 294]\nD: [829, 246, 873, 304]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ed6a292d6cd804eddcda4961ead9b4a2b5e9bae9f81b95a5aabdcde5d500b5d6.jpg", "reference_inputs/images/95c96102e8a9126a8dc2362e391865f4cccb2be1df7cbe5c3bfa84cbdc0dab96.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000048", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000048", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [454, 397, 475, 427]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [765, 262, 866, 326]\nB: [859, 274, 879, 298]\nC: [662, 272, 700, 296]\nD: [748, 269, 768, 290]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/df51f5738b4fdbb02a7e06ddc6bda1de2bcb65219e08638c3a53be5ee82e7c1c.jpg", "reference_inputs/images/74e818dd8ccefda5cdf4d1ea88f33d95a27373669fe205b437f32b3db295ef87.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000049", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000049", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [153, 230, 172, 273]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [806, 83, 848, 105]\nB: [158, 60, 195, 86]\nC: [581, 255, 631, 366]\nD: [776, 403, 807, 422]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5d308c10155d7a9b1cc5b63f8a401ec1ffd3310c2159cfd475b7a1aed61ff552.jpg", "reference_inputs/images/ced792b826b18846a3f799730c81f062334be9bfb26bca981869188d26fed4d2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000050", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000050", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [0, 262, 245, 362]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [506, 474, 526, 492]\nB: [410, 537, 443, 574]\nC: [478, 451, 494, 468]\nD: [550, 537, 577, 574]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2d0299f6a3360b525985d7cfcd2cc8a2488b02ea39fe7494abca0dba331a005c.jpg", "reference_inputs/images/a28580c0acd048e5e50dca36b438dbaa11ae1aafe0cfb9ab9367bcfe807cc7bf.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000051", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000051", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [793, 534, 806, 557]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [585, 258, 598, 296]\nB: [470, 263, 482, 294]\nC: [963, 262, 978, 295]\nD: [178, 271, 310, 307]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/eea0ba64382eff8b05e91d81cc247d4594f546377d9a5972b7845affdcbb97dd.jpg", "reference_inputs/images/29fc872b749e944ac734776f6892fe8e4600cf609cbd0a71f1ff2dda5f3dc4e4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000052", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000052", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [324, 438, 351, 450]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [86, 232, 367, 359]\nB: [658, 264, 673, 294]\nC: [378, 278, 410, 296]\nD: [510, 260, 538, 292]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bacd0526867e3021b3476ecac5fd7a0986826a48bc56cf026195ba716071923d.jpg", "reference_inputs/images/c45f03204bfb05c79b1c7600d3582de3ac706beb732d0c7fa825c450a31e4ec8.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000053", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000053", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [512, 0, 564, 93]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [399, 282, 434, 298]\nB: [318, 278, 354, 296]\nC: [374, 279, 399, 294]\nD: [169, 278, 272, 306]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f6a79034b298f8ff2f1b93aaa565a24635361b61b7454da8eb86027d450c4efa.jpg", "reference_inputs/images/136f851f4ce6fede76cc50fdc3fc6048d01d9a5ec12763732777c13fbb606be9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000054", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000054", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [399, 282, 434, 298]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [512, 0, 564, 93]\nB: [667, 402, 730, 514]\nC: [487, 318, 538, 429]\nD: [326, 470, 346, 494]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/136f851f4ce6fede76cc50fdc3fc6048d01d9a5ec12763732777c13fbb606be9.jpg", "reference_inputs/images/f6a79034b298f8ff2f1b93aaa565a24635361b61b7454da8eb86027d450c4efa.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000057", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000057", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [914, 517, 923, 538]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [865, 257, 882, 306]\nB: [718, 271, 725, 297]\nC: [345, 270, 352, 295]\nD: [380, 267, 383, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/eb1a89bc11789e99b19e77585fd46e9e36fae0d3273a75e9ebfa9d805ddb8203.jpg", "reference_inputs/images/9c6b3f29f4d9176113b4baac3a280ea580c2ac94190eae813932e337951513d2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000058", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000058", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [454, 242, 480, 310]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [135, 78, 182, 100]\nB: [161, 388, 197, 406]\nC: [804, 410, 846, 427]\nD: [582, 566, 622, 575]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/73761be65818a5dc9026c4b691d72ee86c5940efe951d9acb79754e03601e57f.jpg", "reference_inputs/images/6cc5f2a32deb5f6d718ce65a2f84ff9863030ca7431c9373a3d6d1bd209ce13d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000059", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000059", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [759, 432, 787, 458]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [0, 245, 238, 400]\nB: [497, 221, 529, 317]\nC: [237, 278, 271, 293]\nD: [555, 239, 570, 310]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b45e5d4a35ce9f457d62b66c0c04b331a9dc27a35f45653cbb0002e4d287a1e0.jpg", "reference_inputs/images/425cf972cabd45e0522c7a870edb16f10b0dacffe2cf1406bcec77a1b396b7a9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000060", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000060", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [555, 239, 570, 310]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [841, 469, 880, 493]\nB: [510, 155, 537, 171]\nC: [759, 432, 787, 458]\nD: [171, 192, 206, 209]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/425cf972cabd45e0522c7a870edb16f10b0dacffe2cf1406bcec77a1b396b7a9.jpg", "reference_inputs/images/b45e5d4a35ce9f457d62b66c0c04b331a9dc27a35f45653cbb0002e4d287a1e0.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000061", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000061", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [195, 548, 205, 574]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [291, 270, 302, 292]\nB: [205, 268, 218, 292]\nC: [673, 268, 679, 293]\nD: [955, 228, 997, 310]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a4cf592a72e7c039e89f2ce68a4f098467de07ca020316c2c156372cb9a1cfc5.jpg", "reference_inputs/images/9762b261fdab1ae82f5704a59390f7a7a5f789c680e4cf37c3c11ba8d2d2fc0f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000062", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000062", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [500, 270, 517, 289]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [534, 448, 550, 465]\nB: [499, 388, 517, 412]\nC: [280, 538, 297, 561]\nD: [475, 412, 490, 423]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9762b261fdab1ae82f5704a59390f7a7a5f789c680e4cf37c3c11ba8d2d2fc0f.jpg", "reference_inputs/images/a4cf592a72e7c039e89f2ce68a4f098467de07ca020316c2c156372cb9a1cfc5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000063", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000063", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [480, 355, 492, 372]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [446, 282, 480, 290]\nB: [381, 274, 420, 298]\nC: [469, 271, 486, 289]\nD: [0, 257, 188, 352]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/54a45b98db6cf9dcae184341b11c3ada9efb63b5c8c0222a21ea269ead7746f1.jpg", "reference_inputs/images/22e82f3f5da3ca8e0a41507973a1054fc81a0436bd364afb7118546fe3ec6147.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000064", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000064", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [0, 257, 188, 352]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [446, 446, 463, 464]\nB: [542, 480, 563, 504]\nC: [507, 438, 524, 450]\nD: [406, 557, 439, 575]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/22e82f3f5da3ca8e0a41507973a1054fc81a0436bd364afb7118546fe3ec6147.jpg", "reference_inputs/images/54a45b98db6cf9dcae184341b11c3ada9efb63b5c8c0222a21ea269ead7746f1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000065", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000065", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [494, 430, 516, 466]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [167, 252, 188, 296]\nB: [469, 223, 529, 299]\nC: [390, 267, 424, 285]\nD: [343, 273, 424, 306]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0338e6668ccc2980e70ebd4a9e32d2deccdc84389a69907b91810ac1184821af.jpg", "reference_inputs/images/a4ff943138300e14df5dde6b5b4d48e1550533ac77ffe671ed28e02e96353024.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000067", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000067", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [501, 266, 547, 367]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [94, 294, 125, 304]\nB: [122, 291, 157, 310]\nC: [273, 290, 321, 312]\nD: [190, 288, 218, 305]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/439f9ab18292fe5eaa909c9b4e29dd00f69e626bd2c9b4919ec2295bb72369d5.jpg", "reference_inputs/images/02ee26196f052db1fd6094389ca8a203753d4dfdd0196d5aad127f63ffb772e3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000068", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000068", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [442, 521, 471, 554]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [425, 280, 462, 297]\nB: [486, 250, 522, 294]\nC: [594, 262, 678, 312]\nD: [765, 258, 784, 297]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/387eee192cc021d94a93d3ba27331b505992d6b714b2291e6d3f1b53665fad27.jpg", "reference_inputs/images/37fdf1cfa1885a50dc1e3a45f4b4196e30f9e8011ea8b00b2aa25e97696ee22d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000069", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000069", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [478, 362, 490, 381]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [544, 236, 606, 298]\nB: [0, 194, 231, 416]\nC: [438, 260, 468, 292]\nD: [506, 282, 538, 291]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7abe8bf0e8316b5c50cf7a1af93b7def29fa22d68678887ccb5eace7bbf814b0.jpg", "reference_inputs/images/182266c2f040fdee6aaaa0e2c96f76eb54a0cd36c40e98e462f0c0ffbd783088.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000070", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000070", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [544, 236, 606, 298]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [551, 514, 575, 542]\nB: [523, 384, 540, 406]\nC: [480, 446, 495, 462]\nD: [478, 362, 490, 381]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/182266c2f040fdee6aaaa0e2c96f76eb54a0cd36c40e98e462f0c0ffbd783088.jpg", "reference_inputs/images/7abe8bf0e8316b5c50cf7a1af93b7def29fa22d68678887ccb5eace7bbf814b0.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000071", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000071", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [652, 434, 672, 454]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [515, 274, 520, 299]\nB: [880, 270, 890, 301]\nC: [946, 271, 954, 291]\nD: [470, 267, 480, 303]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6a55bff1f7c42f88c8c7cc62ca6d6d44adac0bb5ea56b80b68120229ee6b5703.jpg", "reference_inputs/images/3e209e0da429c596a0bcf30b747b605a4d443dbf2a00e34eee9b83629e766597.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000072", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000072", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [268, 269, 275, 294]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [713, 515, 736, 542]\nB: [164, 122, 198, 141]\nC: [98, 82, 134, 106]\nD: [758, 78, 784, 102]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5656546a0d9e30f3637c2bbbabe2b6fcd0e76f79f385253514533b4f71a004c0.jpg", "reference_inputs/images/8771c85ad7e20bee39b6eb3be777a3478801bbd3fe47bf6dc3ea6bcf510f6924.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000074", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000074", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [502, 270, 558, 309]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [484, 426, 498, 439]\nB: [509, 427, 522, 442]\nC: [550, 549, 596, 575]\nD: [510, 466, 526, 485]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b44ca937189d4b9529c968b0f3bea457bb9b34d3c8f1a0061bfd35bda0a95e8c.jpg", "reference_inputs/images/09d6bbd65d5060eae93487145b4ecd97a46eec4da32909caff0ec56456d840de.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000078", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000078", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [328, 180, 351, 194]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [518, 272, 525, 294]\nB: [398, 214, 440, 347]\nC: [539, 235, 560, 311]\nD: [440, 251, 458, 304]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0270d8b538279fc1cc2d42738a8e13de18ac9681608ba8b8b64e1d8f6abef7bd.jpg", "reference_inputs/images/7bb72823dd1ff767bcfa0af18ffdf886134eb00fa75108eafdef971bd5ab925b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000079", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000079", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [539, 235, 560, 311]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [843, 172, 891, 189]\nB: [785, 432, 826, 449]\nC: [162, 121, 205, 139]\nD: [10, 205, 58, 222]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7bb72823dd1ff767bcfa0af18ffdf886134eb00fa75108eafdef971bd5ab925b.jpg", "reference_inputs/images/0270d8b538279fc1cc2d42738a8e13de18ac9681608ba8b8b64e1d8f6abef7bd.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000080", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000080", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [163, 294, 204, 311]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [106, 250, 123, 300]\nB: [623, 264, 633, 294]\nC: [606, 262, 614, 293]\nD: [14, 248, 32, 302]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/27579f644325a8ad7d9b4eb3774ab00c1617860d245edf1122ec84095e57b823.jpg", "reference_inputs/images/54274a9910117326d71f3e76d951de9920b2bf270df85509c72aa19302d45c54.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000084", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000084", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [530, 448, 546, 465]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [190, 263, 225, 291]\nB: [767, 232, 791, 306]\nC: [342, 262, 424, 314]\nD: [254, 274, 316, 295]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c18ed58b26d234fb469fd862a56269881f2ef4151f5fe299c88cbab19df890eb.jpg", "reference_inputs/images/bf6e1f1858c88fb977f0dbc32936d9c13cdd676f51a1e07b64daf08ce71bb42b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000085", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000085", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [190, 263, 225, 291]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [442, 442, 461, 464]\nB: [474, 361, 486, 379]\nC: [504, 396, 529, 406]\nD: [530, 448, 546, 465]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bf6e1f1858c88fb977f0dbc32936d9c13cdd676f51a1e07b64daf08ce71bb42b.jpg", "reference_inputs/images/c18ed58b26d234fb469fd862a56269881f2ef4151f5fe299c88cbab19df890eb.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000088", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000088", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [826, 227, 858, 320]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [748, 303, 774, 317]\nB: [754, 0, 777, 29]\nC: [91, 407, 138, 427]\nD: [278, 91, 371, 306]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/794cfbcbe7aebf0fd07de502de177b39301e6da824fb87e6bb1e4aa716ca8342.jpg", "reference_inputs/images/e8090a6bd35271c5a185d4f4c043b2a50d9489f7c572b9f626ef642057ddfeca.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000093", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000093", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [609, 218, 1023, 386]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [423, 481, 449, 512]\nB: [503, 414, 532, 427]\nC: [534, 490, 555, 514]\nD: [469, 370, 484, 391]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9f8bd7afecd76acd880fa4f00ef12b38a4ea1e2d76f817d9fdbce48abab400e1.jpg", "reference_inputs/images/6d76248e7087fb5fcf6decf03cdcb4e685fdbb05d11f13983b35973a1a64e154.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000095", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000095", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [453, 242, 652, 320]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [502, 439, 536, 456]\nB: [542, 557, 569, 575]\nC: [465, 382, 480, 406]\nD: [395, 545, 427, 575]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4d94ae80e8ba04008e857d425712d1b5485e3b7c1903015d02007d4e2f266833.jpg", "reference_inputs/images/ff9f872d28e676633b1aaec5c24f8e3560e55512fa951d2b82602e48a167bf6b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000096", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000096", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [726, 265, 751, 278]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [639, 270, 646, 292]\nB: [629, 268, 634, 292]\nC: [0, 251, 22, 300]\nD: [274, 261, 282, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ed35b535eb965f06ad02d8c481285efa318082387be5b3764f24f1a66bd87735.jpg", "reference_inputs/images/40454a7f6ef23fad462d8d4428d9b62cdf2ce712eeaa75846a35968ac6fb3883.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000097", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000097", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [639, 270, 646, 292]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [726, 265, 751, 278]\nB: [154, 498, 194, 526]\nC: [766, 0, 798, 28]\nD: [756, 235, 783, 249]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/40454a7f6ef23fad462d8d4428d9b62cdf2ce712eeaa75846a35968ac6fb3883.jpg", "reference_inputs/images/ed35b535eb965f06ad02d8c481285efa318082387be5b3764f24f1a66bd87735.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000098", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000098", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [481, 418, 494, 436]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [0, 193, 336, 389]\nB: [450, 280, 497, 295]\nC: [1000, 223, 1023, 314]\nD: [506, 272, 542, 300]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f74c703367de7b2fc1295542d39d0c937f0335d7cc3c1046ad03f832839e7815.jpg", "reference_inputs/images/b84ea74379a0dc35ed4efa117f79ffd3eeacf5644d4810946141066cea215ba7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000099", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000099", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [0, 193, 336, 389]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [505, 484, 524, 508]\nB: [481, 418, 494, 436]\nC: [429, 490, 454, 512]\nD: [527, 442, 560, 456]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b84ea74379a0dc35ed4efa117f79ffd3eeacf5644d4810946141066cea215ba7.jpg", "reference_inputs/images/f74c703367de7b2fc1295542d39d0c937f0335d7cc3c1046ad03f832839e7815.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000100", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000100", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [503, 423, 517, 438]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [502, 277, 518, 290]\nB: [752, 260, 762, 294]\nC: [715, 257, 723, 297]\nD: [295, 262, 450, 326]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/369d8c5a94fad57873a397e17382502d73f237070752a22f9e7d59767de6f1e1.jpg", "reference_inputs/images/ff8fa66cbab0447fe511d276663e38da9d137919753d5a15e75769603d65bef0.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000102", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000102", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [403, 235, 424, 308]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [752, 544, 779, 575]\nB: [145, 251, 179, 264]\nC: [772, 384, 808, 402]\nD: [820, 230, 858, 244]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e3e71bc74207124f690a045d4f363510b52853f71bce6a3fedb33ecdb53106b2.jpg", "reference_inputs/images/d58fb1fe47410e3bdcd354c8c21eec20c0c805defa3cc2a24e2f04e9db9bccdb.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000104", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000104", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [764, 326, 797, 341]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [723, 259, 736, 297]\nB: [709, 255, 720, 298]\nC: [672, 265, 686, 294]\nD: [281, 266, 286, 293]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9708bca37386ab6c66ab76e39de82d9e427280bf7e372a1ad40b0a8b541f6a24.jpg", "reference_inputs/images/56d060aef0e030bd67d7d62aafc74f5f6e502ebd2aa875be625aa56703cb2184.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000105", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000105", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [330, 533, 336, 558]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [450, 248, 467, 308]\nB: [32, 242, 54, 310]\nC: [410, 253, 430, 306]\nD: [516, 246, 534, 306]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/32f5b5d0605f2bbeca2a3a4b5930b7b534552a65ced65cb6367ae6531b6d7770.jpg", "reference_inputs/images/9bd1be9135b490482662595471ecb13389cc21f0185e9afd81e9b1cf2152433c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000106", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000106", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [501, 376, 516, 396]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [520, 280, 554, 296]\nB: [210, 206, 374, 310]\nC: [498, 267, 518, 290]\nD: [705, 261, 714, 296]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d0c7a1e600af97e4214a01d83668ccb9038b75752431c9384f9b153978205fed.jpg", "reference_inputs/images/6c5bcc96e9ec88143a2af460a2c36c62a89566b5824bb951b130bc5915679c3b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000107", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000107", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [459, 407, 486, 418]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [510, 280, 526, 292]\nB: [546, 239, 606, 298]\nC: [424, 257, 463, 293]\nD: [264, 275, 405, 308]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ad7e06d19be14c7572094fcd2a8656811827fd7f8c582feb9f85922ac8d4a358.jpg", "reference_inputs/images/bb1234fc75db8b2004f6e5dad03553677922756abe73a2bc76ac3db2608e9339.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000108", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000108", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [546, 239, 606, 298]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [506, 472, 527, 490]\nB: [476, 363, 490, 382]\nC: [429, 486, 451, 511]\nD: [526, 378, 540, 398]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bb1234fc75db8b2004f6e5dad03553677922756abe73a2bc76ac3db2608e9339.jpg", "reference_inputs/images/ad7e06d19be14c7572094fcd2a8656811827fd7f8c582feb9f85922ac8d4a358.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000111", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000111", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [822, 434, 857, 452]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [58, 216, 98, 316]\nB: [557, 232, 586, 318]\nC: [800, 238, 822, 310]\nD: [63, 255, 77, 306]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ab2188839eca588e14ececd553c05ab309b3db806f153f7d78edfdf9250a16d7.jpg", "reference_inputs/images/4ea6834c85dd9da4d0e393b1fbd27477d525a255f2c04252747fb6893e6cca2b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000112", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000112", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [58, 216, 98, 316]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [106, 422, 143, 453]\nB: [822, 434, 857, 452]\nC: [753, 274, 786, 287]\nD: [748, 23, 778, 51]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4ea6834c85dd9da4d0e393b1fbd27477d525a255f2c04252747fb6893e6cca2b.jpg", "reference_inputs/images/ab2188839eca588e14ececd553c05ab309b3db806f153f7d78edfdf9250a16d7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000114", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000114", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [874, 446, 918, 466]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [0, 246, 127, 390]\nB: [906, 247, 934, 308]\nC: [963, 222, 1003, 319]\nD: [558, 224, 586, 316]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/80411caf2c694c923da91acf9ce65757d6d0b3cc4e9f32a494dc213f2e0988f9.jpg", "reference_inputs/images/092ab14103d7daee4393542be420b50621530d158741b01471fd752dac1a81a2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000115", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000115", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [963, 222, 1003, 319]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [874, 446, 918, 466]\nB: [148, 172, 192, 189]\nC: [759, 385, 788, 402]\nD: [146, 413, 180, 436]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/092ab14103d7daee4393542be420b50621530d158741b01471fd752dac1a81a2.jpg", "reference_inputs/images/80411caf2c694c923da91acf9ce65757d6d0b3cc4e9f32a494dc213f2e0988f9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000116", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000116", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [118, 163, 154, 182]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [946, 246, 977, 306]\nB: [795, 243, 818, 305]\nC: [358, 253, 392, 307]\nD: [0, 210, 186, 359]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/34f9529c1e00a701b20a09b60bd22d48dce47dca4e7396f61530ca5e4a3ccfad.jpg", "reference_inputs/images/8d95f0b46b9a39bb273e3130d0a98cce355bab03a4458d815689821b0b28e7fa.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000117", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000117", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [946, 246, 977, 306]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [795, 287, 823, 305]\nB: [76, 253, 124, 270]\nC: [118, 163, 154, 182]\nD: [793, 163, 822, 182]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8d95f0b46b9a39bb273e3130d0a98cce355bab03a4458d815689821b0b28e7fa.jpg", "reference_inputs/images/34f9529c1e00a701b20a09b60bd22d48dce47dca4e7396f61530ca5e4a3ccfad.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000118", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000118", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [756, 235, 783, 249]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [729, 233, 738, 323]\nB: [626, 228, 654, 318]\nC: [181, 223, 219, 315]\nD: [382, 265, 391, 296]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ed35b535eb965f06ad02d8c481285efa318082387be5b3764f24f1a66bd87735.jpg", "reference_inputs/images/8774b1de6e23c1056d6e4a08a53dfa05f070f79847d31a6c33221aded7e686e7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000119", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000119", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [181, 223, 219, 315]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [154, 498, 194, 526]\nB: [726, 265, 751, 278]\nC: [766, 0, 798, 28]\nD: [756, 235, 783, 249]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8774b1de6e23c1056d6e4a08a53dfa05f070f79847d31a6c33221aded7e686e7.jpg", "reference_inputs/images/ed35b535eb965f06ad02d8c481285efa318082387be5b3764f24f1a66bd87735.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000120", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000120", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [408, 529, 441, 573]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [246, 254, 342, 313]\nB: [494, 271, 523, 296]\nC: [827, 247, 845, 303]\nD: [0, 246, 259, 361]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/334b4e4d96aa2cc5e2ef412bf4ca996a78d03fc955f3ccf5b8987a427f0f5fb7.jpg", "reference_inputs/images/9d12e0d7e2f66043454da17dd4d6bf2b971c6c1620f747bf4be58db60909059e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000123", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000123", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [294, 249, 646, 357]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [316, 50, 375, 162]\nB: [589, 333, 643, 442]\nC: [328, 526, 372, 575]\nD: [588, 0, 645, 49]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/afa71b94a423f42d3007c87f85509f0adae7353c2399be4ceeb69c17f0d87d6c.jpg", "reference_inputs/images/a636870c7e4fd90a63cb02038fa2ea19d346afd5a60073645021f783b87298d8.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000124", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000124", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [318, 360, 376, 473]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [390, 280, 431, 297]\nB: [485, 248, 644, 355]\nC: [467, 275, 491, 293]\nD: [229, 269, 338, 313]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b355ce4a747caad0252734b0d1a30745dc9d99b13180da815406c6aa9dbfb60e.jpg", "reference_inputs/images/64aea7386a93b8d411c31c3c476f955f4bb9d442dea86752dfdbfd3a90eec182.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000126", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000126", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [158, 451, 201, 470]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [782, 255, 797, 298]\nB: [829, 243, 846, 301]\nC: [181, 255, 198, 298]\nD: [496, 246, 535, 295]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/35e859c28bf029e3e97f5365979050dd06717e2c31f0c6b35566c29e99c74a4e.jpg", "reference_inputs/images/64ecb14e96931fa12d63d7d52e97e6551d9ab8794bec873526ad42f1cf64fd39.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000128", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000128", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [446, 242, 581, 343]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [570, 471, 594, 489]\nB: [503, 456, 520, 476]\nC: [536, 558, 566, 575]\nD: [512, 421, 526, 438]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/21f2c320c1e570119f3c51a4276d427d1289f16ef62a1d359f7a1f4beaf17ca3.jpg", "reference_inputs/images/35717c2dd7162b69eeec0d1cc1f01170ab4debdc09307139ab4bc23ae14cb0cf.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000130", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000130", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [528, 400, 545, 427]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [460, 267, 479, 290]\nB: [532, 258, 564, 292]\nC: [0, 211, 64, 350]\nD: [422, 274, 467, 303]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a60a62748a3f2e935195f4fb999c581d2c0846a7a84952173733ba07423749b3.jpg", "reference_inputs/images/1b5c7d35d99b4d90fc9acbf22dad919f289e379585dda663df500ef2c2dbccf2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000131", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000131", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [137, 389, 179, 404]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [822, 250, 838, 301]\nB: [60, 250, 78, 301]\nC: [197, 256, 208, 296]\nD: [259, 263, 286, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5de87c83d655713d33a81d292bf084f128ce53e77ca6e352e4e381bbced8fb64.jpg", "reference_inputs/images/fbdc39cc7686c6098f4439103a97584a6a002a045d6e03bb084cdca90a855d25.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000132", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000132", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [183, 486, 216, 513]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [218, 233, 246, 314]\nB: [930, 272, 938, 292]\nC: [574, 241, 594, 268]\nD: [239, 248, 254, 308]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6af64a6f155a53fd7e9405fd0240f3d88f46f26297b3a634601d63726b1edce1.jpg", "reference_inputs/images/5c64e995683b1de21c283c671349dac82351d1b3a64375df3ccdb6bffaedb802.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000133", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000133", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [574, 241, 594, 268]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [754, 55, 779, 87]\nB: [183, 486, 216, 513]\nC: [100, 519, 138, 545]\nD: [162, 274, 183, 287]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5c64e995683b1de21c283c671349dac82351d1b3a64375df3ccdb6bffaedb802.jpg", "reference_inputs/images/6af64a6f155a53fd7e9405fd0240f3d88f46f26297b3a634601d63726b1edce1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000134", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000134", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [432, 509, 441, 530]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [331, 247, 350, 307]\nB: [645, 239, 671, 310]\nC: [617, 247, 656, 298]\nD: [678, 186, 737, 336]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9c7047075779686d932e90770dddff6b67c27ed2467b2c66c6e22ded141760a6.jpg", "reference_inputs/images/7d0ada91494a357afe62a3dcacc31e5200312464c1c5d46f241f69acbb88744a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000135", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000135", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [422, 223, 461, 316]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [122, 443, 161, 462]\nB: [766, 0, 795, 29]\nC: [812, 295, 846, 313]\nD: [168, 341, 208, 355]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/130cd799e21b96c4257f9d4cba36406d83b16b9eb81a6f1ae7913f5be9c9bbad.jpg", "reference_inputs/images/9b07624335b4fa633b68a1113aa88dc561c5ac1dcdfee58fdfa373ca97d99437.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000136", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000136", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [574, 494, 597, 516]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [769, 277, 802, 298]\nB: [274, 258, 289, 299]\nC: [698, 280, 722, 294]\nD: [942, 261, 954, 296]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8bc8e89b82da28dd3eb342c5c2d8ae3aa6bd8756a9bcf90ee17dd356cdad3e46.jpg", "reference_inputs/images/f84e51c89c7aaeb4f164fd1217f5b807fcbcb5695f418ece8dc950712607074b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000137", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000137", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [644, 242, 668, 281]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [711, 46, 744, 70]\nB: [121, 475, 156, 499]\nC: [162, 46, 215, 70]\nD: [102, 565, 126, 575]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b7087cf8dfff20392c5cce16fb1cbc197f4b8a023fd6a1a14e127507670e1033.jpg", "reference_inputs/images/6263899b1ea338624c0248869d9cfc22791beeb4099213fcee8cd99486bbd9a5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000139", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000139", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [10, 205, 58, 222]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [294, 252, 312, 305]\nB: [110, 234, 140, 305]\nC: [1, 216, 46, 350]\nD: [330, 237, 350, 312]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0270d8b538279fc1cc2d42738a8e13de18ac9681608ba8b8b64e1d8f6abef7bd.jpg", "reference_inputs/images/3e7836c5acb2dcdcaa8ae2df4de0058e570a9f85222862e2d16221447f8c8257.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000141", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000141", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [173, 488, 186, 506]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [167, 243, 189, 311]\nB: [1016, 262, 1023, 295]\nC: [886, 266, 898, 294]\nD: [951, 266, 968, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/698a1f1fea025f2ef0067d251e43da911ee6143b00cb60b7e052abd12237e7ba.jpg", "reference_inputs/images/d9f7b70cbce2f492858416fca652a2814092d4f954affedf5bcbcae877e1df12.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000143", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000143", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [344, 523, 349, 548]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [178, 263, 204, 295]\nB: [17, 242, 42, 302]\nC: [336, 267, 342, 292]\nD: [481, 242, 525, 297]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8ff5853e0bd8ef95d064df272a19356897f599fdd94098d87eba1a836939d593.jpg", "reference_inputs/images/f67eec63d5c0c61458bbae0766c70b18ecdbfa0cb0affc46531d9447a2b74645.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000144", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000144", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [517, 459, 537, 474]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [754, 254, 767, 299]\nB: [522, 279, 562, 298]\nC: [32, 167, 330, 324]\nD: [497, 266, 518, 290]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d0c7a1e600af97e4214a01d83668ccb9038b75752431c9384f9b153978205fed.jpg", "reference_inputs/images/135d5add10ec73a8f8dccd1efdf99084b9cf438829994e55973c598984e0abbc.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000146", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000146", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [505, 484, 524, 508]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [271, 275, 366, 308]\nB: [494, 266, 544, 306]\nC: [456, 274, 482, 294]\nD: [544, 279, 603, 297]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f74c703367de7b2fc1295542d39d0c937f0335d7cc3c1046ad03f832839e7815.jpg", "reference_inputs/images/e4d923360efb0a11f4dcd6ff2db5b67792540a723ae2ce828c607a4bb8d7e216.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000147", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000147", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [271, 275, 366, 308]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [481, 418, 494, 436]\nB: [429, 490, 454, 512]\nC: [505, 484, 524, 508]\nD: [527, 442, 560, 456]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e4d923360efb0a11f4dcd6ff2db5b67792540a723ae2ce828c607a4bb8d7e216.jpg", "reference_inputs/images/f74c703367de7b2fc1295542d39d0c937f0335d7cc3c1046ad03f832839e7815.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000148", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000148", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [781, 179, 811, 197]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [440, 210, 473, 322]\nB: [521, 225, 549, 318]\nC: [275, 224, 306, 313]\nD: [968, 281, 1023, 304]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1bd832572b4b448ad338c0f4e1a79903adb20e89675c701d2d4d8178f28548e0.jpg", "reference_inputs/images/16b98ca976b6de186d99c34f240152f162895e25eef46ec2f005aaf3034c6c1a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000150", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000150", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [58, 245, 100, 314]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [774, 82, 806, 105]\nB: [146, 536, 181, 562]\nC: [753, 536, 786, 563]\nD: [703, 463, 734, 488]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6bde287cf09eccf13f1a7282d5f5bbecd74d66d92d3f8b16409fa292a063c87d.jpg", "reference_inputs/images/a48ac53625eae2be6e6ab914a9040ecbd84ad8ba0efa5bb37f0c2237317ba31f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000151", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000151", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [746, 219, 769, 230]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [380, 275, 398, 316]\nB: [522, 296, 529, 319]\nC: [538, 302, 554, 321]\nD: [614, 318, 637, 330]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1ee50cfdaac707213827426bf2a39ffa6b1fbf9a114cbcb4670402f1cb9ae1fc.jpg", "reference_inputs/images/1bf7d9011d395be89f19ecb09013d2176af7238d43a34260ff514796416a497c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000152", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000152", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [274, 245, 292, 304]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [226, 83, 259, 102]\nB: [164, 44, 210, 64]\nC: [811, 182, 843, 197]\nD: [554, 456, 570, 477]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9bc1dcf84743d3725a7ba6f9333e43c2e3c72e853250d06fcf1cc8ff12eac131.jpg", "reference_inputs/images/b31703345ab53831ede95ce678077e9c64ec313f7611b3342fce89c5672d0399.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000154", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000154", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [834, 254, 846, 298]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [143, 161, 178, 177]\nB: [144, 386, 180, 402]\nC: [864, 513, 909, 542]\nD: [755, 113, 782, 133]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f715261c31086ae963af6c95167737fcc81cedadfdf0f8460aaa5262d2a8969f.jpg", "reference_inputs/images/5fcef0ed1e787fc5faa2ef566001887a3b83ebedf6e08c51a36e92a864bcc37c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000156", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000156", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [426, 246, 445, 307]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [158, 162, 191, 182]\nB: [122, 262, 162, 276]\nC: [83, 326, 127, 342]\nD: [163, 362, 200, 376]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/53f6413a5aece953d3d3f19040ae4fcc8e6ef179c92e51b1341438a643b7e920.jpg", "reference_inputs/images/2cc8a54174351d0059fbddcddde72408fcaefe0c3a706b17b11c8bb9852967bf.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000158", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000158", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [382, 250, 733, 383]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [474, 373, 487, 395]\nB: [505, 526, 532, 555]\nC: [528, 392, 545, 416]\nD: [400, 551, 431, 575]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1c045cc9d56ac5597dd52c6857718f20cd7716c7605e6a518351ebcde2bc3b2d.jpg", "reference_inputs/images/6a5e13694f5f116e2e93f37ae27001d2424ab7e44cf25ed4c982edd446cbccd4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000160", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000160", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [505, 267, 554, 306]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [436, 462, 455, 483]\nB: [508, 556, 534, 575]\nC: [642, 514, 658, 535]\nD: [502, 436, 527, 475]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2c4d8a2bd0304a00d3fcdc28715ad25aedb7c02628a93f4a466fc02ebe7ae0fe.jpg", "reference_inputs/images/4dabf19c91624ce95cd33715a4547c90755ab3181ee934441d405dcb02c76b3b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000161", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000161", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [893, 297, 1001, 344]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [384, 279, 407, 294]\nB: [305, 282, 351, 298]\nC: [461, 266, 483, 282]\nD: [430, 274, 488, 311]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9244160e155b63bdbce63c0444e48588782c5415701fca6dadd7b04750cdf985.jpg", "reference_inputs/images/43b8a14348f565382ab0cc4fbbf8e087a90487c0764d3c3e5bec20884a3cdbb8.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000162", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000162", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [59, 234, 106, 312]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [791, 109, 828, 126]\nB: [173, 506, 207, 534]\nC: [507, 474, 540, 494]\nD: [326, 0, 371, 18]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/43b8a14348f565382ab0cc4fbbf8e087a90487c0764d3c3e5bec20884a3cdbb8.jpg", "reference_inputs/images/9244160e155b63bdbce63c0444e48588782c5415701fca6dadd7b04750cdf985.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000167", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000167", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [146, 482, 190, 503]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [500, 281, 542, 294]\nB: [795, 252, 814, 302]\nC: [50, 246, 78, 303]\nD: [398, 250, 451, 295]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2b1002b17b14d59cab12fe2557f63347dd40fedc814306ea684da7b07f570236.jpg", "reference_inputs/images/c1116c823eb3027d74934cdd532864e0cb992a42a4af3ebcabca0ac643e91343.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000171", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000171", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [494, 262, 519, 291]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [424, 446, 459, 501]\nB: [817, 556, 885, 575]\nC: [502, 366, 515, 383]\nD: [516, 428, 532, 439]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8fa263fd0a2bb14dc52206fb1d755d9a5f3fdea1c2da06223795028c853d0a3e.jpg", "reference_inputs/images/35cd9180f440e70c746afd7565e001662570370fa734d09ad27e609693f282e9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000172", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000172", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [773, 402, 804, 424]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [240, 263, 263, 296]\nB: [668, 267, 675, 293]\nC: [282, 267, 290, 294]\nD: [715, 250, 727, 300]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5efc52003cf953f94e3968827080733c9c7d3c87a5365c1c13b75ff2f9a21f03.jpg", "reference_inputs/images/a863c0425374d378735269f9a18b0e786addc41c4cfe1ec95b61f4e2c1feb0bd.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000176", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000176", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [722, 275, 823, 314]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [434, 483, 454, 507]\nB: [175, 490, 246, 541]\nC: [509, 436, 522, 449]\nD: [730, 566, 789, 575]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8d37f800ffbf71905c3386e8f60ca28c2dadd52fc67793a1c4e7b790f4df4fcb.jpg", "reference_inputs/images/efecc6ea655d2baf9c731021df4e0e26e6ddae097bd52949d1ab78c194d2649f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000179", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000179", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [697, 267, 705, 296]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [460, 402, 479, 435]\nB: [129, 493, 135, 514]\nC: [173, 488, 186, 506]\nD: [715, 490, 720, 510]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/511dc7d282931166d0a31a405453bd26ae804a4dc5214093bb6a1982726f200e.jpg", "reference_inputs/images/698a1f1fea025f2ef0067d251e43da911ee6143b00cb60b7e052abd12237e7ba.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000180", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000180", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [186, 44, 221, 72]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [226, 261, 238, 297]\nB: [235, 263, 250, 294]\nC: [360, 269, 366, 292]\nD: [98, 246, 133, 303]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7fe81ffe48ac7047ad4182035e62772d238bb9f1dfef4032e169297c0ac7b451.jpg", "reference_inputs/images/861fd07722f02e7775960ebcbb39ddd697e58e8a2aa057c62f68cc8e41bceecf.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000181", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000181", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [226, 261, 238, 297]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [186, 44, 221, 72]\nB: [503, 78, 552, 181]\nC: [848, 136, 886, 157]\nD: [167, 490, 198, 513]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/861fd07722f02e7775960ebcbb39ddd697e58e8a2aa057c62f68cc8e41bceecf.jpg", "reference_inputs/images/7fe81ffe48ac7047ad4182035e62772d238bb9f1dfef4032e169297c0ac7b451.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000183", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000183", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [899, 206, 950, 328]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [705, 146, 734, 166]\nB: [143, 223, 178, 238]\nC: [170, 0, 197, 17]\nD: [754, 223, 788, 238]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3f33e168375781384fa6ff8115c76325dcb459dd1f23d2582f540877beb2c83e.jpg", "reference_inputs/images/110cd9767a4801cd4b1e09a67850b3b7fa467db73588ae8e67663aa8562adf99.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000184", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000184", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [467, 433, 481, 447]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [702, 262, 712, 296]\nB: [318, 270, 398, 310]\nC: [363, 266, 406, 299]\nD: [498, 274, 521, 293]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/760df4126276baeb1783940be483f8faf7c7cad7e4581beac622ddf0980c251b.jpg", "reference_inputs/images/b4d21a7e3f68d56255291f70c6384a51e8da3d9f3012b231848b23a57471c549.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000185", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000185", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [713, 515, 736, 542]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [103, 215, 136, 318]\nB: [822, 264, 830, 294]\nC: [411, 268, 418, 293]\nD: [895, 282, 916, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8771c85ad7e20bee39b6eb3be777a3478801bbd3fe47bf6dc3ea6bcf510f6924.jpg", "reference_inputs/images/62bc1d89c2ab42acf8235ee12e229525f31983d1e1f0d596665befbb9a4745c8.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000186", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000186", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [822, 264, 830, 294]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [164, 122, 198, 141]\nB: [713, 515, 736, 542]\nC: [758, 78, 784, 102]\nD: [98, 82, 134, 106]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/62bc1d89c2ab42acf8235ee12e229525f31983d1e1f0d596665befbb9a4745c8.jpg", "reference_inputs/images/8771c85ad7e20bee39b6eb3be777a3478801bbd3fe47bf6dc3ea6bcf510f6924.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000188", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000188", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [958, 249, 978, 302]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [300, 546, 306, 564]\nB: [195, 548, 205, 574]\nC: [717, 548, 721, 574]\nD: [280, 538, 297, 561]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/38b794c38af6958c62c3bf8c175f24dd96ded835eb611aef8431b16ece19e7c4.jpg", "reference_inputs/images/a4cf592a72e7c039e89f2ce68a4f098467de07ca020316c2c156372cb9a1cfc5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000189", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000189", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [753, 274, 786, 287]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [635, 268, 641, 292]\nB: [178, 259, 190, 297]\nC: [779, 257, 791, 296]\nD: [675, 266, 698, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ab2188839eca588e14ececd553c05ab309b3db806f153f7d78edfdf9250a16d7.jpg", "reference_inputs/images/72cd0a8adb1640bf26b9f9f7b03503233379e3d16e3754adda473802969f4c0f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000190", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000190", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [675, 266, 698, 294]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [753, 274, 786, 287]\nB: [106, 422, 143, 453]\nC: [822, 434, 857, 452]\nD: [748, 23, 778, 51]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/72cd0a8adb1640bf26b9f9f7b03503233379e3d16e3754adda473802969f4c0f.jpg", "reference_inputs/images/ab2188839eca588e14ececd553c05ab309b3db806f153f7d78edfdf9250a16d7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000191", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000191", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [417, 503, 445, 540]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [354, 262, 438, 314]\nB: [500, 270, 534, 296]\nC: [546, 280, 565, 292]\nD: [436, 276, 454, 292]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7abe8bf0e8316b5c50cf7a1af93b7def29fa22d68678887ccb5eace7bbf814b0.jpg", "reference_inputs/images/1ac8ae79665ede165e788f39ca756a98d62b27c8dc1a0d5c67b311d5992b3466.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000192", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000192", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [747, 232, 787, 248]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [426, 241, 460, 307]\nB: [630, 253, 648, 304]\nC: [542, 274, 548, 294]\nD: [554, 243, 566, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f2ce39d37b6fdce7c4b61a39b23f714c97ca02184f46e2281cfe179863206144.jpg", "reference_inputs/images/57c3f70f35fc88b642076e72a90f518e3004de43237b3e468d97c6500ead1958.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000195", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000195", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [120, 549, 157, 574]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [425, 281, 446, 294]\nB: [614, 272, 621, 296]\nC: [452, 197, 488, 332]\nD: [818, 245, 838, 310]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1dd5e7838e2c03695637a18be71970132c07aa954e052efb5afe736e1b8ede06.jpg", "reference_inputs/images/772ae829aab097e813edc986df6337aa53a4dfb46188b630cff56caf74228b46.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000198", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000198", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [76, 253, 124, 270]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [126, 247, 149, 308]\nB: [221, 249, 244, 310]\nC: [853, 243, 876, 310]\nD: [892, 268, 1023, 338]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/34f9529c1e00a701b20a09b60bd22d48dce47dca4e7396f61530ca5e4a3ccfad.jpg", "reference_inputs/images/8c76db10b89a1cf3ac66190f7fe41fc36d0d46416cce3ba018f1edc118ee675b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000199", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000199", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [540, 489, 560, 514]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [545, 277, 570, 294]\nB: [0, 218, 316, 446]\nC: [463, 282, 483, 291]\nD: [496, 265, 517, 290]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/977acfa7137e557053b937cdac0f694f0400cda381f38b0fcb1b44864ca8defa.jpg", "reference_inputs/images/bcdc1e69c0a36fdd6f81b6dc138bac8948640e1f401b454a6715a6941bd5a4a5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000200", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000200", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [545, 277, 570, 294]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [462, 406, 476, 422]\nB: [540, 489, 560, 514]\nC: [497, 406, 516, 434]\nD: [469, 435, 486, 450]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bcdc1e69c0a36fdd6f81b6dc138bac8948640e1f401b454a6715a6941bd5a4a5.jpg", "reference_inputs/images/977acfa7137e557053b937cdac0f694f0400cda381f38b0fcb1b44864ca8defa.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000201", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000201", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [552, 236, 570, 309]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [166, 549, 215, 574]\nB: [815, 412, 856, 429]\nC: [7, 47, 62, 74]\nD: [754, 275, 779, 298]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5a44f25e591592c2db02b8e83f9d3cf80ff7f08e9ed31c0e2a11bae27caf66dd.jpg", "reference_inputs/images/8a9a5431ad6810074181893a198207987111ee04ad9a44a51879db027ca2565a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000205", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000205", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [991, 237, 1021, 314]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [27, 134, 66, 150]\nB: [783, 193, 820, 208]\nC: [762, 178, 794, 197]\nD: [246, 541, 272, 572]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c8e861342cff147472a889aca5f84bc5c0bc693aae247f0b0ff24c88927645db.jpg", "reference_inputs/images/8ca314213564a7362598643ee7e9bc0b091f5a0815d448b65114ea6449619618.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000207", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000207", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [316, 294, 374, 403]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [341, 265, 402, 302]\nB: [518, 282, 539, 292]\nC: [105, 250, 151, 299]\nD: [439, 282, 460, 291]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1045dc14fabe226b9f5e057b367632be473563d2f1157262f947e0220d68634a.jpg", "reference_inputs/images/121ca0a666b27ee44d92209d4b2bcecda33df73f1f6332b6dff01271e6619a5e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000208", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000208", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [717, 258, 730, 297]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [708, 138, 749, 154]\nB: [763, 325, 793, 355]\nC: [98, 454, 143, 475]\nD: [87, 265, 125, 302]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/121ca0a666b27ee44d92209d4b2bcecda33df73f1f6332b6dff01271e6619a5e.jpg", "reference_inputs/images/1045dc14fabe226b9f5e057b367632be473563d2f1157262f947e0220d68634a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000210", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000210", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [504, 51, 555, 166]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [508, 276, 534, 295]\nB: [462, 282, 498, 293]\nC: [351, 255, 439, 312]\nD: [651, 260, 887, 352]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ed6a292d6cd804eddcda4961ead9b4a2b5e9bae9f81b95a5aabdcde5d500b5d6.jpg", "reference_inputs/images/c33e1aeee82405ba47f657e32649e2e4e8d1e73b08d660d0fe811f4195c3ad2e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000211", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000211", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [225, 260, 235, 294]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [170, 13, 196, 36]\nB: [92, 330, 136, 348]\nC: [751, 522, 786, 548]\nD: [121, 123, 158, 140]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c33e1aeee82405ba47f657e32649e2e4e8d1e73b08d660d0fe811f4195c3ad2e.jpg", "reference_inputs/images/ed6a292d6cd804eddcda4961ead9b4a2b5e9bae9f81b95a5aabdcde5d500b5d6.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000212", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000212", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [702, 231, 745, 318]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [170, 0, 197, 17]\nB: [143, 223, 178, 238]\nC: [705, 146, 734, 166]\nD: [754, 223, 788, 238]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f7d45e66b52b56f9d2fe3a416f385af8583f41388a32feefa9e92bff4b30e63f.jpg", "reference_inputs/images/110cd9767a4801cd4b1e09a67850b3b7fa467db73588ae8e67663aa8562adf99.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000213", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000213", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [91, 188, 129, 210]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [778, 250, 800, 303]\nB: [977, 247, 987, 305]\nC: [746, 247, 778, 305]\nD: [937, 258, 946, 297]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/903023c225c149cedd651170969d99e1abf3ce6bbe19bfa15fba8caa56e5ca42.jpg", "reference_inputs/images/6668b014712035c16c55c9d346ec9d65f9aea0505cf77f2f90be5d06ec528252.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000215", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000215", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [508, 238, 536, 313]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [143, 331, 180, 345]\nB: [788, 390, 827, 404]\nC: [826, 177, 838, 188]\nD: [588, 0, 645, 49]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1863ae1679d7711a8275fbc629d40f0f2d6d22193fe8db6f3890c46eda1d080a.jpg", "reference_inputs/images/a636870c7e4fd90a63cb02038fa2ea19d346afd5a60073645021f783b87298d8.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000216", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000216", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [42, 446, 91, 467]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [36, 250, 58, 298]\nB: [906, 250, 928, 298]\nC: [166, 256, 172, 299]\nD: [690, 262, 705, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/80411caf2c694c923da91acf9ce65757d6d0b3cc4e9f32a494dc213f2e0988f9.jpg", "reference_inputs/images/9ee52d4b590978577fe3a21b750b8093f46ed3b4efce315bf8098e1feb0e6325.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000217", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000217", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [278, 262, 285, 294]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [146, 413, 180, 436]\nB: [148, 172, 192, 189]\nC: [42, 446, 91, 467]\nD: [874, 446, 918, 466]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9ee52d4b590978577fe3a21b750b8093f46ed3b4efce315bf8098e1feb0e6325.jpg", "reference_inputs/images/80411caf2c694c923da91acf9ce65757d6d0b3cc4e9f32a494dc213f2e0988f9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000218", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000218", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [874, 71, 920, 93]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [211, 242, 230, 304]\nB: [128, 227, 170, 317]\nC: [118, 260, 130, 296]\nD: [304, 270, 312, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d185a65e2420ed7d26f40dd0df1b158c836a094d8997559febbdf0997ce1e750.jpg", "reference_inputs/images/e637bb7fcd81a6fb6bdc4e262f86cf16f78627c8a83cd6bccefdd9e548661bdb.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000219", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000219", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [438, 282, 461, 292]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [505, 392, 521, 416]\nB: [543, 492, 569, 514]\nC: [476, 457, 494, 482]\nD: [418, 513, 449, 539]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0babe2b74ff3bef3cdac98263fe4524186cb1fd3c2bd927891a0e7a8de627389.jpg", "reference_inputs/images/3e21d07491e09cd36408a4cbcdacf1e77d7462660023e5288805dc80014a0207.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000220", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000220", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [52, 147, 83, 165]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [400, 250, 427, 307]\nB: [464, 252, 471, 303]\nC: [659, 267, 670, 297]\nD: [717, 259, 730, 298]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/903023c225c149cedd651170969d99e1abf3ce6bbe19bfa15fba8caa56e5ca42.jpg", "reference_inputs/images/a5d0e7cbf9767cb7cc710572ec696712ed5a45fc51ec05aba895560f75129c58.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000222", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000222", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [750, 436, 777, 454]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [510, 266, 516, 292]\nB: [156, 271, 166, 292]\nC: [64, 255, 78, 269]\nD: [524, 267, 532, 293]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/beed69582179f7e4769d3cc45dd0fdff7a8efafbfa7a4eb2f9e0c8f4f69a69ca.jpg", "reference_inputs/images/79eb13a479257c25e7f5b6897a0354841cbe1086dd11e973156ea1192b073fc2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000227", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000227", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [526, 378, 540, 398]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [542, 246, 590, 295]\nB: [322, 277, 428, 302]\nC: [434, 260, 467, 292]\nD: [874, 259, 885, 296]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ad7e06d19be14c7572094fcd2a8656811827fd7f8c582feb9f85922ac8d4a358.jpg", "reference_inputs/images/4a9235ef98cd0be9d0efdcdc3660811c02b4055a0d190afdc8cf695abe4f7bee.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000228", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000228", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [434, 260, 467, 292]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [506, 472, 527, 490]\nB: [459, 407, 486, 418]\nC: [526, 378, 540, 398]\nD: [476, 363, 490, 382]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4a9235ef98cd0be9d0efdcdc3660811c02b4055a0d190afdc8cf695abe4f7bee.jpg", "reference_inputs/images/ad7e06d19be14c7572094fcd2a8656811827fd7f8c582feb9f85922ac8d4a358.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000231", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000231", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [410, 517, 417, 542]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [268, 253, 282, 302]\nB: [414, 270, 420, 295]\nC: [56, 266, 65, 293]\nD: [38, 267, 47, 293]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cfd3ae4956d0d912e42e8b069a600698afb87d5f906b5079bc3f973feb58a839.jpg", "reference_inputs/images/8f6e1d91122c91578b18dffedaa92d09784d3e28552882ecd25afe2c34326083.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000232", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000232", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [446, 281, 474, 294]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [454, 428, 470, 447]\nB: [458, 468, 480, 490]\nC: [494, 430, 516, 466]\nD: [549, 555, 576, 575]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/34d027dbab11397f568f524e93a1358bb23b7502da274b0c04a1dba07548b837.jpg", "reference_inputs/images/0338e6668ccc2980e70ebd4a9e32d2deccdc84389a69907b91810ac1184821af.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000233", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000233", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [402, 31, 453, 146]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [507, 277, 526, 291]\nB: [550, 224, 630, 302]\nC: [0, 253, 105, 342]\nD: [450, 275, 478, 296]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0fc31c1a19b1bb2cdd002913a35bc79ef3f253358915b124d93a788326e3387b.jpg", "reference_inputs/images/068981882776608705076f4c4c3e998b66611be3d5507c1608daae3a21fad37c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000234", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000234", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [761, 55, 789, 82]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [706, 222, 741, 318]\nB: [734, 246, 754, 306]\nC: [307, 222, 337, 314]\nD: [606, 263, 618, 299]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/94a42c416d039b03a48924175f0fe6b51d54b3a84c1a26fee68d14316a486a57.jpg", "reference_inputs/images/cd6a290390b91df5adfef18059245e2cc6e94d1dd09f6b3f5d9825f1eee65cfb.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000235", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000235", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [734, 246, 754, 306]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [874, 125, 920, 142]\nB: [761, 55, 789, 82]\nC: [39, 126, 90, 142]\nD: [873, 396, 918, 413]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cd6a290390b91df5adfef18059245e2cc6e94d1dd09f6b3f5d9825f1eee65cfb.jpg", "reference_inputs/images/94a42c416d039b03a48924175f0fe6b51d54b3a84c1a26fee68d14316a486a57.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000239", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000239", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [534, 425, 549, 440]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [443, 269, 480, 298]\nB: [686, 202, 1000, 350]\nC: [274, 275, 367, 307]\nD: [577, 277, 637, 302]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1c3d4109a38cb72dd2d349b03eb89d2366a1cfb46285ae129959aae657de9447.jpg", "reference_inputs/images/e8f86a144ff07fab1b6ed4595caeffbf95d4782f342c62d61c804499bb0065bf.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000240", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000240", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [686, 202, 1000, 350]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [508, 469, 525, 487]\nB: [507, 358, 519, 374]\nC: [534, 425, 549, 440]\nD: [450, 444, 466, 460]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e8f86a144ff07fab1b6ed4595caeffbf95d4782f342c62d61c804499bb0065bf.jpg", "reference_inputs/images/1c3d4109a38cb72dd2d349b03eb89d2366a1cfb46285ae129959aae657de9447.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000244", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000244", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [650, 519, 655, 542]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [383, 264, 392, 294]\nB: [891, 258, 902, 297]\nC: [843, 269, 850, 292]\nD: [914, 262, 924, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6d76248e7087fb5fcf6decf03cdcb4e685fdbb05d11f13983b35973a1a64e154.jpg", "reference_inputs/images/74f386330c7d9af7076f742a7d66db3df65f2fcc5027d9e6f0c2bb2d6c20ab41.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000247", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000247", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [350, 264, 370, 298]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [843, 219, 879, 233]\nB: [654, 340, 721, 456]\nC: [914, 390, 960, 406]\nD: [233, 494, 273, 517]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/16652bd8a0749e2b8903fd22753cff94fa7d4fe455aeb70555cc2ffe81a55955.jpg", "reference_inputs/images/7d3719bc63e8a194e6b5bade4d3c3bfbb3b34170c480ec70ad99fbee5cfcb1c4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000250", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000250", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [119, 438, 156, 455]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [174, 258, 188, 298]\nB: [737, 260, 747, 297]\nC: [46, 246, 94, 302]\nD: [945, 209, 983, 318]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5efc52003cf953f94e3968827080733c9c7d3c87a5365c1c13b75ff2f9a21f03.jpg", "reference_inputs/images/1c86cc5951e53a469d0dfa1d49811abb2200a75b723703e0b03968751a5c81cc.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000252", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000252", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [106, 262, 148, 276]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [354, 247, 372, 307]\nB: [280, 249, 298, 309]\nC: [371, 251, 388, 307]\nD: [262, 253, 278, 306]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8a36ca2e5fe4225b02d5b48a0cbef01de524cc951818afcfdff8047020c582b5.jpg", "reference_inputs/images/91613324c1c5a055df08a78e89bc88cb93a4e51fd5b21ca12316859c761a71ab.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000253", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000253", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [354, 247, 372, 307]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [82, 211, 126, 226]\nB: [106, 262, 148, 276]\nC: [90, 46, 130, 75]\nD: [779, 498, 810, 526]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/91613324c1c5a055df08a78e89bc88cb93a4e51fd5b21ca12316859c761a71ab.jpg", "reference_inputs/images/8a36ca2e5fe4225b02d5b48a0cbef01de524cc951818afcfdff8047020c582b5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000258", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000258", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [87, 265, 125, 302]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [542, 253, 558, 294]\nB: [122, 247, 152, 314]\nC: [958, 242, 983, 303]\nD: [164, 207, 549, 358]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1045dc14fabe226b9f5e057b367632be473563d2f1157262f947e0220d68634a.jpg", "reference_inputs/images/05653c6340d7df55442d03ed3d8119b93e8374d103577b227acac495e750863c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000260", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000260", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [686, 549, 691, 574]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [671, 266, 679, 294]\nB: [872, 234, 896, 307]\nC: [655, 266, 662, 292]\nD: [906, 243, 924, 301]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/369d8c5a94fad57873a397e17382502d73f237070752a22f9e7d59767de6f1e1.jpg", "reference_inputs/images/52169b7a0b342f2d63699bcc716256cbbe071a240289254b98442a1483faace4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000261", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000261", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [368, 274, 422, 303]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [434, 472, 457, 501]\nB: [441, 515, 464, 546]\nC: [686, 549, 691, 574]\nD: [503, 423, 517, 438]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/52169b7a0b342f2d63699bcc716256cbbe071a240289254b98442a1483faace4.jpg", "reference_inputs/images/369d8c5a94fad57873a397e17382502d73f237070752a22f9e7d59767de6f1e1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000262", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000262", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [642, 514, 658, 535]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [684, 273, 691, 296]\nB: [633, 276, 646, 294]\nC: [352, 274, 366, 293]\nD: [990, 246, 1016, 310]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4dabf19c91624ce95cd33715a4547c90755ab3181ee934441d405dcb02c76b3b.jpg", "reference_inputs/images/c5b600c09a6d37cd46165ccf13dda6e2f27b7d463018639951817900fbdef907.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000265", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000265", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [306, 238, 329, 314]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [171, 192, 206, 209]\nB: [759, 432, 787, 458]\nC: [510, 155, 537, 171]\nD: [187, 290, 219, 304]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3c1eed17076a7921cc0af79964c28dec9371a24e9179fdf879f58928ac06acc9.jpg", "reference_inputs/images/b45e5d4a35ce9f457d62b66c0c04b331a9dc27a35f45653cbb0002e4d287a1e0.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000266", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000266", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [146, 350, 190, 367]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [198, 246, 235, 311]\nB: [960, 262, 970, 294]\nC: [474, 256, 494, 302]\nD: [454, 250, 477, 306]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4bd927b0b51934721abede2f40236bb390f0971b1dcb47b3db9fdc20f9de8603.jpg", "reference_inputs/images/17f870c3e80b4ea658ab9fad190bfd9efb07496fc221e0102bd1988432097401.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000268", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000268", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [667, 244, 686, 307]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [54, 478, 98, 503]\nB: [798, 307, 831, 327]\nC: [123, 481, 160, 504]\nD: [586, 0, 635, 15]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8c5f73d844824d15dda918f79413219de0ec916fc701703ec066bad68bcb72e8.jpg", "reference_inputs/images/5ddafb32c4966481d1800c1d88b81b79c74baa0263019196d4f686c8aa5348ca.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000270", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000270", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [500, 406, 551, 518]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [0, 244, 182, 429]\nB: [390, 279, 431, 297]\nC: [442, 279, 460, 291]\nD: [495, 247, 573, 315]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/686afe55c2fa06299ea2faa81211ce9eaa6045b06db4b14b5dbd4117d79dc52f.jpg", "reference_inputs/images/176edff7c377ecfb0a85bbb5fe6f909debb69e63dd867a4d9757b6b09591aec1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000271", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000271", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [495, 247, 573, 315]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [505, 0, 552, 88]\nB: [500, 406, 551, 518]\nC: [325, 0, 378, 54]\nD: [52, 151, 97, 170]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/176edff7c377ecfb0a85bbb5fe6f909debb69e63dd867a4d9757b6b09591aec1.jpg", "reference_inputs/images/686afe55c2fa06299ea2faa81211ce9eaa6045b06db4b14b5dbd4117d79dc52f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000273", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000273", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [113, 278, 158, 295]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [303, 254, 326, 306]\nB: [275, 264, 291, 308]\nC: [649, 240, 672, 307]\nD: [319, 260, 330, 302]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6424a6e0b499a17f7382d0f8f75b062bd2157088fe8cf465fa6806e6fb5567e3.jpg", "reference_inputs/images/e304d7a9e22f0336047f9a4c01a819c39ffd0e777ada57b545c1312a956f2a58.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000280", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000280", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [885, 219, 916, 322]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [720, 326, 746, 342]\nB: [791, 162, 821, 182]\nC: [158, 162, 191, 182]\nD: [163, 362, 200, 376]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e896068cd5e02e2ec5d94cbaf3f2eafa7c90a9b27fc4198ea4a96cd1918c4c7d.jpg", "reference_inputs/images/2cc8a54174351d0059fbddcddde72408fcaefe0c3a706b17b11c8bb9852967bf.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000281", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000281", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [26, 226, 66, 314]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [815, 412, 856, 429]\nB: [166, 549, 215, 574]\nC: [7, 47, 62, 74]\nD: [754, 275, 779, 298]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d704d3ce55f0f888be6555c08731f1558642b777791b4fa2cf0dd70814ad2ea9.jpg", "reference_inputs/images/8a9a5431ad6810074181893a198207987111ee04ad9a44a51879db027ca2565a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000282", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000282", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [674, 554, 683, 575]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [637, 269, 642, 291]\nB: [729, 259, 740, 295]\nC: [134, 254, 157, 299]\nD: [77, 251, 94, 299]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/54a45b98db6cf9dcae184341b11c3ada9efb63b5c8c0222a21ea269ead7746f1.jpg", "reference_inputs/images/389a42f43ad667172d6855c4182d3f7f74d9b73cea54b3b04b85012cce008507.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000289", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000289", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [189, 58, 230, 80]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [233, 223, 274, 318]\nB: [506, 238, 534, 314]\nC: [992, 243, 1020, 302]\nD: [814, 241, 854, 309]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f2ce39d37b6fdce7c4b61a39b23f714c97ca02184f46e2281cfe179863206144.jpg", "reference_inputs/images/6fbf9e0872d4e16add6d278a01df5aa9f4b19360391d2d9887a86189ddce823b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000290", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000290", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [506, 238, 534, 314]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [189, 58, 230, 80]\nB: [866, 148, 909, 169]\nC: [111, 251, 151, 268]\nD: [747, 232, 787, 248]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6fbf9e0872d4e16add6d278a01df5aa9f4b19360391d2d9887a86189ddce823b.jpg", "reference_inputs/images/f2ce39d37b6fdce7c4b61a39b23f714c97ca02184f46e2281cfe179863206144.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000291", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000291", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [95, 216, 134, 230]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [806, 253, 821, 300]\nB: [475, 280, 493, 292]\nC: [142, 253, 158, 300]\nD: [257, 266, 266, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/73b89e2237847f5cca0725df1591c03e7b3ef51668e61483992f6a44fab9033a.jpg", "reference_inputs/images/267140f4919911df6db0b7cf8ed7c04c6d81dac27a9d963632b49ae48d93239f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000296", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000296", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [164, 222, 209, 237]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [266, 260, 274, 294]\nB: [694, 266, 700, 292]\nC: [284, 266, 304, 294]\nD: [100, 250, 118, 302]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b6e69bb6b20e30c6251d467415a026ad170366486c78c77a0625195e38cab5b5.jpg", "reference_inputs/images/214450fcb9ec8f693510b00eb2718a24d336b00136b94d6785fc63ff6db98864.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000297", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000297", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [284, 266, 304, 294]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [823, 518, 856, 541]\nB: [147, 526, 180, 559]\nC: [818, 76, 858, 97]\nD: [164, 222, 209, 237]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/214450fcb9ec8f693510b00eb2718a24d336b00136b94d6785fc63ff6db98864.jpg", "reference_inputs/images/b6e69bb6b20e30c6251d467415a026ad170366486c78c77a0625195e38cab5b5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000299", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000299", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [608, 242, 628, 312]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [750, 436, 777, 454]\nB: [722, 0, 746, 27]\nC: [509, 236, 556, 338]\nD: [720, 442, 745, 465]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9b63b95141de51fe7f4ef28a9d59d6e667353c774669cdd16e4143b1a44e07b0.jpg", "reference_inputs/images/beed69582179f7e4769d3cc45dd0fdff7a8efafbfa7a4eb2f9e0c8f4f69a69ca.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000301", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000301", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [225, 220, 265, 315]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [842, 151, 882, 168]\nB: [582, 0, 639, 65]\nC: [190, 383, 230, 398]\nD: [760, 116, 789, 142]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e157a1b5820d9d1cdf2a066f3968259272d72469f5e1900fe40298d50cd68970.jpg", "reference_inputs/images/16b926098bbd131fdcd2cf63db01c3e7128d0bc977544e3e0e8485122b146dc8.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000305", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000305", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [653, 258, 665, 297]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [779, 66, 830, 86]\nB: [766, 471, 796, 503]\nC: [680, 412, 704, 430]\nD: [142, 2, 178, 34]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bc9ecf1a01f58e317cba1408870ca8cd9224a3287a4cfcfb65985c9815fc847b.jpg", "reference_inputs/images/32ec1cfd3248914be2bb9127bbec2417bc640c23321fbce528d484cc3448edf0.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000307", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000307", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [45, 251, 64, 309]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [81, 514, 124, 538]\nB: [100, 124, 147, 145]\nC: [85, 519, 128, 550]\nD: [92, 368, 131, 383]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/008072f8938f8f1a63f207a5da458f8c1016122eb4926ea5e9575db7ace1132e.jpg", "reference_inputs/images/5227f1514bc26367b9a685962255b280ec8c82f911ad7282b2912b36f5b3dc4c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000308", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000308", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [796, 19, 824, 60]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [722, 258, 730, 276]\nB: [217, 254, 230, 272]\nC: [704, 254, 710, 278]\nD: [624, 242, 638, 282]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/99eeaadc2838f3cd9673b0ea6ee84276657a1a1d002c3dda5d6e866228c9a89a.jpg", "reference_inputs/images/db28015dd827c7a87eca54c8bed03ee4194eed92a3b8f6f40966f1f3a8f4b38c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000309", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000309", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [166, 178, 197, 192]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [98, 253, 125, 298]\nB: [297, 270, 302, 291]\nC: [645, 268, 652, 291]\nD: [2, 242, 33, 307]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/66e917615d59dc2f75de49fdf708888f316aba06463d0b7fe2f2cdb39aed6e65.jpg", "reference_inputs/images/5a50298408b5a44380b325f06a63c9e6f92a85c81613f17465ad2b8e5b4836a3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000312", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000312", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [238, 228, 286, 315]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [510, 460, 526, 479]\nB: [798, 506, 810, 527]\nC: [638, 530, 644, 553]\nD: [477, 433, 494, 447]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8d417a17c18faf40fc765c8b23d0f70cca61632b0e3f1a324e97c9d544bb2988.jpg", "reference_inputs/images/6866c21286157ed1a98a7a54f7f47485a2c683f40fea7eb1b587906f1433733e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000313", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000313", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [92, 368, 131, 383]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [210, 250, 231, 308]\nB: [38, 253, 56, 309]\nC: [523, 242, 542, 306]\nD: [851, 250, 878, 303]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5227f1514bc26367b9a685962255b280ec8c82f911ad7282b2912b36f5b3dc4c.jpg", "reference_inputs/images/b4124de2733458d6b02f2a8d6824132a5a497241fa556c58ee1b573e171b45bc.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000316", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000316", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [657, 230, 686, 311]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [162, 274, 183, 287]\nB: [100, 519, 138, 545]\nC: [754, 55, 779, 87]\nD: [183, 486, 216, 513]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d3a0110b6a67af346162179241e854b0f3917a8527d17d0238e525236b325fc3.jpg", "reference_inputs/images/6af64a6f155a53fd7e9405fd0240f3d88f46f26297b3a634601d63726b1edce1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000317", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000317", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [361, 249, 383, 312]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [154, 120, 188, 138]\nB: [158, 445, 193, 466]\nC: [750, 409, 776, 426]\nD: [91, 409, 129, 425]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/baefc67a32eccdf965e3e0c36e8575e991982312d417953f153578450baf20b8.jpg", "reference_inputs/images/163d53839b6d2b76ccde8abff6da2726cb12953bdfcc951e7e20911117043a7f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000318", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000318", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [470, 490, 490, 514]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [536, 279, 553, 291]\nB: [499, 269, 516, 289]\nC: [404, 272, 459, 306]\nD: [252, 268, 272, 293]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bfbff7adbebc42ae595d39371347b63219639483ac4c74142b2bc61623dabbd5.jpg", "reference_inputs/images/06ecb69a21b3792b108f8f5945a2d8ba57f1cf4758d640de4aaa28301ab4558b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000321", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000321", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [208, 62, 242, 85]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [970, 178, 1023, 344]\nB: [615, 269, 622, 296]\nC: [691, 203, 788, 333]\nD: [943, 273, 954, 298]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/dedfb06ec513842c3556edc6180d1cbc0ae0754bcd127e769bfb1729ff588f32.jpg", "reference_inputs/images/f288b97252aa8d63814994094c198f70d705ed1b8a7f816ac192773ad6f496dd.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000322", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000322", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [970, 178, 1023, 344]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [208, 62, 242, 85]\nB: [276, 98, 302, 121]\nC: [858, 168, 896, 186]\nD: [631, 470, 649, 490]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f288b97252aa8d63814994094c198f70d705ed1b8a7f816ac192773ad6f496dd.jpg", "reference_inputs/images/dedfb06ec513842c3556edc6180d1cbc0ae0754bcd127e769bfb1729ff588f32.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_20m/000324", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_20m/all_questions.json", "question_id": "000324", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [746, 262, 757, 296]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [122, 443, 161, 462]\nB: [812, 295, 846, 313]\nC: [168, 341, 208, 355]\nD: [766, 0, 795, 29]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/67d3c4dfaf5334baa46fdb18857f233075e29e32dcbad3903bb91dcb5950e373.jpg", "reference_inputs/images/9b07624335b4fa633b68a1113aa88dc561c5ac1dcdfee58fdfa373ca97d99437.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000006", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [510, 239, 538, 307]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [274, 412, 289, 426]\nB: [301, 156, 321, 166]\nC: [195, 551, 216, 571]\nD: [733, 0, 751, 19]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b33673b15d0fa3dc73b6257e27c913af7c3d9becfaea84927dfcd5d4001a167c.jpg", "reference_inputs/images/fc8f5d7acbd30eec6dd9a663f4ff0c170d5430d8a7c9e0cdb3a26b0bc0bbebc8.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000010", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [476, 410, 490, 432]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [460, 267, 479, 290]\nB: [872, 279, 893, 293]\nC: [414, 281, 466, 294]\nD: [536, 261, 560, 279]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f5158663b6493a36ff24cee09cee5d0205729b56889b7518818f61b53f9e3b4b.jpg", "reference_inputs/images/b683797c51a05883b7e76e4ff54291fc1eb45bd00558f4d84c85fd38534d3947.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000011", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [648, 447, 659, 465]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [345, 270, 352, 295]\nB: [380, 267, 383, 294]\nC: [718, 271, 725, 297]\nD: [865, 257, 882, 306]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/65ab20dcbba598480d730324f3c9ceffe765f5fccfc0bf381aeb02abec648f0f.jpg", "reference_inputs/images/9c6b3f29f4d9176113b4baac3a280ea580c2ac94190eae813932e337951513d2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000014", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [577, 270, 582, 291]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [731, 1, 746, 23]\nB: [720, 205, 736, 214]\nC: [291, 140, 313, 152]\nD: [336, 164, 350, 175]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5940e2d6e51e07fc71ca2c19e8d33a5a3ee00569378163517bc52e26913430ab.jpg", "reference_inputs/images/19a1b73f12288dcce7ae3ca0cbe6db93634d4b933c6efb82423b324d681afd14.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000015", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [521, 409, 534, 427]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [506, 282, 539, 292]\nB: [94, 252, 134, 301]\nC: [552, 224, 629, 302]\nD: [425, 258, 463, 293]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/24a2eecddb166a028d5f9a79b38048fb8bd9441a5c5dec2077a6e3485ed9e041.jpg", "reference_inputs/images/4ae21f7784b345ef7e5090eb8889ba1faa07029fe667db9eead51be6bac0ce7e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000016", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [454, 242, 480, 310]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [382, 0, 418, 33]\nB: [269, 158, 291, 169]\nC: [710, 375, 728, 385]\nD: [284, 361, 301, 371]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/73761be65818a5dc9026c4b691d72ee86c5940efe951d9acb79754e03601e57f.jpg", "reference_inputs/images/52099689106298857ec9115810cece30ac6c15664bf475beb76baf80766f357e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000019", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [40, 242, 94, 298]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [453, 510, 470, 526]\nB: [487, 471, 499, 489]\nC: [533, 497, 549, 510]\nD: [506, 417, 520, 437]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/016d15709c606ef2934ed869e77614cc7513cfc46d8e48ffa81ac3831a4acc4d.jpg", "reference_inputs/images/a0241464dec3f4225f6d5e24ec491dd01f32833a9eb8d175a131b63a9175f0c1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000020", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [137, 432, 161, 446]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [474, 256, 494, 302]\nB: [198, 246, 235, 311]\nC: [454, 250, 477, 306]\nD: [960, 262, 970, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/beb4a4da294a37a291b7007f52317ea5ec3c945d61ec032b2127c546ed28efa1.jpg", "reference_inputs/images/17f870c3e80b4ea658ab9fad190bfd9efb07496fc221e0102bd1988432097401.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000021", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [198, 246, 235, 311]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [749, 22, 770, 42]\nB: [570, 565, 597, 575]\nC: [137, 432, 161, 446]\nD: [274, 541, 294, 558]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/17f870c3e80b4ea658ab9fad190bfd9efb07496fc221e0102bd1988432097401.jpg", "reference_inputs/images/beb4a4da294a37a291b7007f52317ea5ec3c945d61ec032b2127c546ed28efa1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000024", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000024", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [446, 367, 479, 440]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [888, 266, 1023, 399]\nB: [496, 246, 535, 295]\nC: [502, 261, 572, 314]\nD: [254, 235, 423, 333]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c236d9347f55c0c3ec5ba178d174cdfed7be3be0fb69abec27fac1b60b6cc6cf.jpg", "reference_inputs/images/64ecb14e96931fa12d63d7d52e97e6551d9ab8794bec873526ad42f1cf64fd39.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000025", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [782, 255, 797, 298]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [281, 205, 296, 215]\nB: [697, 205, 716, 215]\nC: [691, 520, 704, 537]\nD: [498, 0, 538, 6]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/64ecb14e96931fa12d63d7d52e97e6551d9ab8794bec873526ad42f1cf64fd39.jpg", "reference_inputs/images/c236d9347f55c0c3ec5ba178d174cdfed7be3be0fb69abec27fac1b60b6cc6cf.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000026", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [288, 502, 311, 514]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [102, 247, 142, 314]\nB: [397, 260, 408, 300]\nC: [0, 242, 18, 312]\nD: [382, 270, 392, 298]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7ef8435af79d5039b6ea4551e5399f6861d16b59baa994c4a98a2e33c187dd61.jpg", "reference_inputs/images/a95ddbdca69a13640b9d9595bb317ae802f40659a5f39d4e97c1ceba66babbbb.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000027", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [102, 247, 142, 314]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [741, 62, 764, 78]\nB: [288, 502, 311, 514]\nC: [76, 430, 110, 443]\nD: [10, 458, 44, 470]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a95ddbdca69a13640b9d9595bb317ae802f40659a5f39d4e97c1ceba66babbbb.jpg", "reference_inputs/images/7ef8435af79d5039b6ea4551e5399f6861d16b59baa994c4a98a2e33c187dd61.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000029", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [480, 491, 494, 513]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [508, 276, 534, 295]\nB: [351, 255, 439, 312]\nC: [651, 260, 887, 352]\nD: [462, 282, 498, 293]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6321bf478e6b07f7aac4caaad696c33ea7ba3bfde825bc08e5bfdc4c07385130.jpg", "reference_inputs/images/c33e1aeee82405ba47f657e32649e2e4e8d1e73b08d660d0fe811f4195c3ad2e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000030", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [956, 489, 976, 504]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [126, 270, 306, 321]\nB: [495, 248, 570, 314]\nC: [570, 277, 604, 298]\nD: [374, 270, 418, 298]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9dd94905f247910ef62de3c8e5a8092a01734659fe932b42faf400b0acc57c3a.jpg", "reference_inputs/images/170de95388c1beacd301d8b4deddad64ffae0d5a318f315340f7ecf49a4a5c07.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000031", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000031", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [126, 270, 306, 321]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [981, 537, 1003, 560]\nB: [956, 489, 976, 504]\nC: [991, 387, 1005, 403]\nD: [966, 454, 982, 470]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/170de95388c1beacd301d8b4deddad64ffae0d5a318f315340f7ecf49a4a5c07.jpg", "reference_inputs/images/9dd94905f247910ef62de3c8e5a8092a01734659fe932b42faf400b0acc57c3a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000032", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [475, 478, 490, 493]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [314, 262, 386, 304]\nB: [37, 261, 332, 344]\nC: [690, 262, 701, 295]\nD: [429, 170, 538, 314]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6dbfd15899345da323d2701eef481ffd7f42d3c6927e0fa020b0dae7b21106da.jpg", "reference_inputs/images/dd84ec2bafb16bb213460e7a124814e17d0f7d9c46fed3791c34b8ea12eba9b2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000039", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000039", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [240, 231, 257, 245]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [152, 261, 163, 296]\nB: [753, 258, 765, 298]\nC: [254, 261, 264, 295]\nD: [162, 258, 173, 298]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5b432ec76c85bb5a57956a33adb0353060181098b08c679dfb6e3d264a7c7ff0.jpg", "reference_inputs/images/9a6e5f9f30ee237528c1c8bea12e3e2df67813b2f7eaf7df22f916e035be8caa.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000042", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000042", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [516, 487, 530, 500]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [680, 214, 1023, 390]\nB: [518, 282, 537, 291]\nC: [374, 269, 419, 298]\nD: [624, 270, 634, 292]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/464950ecad6dd709838bea0f9542bf17c5025d41f5e0c53dd783e936f9e8c24e.jpg", "reference_inputs/images/f3d0a4d31d16c204ff9a76a17bcb947c6cb4b7bbb5cf7f59f6e1d7ad3080d41a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000043", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000043", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [234, 246, 254, 254]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [516, 246, 534, 306]\nB: [450, 248, 467, 308]\nC: [531, 250, 546, 306]\nD: [719, 247, 740, 306]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3c008430886d13cc06b3af1ca92b44323c9fdc224aa91af5bf1a822dea3e92b5.jpg", "reference_inputs/images/9bd1be9135b490482662595471ecb13389cc21f0185e9afd81e9b1cf2152433c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000044", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000044", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [719, 247, 740, 306]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [239, 139, 257, 155]\nB: [231, 238, 252, 246]\nC: [250, 279, 270, 287]\nD: [690, 436, 704, 450]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9bd1be9135b490482662595471ecb13389cc21f0185e9afd81e9b1cf2152433c.jpg", "reference_inputs/images/3c008430886d13cc06b3af1ca92b44323c9fdc224aa91af5bf1a822dea3e92b5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000046", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000046", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [440, 122, 473, 196]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [69, 262, 265, 326]\nB: [508, 277, 525, 291]\nC: [458, 277, 482, 294]\nD: [544, 235, 607, 298]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7178e1b2d214b3c18b7b708a46724106443903040c88c0642dfca51bb5d6326b.jpg", "reference_inputs/images/aa247e83cbfa6a55e87f479c5d2e650e3f2266f2b3bb012623eb5ab140aa8602.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000048", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000048", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [670, 462, 686, 475]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [188, 250, 329, 316]\nB: [2, 251, 40, 302]\nC: [424, 282, 450, 293]\nD: [518, 281, 546, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2d80bc60677dd8b2a63c877219f78968fdfbf756f5998945b085e9e6ae8f4dab.jpg", "reference_inputs/images/948eb1874b239d46c2fb2d23b88f7be52db81266aea190904fab7655e01e84d4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000051", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000051", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [571, 280, 592, 292]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [612, 288, 643, 355]\nB: [390, 50, 399, 71]\nC: [505, 68, 535, 137]\nD: [512, 506, 546, 575]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2a4ced4ecbda3f0ad9bfde9034be5e0a49ec9bab9c712884affd3ad08ecb94fb.jpg", "reference_inputs/images/72e6066675405adba0630b30e491e278d7c25335b22456974782fe3e844d5eda.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000052", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000052", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [543, 558, 560, 575]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [811, 225, 1023, 362]\nB: [0, 211, 64, 350]\nC: [532, 258, 564, 292]\nD: [422, 274, 467, 303]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f0897f975c4069cd51679c11c42f130c4256ac116bcb15922beda8701b62fa59.jpg", "reference_inputs/images/1b5c7d35d99b4d90fc9acbf22dad919f289e379585dda663df500ef2c2dbccf2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000053", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000053", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [532, 258, 564, 292]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [486, 490, 498, 506]\nB: [522, 424, 537, 445]\nC: [543, 558, 560, 575]\nD: [481, 396, 493, 414]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1b5c7d35d99b4d90fc9acbf22dad919f289e379585dda663df500ef2c2dbccf2.jpg", "reference_inputs/images/f0897f975c4069cd51679c11c42f130c4256ac116bcb15922beda8701b62fa59.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000054", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000054", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [668, 494, 694, 509]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [706, 262, 738, 295]\nB: [80, 252, 109, 300]\nC: [855, 249, 871, 299]\nD: [654, 266, 662, 293]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a1d1537e3be8850fe6f3702a692d2159724b057ca4725db2e331af7a19f6bdbf.jpg", "reference_inputs/images/fc2f455a3a134b67bdb86f8967c4e888682308e77817c4a1c0ede8e7b7ffb7cb.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000055", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000055", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [706, 262, 738, 295]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [715, 294, 735, 302]\nB: [668, 494, 694, 509]\nC: [674, 35, 686, 54]\nD: [664, 341, 682, 350]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fc2f455a3a134b67bdb86f8967c4e888682308e77817c4a1c0ede8e7b7ffb7cb.jpg", "reference_inputs/images/a1d1537e3be8850fe6f3702a692d2159724b057ca4725db2e331af7a19f6bdbf.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000057", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000057", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [390, 279, 430, 297]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [438, 83, 472, 157]\nB: [383, 470, 421, 544]\nC: [500, 458, 533, 534]\nD: [542, 220, 614, 291]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4bef85cf9114015576227a64dbbbb0fb2abfbf1cc0ae73d509da424e5b2b3f87.jpg", "reference_inputs/images/94afe022a1b3abd64effa4e05540b2eb9a3a7625b0e9302f064304f48566d76b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000058", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000058", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [668, 450, 690, 462]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [534, 260, 545, 299]\nB: [19, 232, 66, 310]\nC: [766, 230, 810, 318]\nD: [972, 263, 986, 302]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ea0b07874c421a30ce3bde9476a6b09e6eeb1f527b7870a89b564f5017a1476a.jpg", "reference_inputs/images/d396ca39d542235fbcc69b5858e5f5a773de1152513842df5f369577757d93ca.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000059", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000059", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [534, 260, 545, 299]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [919, 384, 952, 394]\nB: [259, 186, 283, 197]\nC: [713, 88, 730, 103]\nD: [668, 450, 690, 462]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d396ca39d542235fbcc69b5858e5f5a773de1152513842df5f369577757d93ca.jpg", "reference_inputs/images/ea0b07874c421a30ce3bde9476a6b09e6eeb1f527b7870a89b564f5017a1476a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000061", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000061", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [610, 220, 636, 320]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [686, 134, 703, 146]\nB: [716, 499, 734, 514]\nC: [666, 545, 682, 563]\nD: [718, 29, 736, 46]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/576d8be8f3dbf01571e62e0948abed9b4f3a1622dc4e9a7e3af95ef90daf400e.jpg", "reference_inputs/images/fa6d3a482a58eb0772ff01e0b51f94db60268aa1c2e36ef81f7015f2b296605a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000062", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000062", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [237, 286, 254, 306]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [862, 250, 897, 304]\nB: [962, 260, 1023, 338]\nC: [645, 242, 667, 306]\nD: [547, 243, 571, 306]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1432e33aabdc6f9551f47e1894e23338857b9c07f84cbe0f83f482c8a61515cd.jpg", "reference_inputs/images/d94eb62b243c1682b6d8d0632a483769df092cb170aaaf44b82e0cf25a2ef858.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000063", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000063", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [862, 250, 897, 304]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [257, 0, 273, 22]\nB: [700, 0, 714, 22]\nC: [237, 286, 254, 306]\nD: [244, 407, 265, 419]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d94eb62b243c1682b6d8d0632a483769df092cb170aaaf44b82e0cf25a2ef858.jpg", "reference_inputs/images/1432e33aabdc6f9551f47e1894e23338857b9c07f84cbe0f83f482c8a61515cd.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000066", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000066", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [274, 541, 294, 558]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [298, 262, 376, 305]\nB: [227, 265, 246, 294]\nC: [196, 259, 220, 297]\nD: [10, 247, 33, 303]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/beb4a4da294a37a291b7007f52317ea5ec3c945d61ec032b2127c546ed28efa1.jpg", "reference_inputs/images/bd6ea9d7f0ba2700e5a7f1012bb7ccc6ed5ade16fe5b4c58580835f1e69945c8.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000067", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000067", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [507, 518, 524, 534]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [399, 280, 460, 295]\nB: [477, 268, 601, 327]\nC: [936, 276, 983, 296]\nD: [0, 266, 120, 342]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7ca6b03f8931b1ab6f7da5d8418422538d148f0d692e684c2e49e2790acb8b6c.jpg", "reference_inputs/images/d805d4a2c5ab72663e1777357194e745b5bdfd434697b3d4f6a88ba3b05e17b3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000068", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000068", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [399, 280, 460, 295]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [507, 518, 524, 534]\nB: [525, 416, 537, 436]\nC: [479, 398, 492, 417]\nD: [466, 447, 490, 458]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d805d4a2c5ab72663e1777357194e745b5bdfd434697b3d4f6a88ba3b05e17b3.jpg", "reference_inputs/images/7ca6b03f8931b1ab6f7da5d8418422538d148f0d692e684c2e49e2790acb8b6c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000069", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000069", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [794, 274, 799, 297]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [754, 144, 770, 154]\nB: [287, 418, 304, 433]\nC: [748, 278, 829, 335]\nD: [871, 161, 897, 182]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6e866ae608c1490d728a04e4302a716c447fcdebcd4daf09bd5da54aee35bfc4.jpg", "reference_inputs/images/43d5746cf60013f61d81966c0c6e49fe26429e21c2249a6021ac2a039deb0240.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000070", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000070", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [885, 238, 911, 311]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [674, 35, 686, 54]\nB: [664, 341, 682, 350]\nC: [668, 494, 694, 509]\nD: [715, 294, 735, 302]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e4739660eee3426aaaee807c0a42cbc1c41698bd203b45348c9d2c48984902ee.jpg", "reference_inputs/images/a1d1537e3be8850fe6f3702a692d2159724b057ca4725db2e331af7a19f6bdbf.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000071", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000071", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [350, 264, 370, 298]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [626, 26, 636, 47]\nB: [247, 0, 263, 18]\nC: [504, 43, 512, 60]\nD: [558, 474, 566, 498]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/16652bd8a0749e2b8903fd22753cff94fa7d4fe455aeb70555cc2ffe81a55955.jpg", "reference_inputs/images/9387fb07c1ea73e8176f603aa4bc4630b33ca3cd16a48c5c1865d623fa341a4d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000073", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000073", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [746, 234, 770, 310]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [286, 404, 306, 416]\nB: [189, 457, 210, 473]\nC: [387, 469, 418, 537]\nD: [735, 436, 756, 450]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b0af3bb36620b6a360c476abd2224bdf5db62721179541bcc51903f4cb00b64f.jpg", "reference_inputs/images/5c528b787bb0bb05fa3e469f6e93e9a80d6e2cc70504ec83808d14638d39d228.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000074", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000074", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [702, 231, 745, 318]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [291, 111, 302, 122]\nB: [669, 263, 690, 272]\nC: [274, 263, 297, 272]\nD: [640, 213, 655, 223]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f7d45e66b52b56f9d2fe3a416f385af8583f41388a32feefa9e92bff4b30e63f.jpg", "reference_inputs/images/676873394784abd35cfd3013a7985d1be5dae73b6d656086834092c90cdd90ab.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000075", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000075", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [188, 251, 210, 262]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [797, 254, 813, 299]\nB: [178, 262, 188, 294]\nC: [304, 264, 312, 294]\nD: [744, 262, 754, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/01bec1e6063f2abdaa5a2a82f0cc79873e8527ba96cf4d200ccc3916f0c1bc18.jpg", "reference_inputs/images/502fe2b7bf5d1fe5c44e96178c57cdc8f6cda7fad2cf013c6853ede35a442838.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000076", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000076", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [744, 262, 754, 294]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [188, 251, 210, 262]\nB: [693, 401, 714, 410]\nC: [286, 197, 305, 205]\nD: [736, 230, 757, 239]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/502fe2b7bf5d1fe5c44e96178c57cdc8f6cda7fad2cf013c6853ede35a442838.jpg", "reference_inputs/images/01bec1e6063f2abdaa5a2a82f0cc79873e8527ba96cf4d200ccc3916f0c1bc18.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000081", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000081", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [798, 479, 813, 498]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [151, 271, 170, 287]\nB: [0, 270, 42, 313]\nC: [112, 276, 135, 290]\nD: [192, 279, 217, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/10e0caccd78914ca0540a51e20bb43e216355caae6de0b951c41c6400442c2ba.jpg", "reference_inputs/images/3f6b8fcca774c8e8aea257100830169913e8608cc0bcabc19239f3bc3545c962.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000082", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000082", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [112, 276, 135, 290]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [878, 495, 898, 511]\nB: [822, 438, 848, 450]\nC: [798, 479, 813, 498]\nD: [773, 398, 785, 416]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3f6b8fcca774c8e8aea257100830169913e8608cc0bcabc19239f3bc3545c962.jpg", "reference_inputs/images/10e0caccd78914ca0540a51e20bb43e216355caae6de0b951c41c6400442c2ba.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000084", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000084", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [553, 258, 564, 300]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [116, 77, 143, 91]\nB: [212, 46, 237, 62]\nC: [118, 507, 146, 524]\nD: [236, 475, 255, 492]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d1a2e0d8a7b1234553fb0f2b6b49098b855cc90c859e55d2b0ab0d19a2abd575.jpg", "reference_inputs/images/f21e2f3ad9c76b4b7cd93f585cefa4c1a8e7d0563290f1be9515c9b7766eb3ff.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000085", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000085", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [529, 536, 545, 558]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [685, 270, 744, 310]\nB: [595, 274, 626, 294]\nC: [740, 276, 763, 295]\nD: [659, 270, 678, 289]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c2ac00d149129e254d16e74d07510cd0bfc6b336586669eb5e52ae72e11897d0.jpg", "reference_inputs/images/74f386330c7d9af7076f742a7d66db3df65f2fcc5027d9e6f0c2bb2d6c20ab41.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000086", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000086", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [659, 270, 678, 289]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [440, 530, 459, 557]\nB: [529, 536, 545, 558]\nC: [502, 457, 527, 469]\nD: [472, 406, 485, 428]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/74f386330c7d9af7076f742a7d66db3df65f2fcc5027d9e6f0c2bb2d6c20ab41.jpg", "reference_inputs/images/c2ac00d149129e254d16e74d07510cd0bfc6b336586669eb5e52ae72e11897d0.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000087", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000087", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [166, 222, 208, 314]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [261, 398, 283, 409]\nB: [284, 329, 310, 338]\nC: [712, 301, 729, 312]\nD: [681, 110, 695, 122]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a4819079d3821c574871437de2ea57a167191c3f99cd65d3aac541cb6ba59da1.jpg", "reference_inputs/images/f64871178c49e7c522c6b88ab61796420ba48e16d787c5ab3deaac380c90cc48.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000088", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000088", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [913, 114, 984, 147]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [444, 282, 472, 296]\nB: [242, 272, 250, 293]\nC: [770, 229, 810, 319]\nD: [331, 278, 382, 302]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ebb9d6feb3877e846aa11fbb3a98712e537e5a445bc68542527fce62c6ac1ee1.jpg", "reference_inputs/images/7851f9a2784ea4549a5d2fa4c73ca071e9fc113790d62d9784b7186ed916dc5c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000089", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000089", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [696, 266, 718, 300]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [466, 20, 475, 42]\nB: [604, 467, 615, 485]\nC: [589, 474, 598, 495]\nD: [564, 57, 574, 73]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b40fb3da6d31f67dfc0380d6881e015176910045bdf7769a44ee54f580242034.jpg", "reference_inputs/images/7d1a04beac0a65e8d5644b0e96e0ffa9971c7899972461738dbebcdb4f5e7b4e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000090", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000090", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [528, 110, 597, 162]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [226, 278, 258, 295]\nB: [0, 270, 110, 323]\nC: [353, 274, 370, 288]\nD: [650, 253, 680, 302]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0dc5e9dd4c1857374a67d4565eaa428149d3d5ceb0c6e075ca173641439171ec.jpg", "reference_inputs/images/77d355cbe295b4ce12ba03e43795081814bf235837271b4caadc6eb5ded3c1d5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000091", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000091", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [503, 400, 514, 417]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [708, 219, 1023, 382]\nB: [478, 237, 526, 298]\nC: [46, 243, 74, 305]\nD: [385, 278, 442, 302]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4bbaf7b875b927803278ab0e3bbe127dc986d7e632e71250f020a2395eb683e5.jpg", "reference_inputs/images/d58c9ae08eef29500a5bf2870e2b136cf317de0992e07d5be2a8fc5ab35b65a2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000093", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000093", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [624, 535, 634, 559]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [633, 276, 646, 294]\nB: [352, 274, 366, 293]\nC: [793, 206, 894, 334]\nD: [990, 246, 1016, 310]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d48c4d949dccb09b3810f2e3a542562e8a29926ed4a51421be786d44c547bfa1.jpg", "reference_inputs/images/c5b600c09a6d37cd46165ccf13dda6e2f27b7d463018639951817900fbdef907.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000097", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000097", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [461, 467, 480, 506]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [465, 268, 482, 290]\nB: [440, 276, 475, 299]\nC: [530, 260, 559, 291]\nD: [674, 258, 855, 332]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f221b3b69de7e07be0c49324d9ef559f3a8f8d4b2de7913d4e7f3da6c599413a.jpg", "reference_inputs/images/43b0a3fe42f2acc793e99dea4dd8389f524b8f3d47f09236481fb1cc1befeb13.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000100", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000100", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [455, 463, 475, 501]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [498, 268, 517, 290]\nB: [519, 281, 548, 294]\nC: [290, 225, 402, 304]\nD: [766, 253, 783, 300]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a1c511f49696a445fff0f4cd0794fc2a08afe9322f7069f11eeea41c460a318f.jpg", "reference_inputs/images/abcd00b1ecf273170e58d374e635af5b0486c70f398e13ec84205b56af863cca.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000101", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000101", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [346, 206, 727, 357]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [418, 385, 456, 458]\nB: [672, 91, 685, 111]\nC: [383, 168, 419, 241]\nD: [474, 243, 528, 377]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e79483d5292ccf91b183bf38925461115a01c0a659d376c28c30e309b4321ab2.jpg", "reference_inputs/images/edbffda0f2b67c463fd64cae867842b22130d1b4b709eeaf66f07cd968060492.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000103", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000103", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [528, 455, 545, 483]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [349, 234, 430, 301]\nB: [507, 278, 535, 297]\nC: [597, 113, 867, 327]\nD: [471, 280, 499, 293]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/645c9bf1b25cd76a2843ff2e95d1d9d7f6d29a09e1a67527a227fd74797152f7.jpg", "reference_inputs/images/dbdbafd1873f8374fc2da3b5dc8e39767781e722be8ed25fdcf2da41eac4a224.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000104", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000104", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [768, 205, 797, 286]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [338, 70, 354, 89]\nB: [754, 293, 775, 300]\nC: [434, 22, 445, 40]\nD: [748, 307, 770, 320]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9dda419c3bfb342e9f766eac02b43633a9db8704e47d2067503de3aebe5e004a.jpg", "reference_inputs/images/17e8acdee8139296e74054ad46f0b8e4ac3f31e20e58c6ad512a0f49c246c007.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000105", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000105", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [755, 290, 774, 298]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [72, 230, 122, 314]\nB: [893, 268, 905, 299]\nC: [498, 235, 526, 314]\nD: [432, 242, 450, 307]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e52c16d1d979c88426a29b8a8a019a1af495fab894de56409f5f420abf1b6662.jpg", "reference_inputs/images/46a777e3a51d6c39f60dd75a2a9a9f7a3b523703d394b82437cbb2aba0121458.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000107", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000107", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [467, 454, 479, 469]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [0, 246, 259, 361]\nB: [246, 254, 342, 313]\nC: [494, 271, 523, 296]\nD: [827, 247, 845, 303]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a0615fc5d592de64278ec88960745f2edc782ab5e116e4a4c1da9eb5e026d467.jpg", "reference_inputs/images/9d12e0d7e2f66043454da17dd4d6bf2b971c6c1620f747bf4be58db60909059e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000110", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000110", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [444, 277, 477, 298]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [500, 405, 538, 479]\nB: [558, 289, 597, 362]\nC: [446, 306, 477, 373]\nD: [766, 101, 836, 134]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b800d7a40eea13b4f7e2a8f5023f5f5668249c4876513c6be9737791b96918ee.jpg", "reference_inputs/images/53ad558c6a7324be4c3615e1313f2fcdd5099cfd3e3c380156235603895261b6.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000112", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000112", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [334, 518, 371, 536]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [276, 261, 539, 354]\nB: [962, 269, 992, 298]\nC: [701, 276, 710, 298]\nD: [852, 278, 898, 306]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a6d4189bd6cf025b8ad0727de720ce78c370467fdcb90a9a16b44d4dfbfc1062.jpg", "reference_inputs/images/48f58850513c81ed977c5c8c574a23c6de9d7218bc84a9d8d11a36c15e3b65cf.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000114", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000114", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [663, 278, 722, 306]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [880, 280, 954, 311]\nB: [0, 461, 66, 500]\nC: [364, 0, 420, 96]\nD: [490, 458, 562, 497]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9dabade18b21c6aa67c58127d7a894f94f432a931037529764f20ab13e6c9cdc.jpg", "reference_inputs/images/0fef7bf43103cdecd4e2c7ba577149ffb74ee952c482f428f478d115ae197001.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000115", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000115", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [502, 525, 536, 544]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [420, 258, 455, 293]\nB: [853, 278, 930, 297]\nC: [481, 262, 585, 303]\nD: [841, 246, 858, 300]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4ed1be0fb3ebdf7e2a48b580325d5608c69ed643f203b18102b83cd836ca3c64.jpg", "reference_inputs/images/aafc73a7aa22bec3f1d4f00952a95fe97681a40c8c7274f4dc3e6d3f563faebf.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000116", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000116", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [628, 194, 677, 271]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [417, 283, 445, 295]\nB: [370, 283, 401, 295]\nC: [222, 274, 267, 300]\nD: [386, 269, 417, 292]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e5cdc2ffbe31b0bf6dc8b74abe10ba04fd31c30e5a07a413559f73a37a56dd6e.jpg", "reference_inputs/images/8d350c70f2db8cfcc325e55d17672ba8feba9ccb2d80d49bf961be1d03c18840.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000118", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000118", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [496, 515, 528, 575]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [424, 279, 445, 294]\nB: [386, 279, 413, 296]\nC: [470, 283, 501, 298]\nD: [311, 278, 388, 306]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4b2fd8d6119b02597cdc72ae4e2b836c93dad3a4121c6f7b36fcfe018d04acf4.jpg", "reference_inputs/images/fd1a741907a15ac959be34c50c789abef4ee72af9a0929d747d349ffbe6f6efb.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000121", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000121", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [114, 260, 516, 360]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [562, 72, 601, 146]\nB: [382, 148, 420, 222]\nC: [387, 465, 418, 533]\nD: [563, 336, 598, 408]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ffc99a2f94c1bacb24b4a7edbe241b6633b8e78ce33b693a974a5e5d4b497c19.jpg", "reference_inputs/images/a41e4ff6460129e4dc8ecebf164fd651da80c2cf4e41eb8c4f8a1269e203aea6.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000124", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000124", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [434, 22, 445, 40]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [585, 258, 598, 296]\nB: [963, 262, 978, 295]\nC: [178, 271, 310, 307]\nD: [470, 263, 482, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/17e8acdee8139296e74054ad46f0b8e4ac3f31e20e58c6ad512a0f49c246c007.jpg", "reference_inputs/images/29fc872b749e944ac734776f6892fe8e4600cf609cbd0a71f1ff2dda5f3dc4e4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000127", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000127", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [220, 256, 226, 297]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [680, 108, 696, 121]\nB: [280, 434, 300, 449]\nC: [234, 54, 254, 71]\nD: [116, 86, 145, 100]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4183a0b6de334d0e851eef8ba64a5005f51b443781882f4bdd69e0315019eef6.jpg", "reference_inputs/images/b54de1266305b9b20f5b628edb6cc3be8bf303f5eaaa4ba4235d674b186ebab7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000132", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000132", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [303, 245, 325, 308]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [692, 12, 706, 32]\nB: [240, 375, 262, 386]\nC: [848, 387, 872, 403]\nD: [272, 39, 288, 61]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/010870be1ff3452306b89e08d9db3ca31baec1660cdaf57e6ddea918045c7365.jpg", "reference_inputs/images/54fe527ada2d1060e9abe51e500c229043388ec7321da0fcfb22ea61353b10b0.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000134", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000134", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [458, 246, 478, 306]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [258, 416, 275, 433]\nB: [642, 139, 667, 150]\nC: [286, 139, 314, 150]\nD: [231, 480, 254, 496]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/55aca58644f2001dd1fad5ed84021bd3b993bd877b87301b5ec3265ac716a94a.jpg", "reference_inputs/images/615f44861ee6ee580e557aa2ff2f6a04cff6e3258f6acafed97c60055df847c6.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000135", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000135", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [444, 262, 471, 291]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [966, 454, 982, 470]\nB: [981, 537, 1003, 560]\nC: [956, 489, 976, 504]\nD: [991, 387, 1005, 403]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f715261c31086ae963af6c95167737fcc81cedadfdf0f8460aaa5262d2a8969f.jpg", "reference_inputs/images/9dd94905f247910ef62de3c8e5a8092a01734659fe932b42faf400b0acc57c3a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000136", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000136", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [446, 259, 452, 301]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [116, 86, 145, 100]\nB: [234, 54, 254, 71]\nC: [680, 108, 696, 121]\nD: [280, 434, 300, 449]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/adc9f84d780eac8c352b4449a1a123832c0703b5e33227e663e90dc2368f5759.jpg", "reference_inputs/images/b54de1266305b9b20f5b628edb6cc3be8bf303f5eaaa4ba4235d674b186ebab7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000137", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000137", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [899, 206, 950, 328]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [291, 111, 302, 122]\nB: [640, 213, 655, 223]\nC: [669, 263, 690, 272]\nD: [274, 263, 297, 272]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3f33e168375781384fa6ff8115c76325dcb459dd1f23d2582f540877beb2c83e.jpg", "reference_inputs/images/676873394784abd35cfd3013a7985d1be5dae73b6d656086834092c90cdd90ab.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000139", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000139", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [462, 476, 477, 488]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [508, 275, 535, 296]\nB: [443, 269, 480, 298]\nC: [274, 275, 367, 307]\nD: [577, 277, 637, 302]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4e33cdd8ea49b40c1dc79aa8a47c3247f0846a5bfc6bf23c11356d4b495d047c.jpg", "reference_inputs/images/e8f86a144ff07fab1b6ed4595caeffbf95d4782f342c62d61c804499bb0065bf.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000140", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000140", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [686, 202, 1000, 350]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [449, 530, 465, 553]\nB: [507, 403, 519, 420]\nC: [534, 503, 549, 521]\nD: [462, 476, 477, 488]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e8f86a144ff07fab1b6ed4595caeffbf95d4782f342c62d61c804499bb0065bf.jpg", "reference_inputs/images/4e33cdd8ea49b40c1dc79aa8a47c3247f0846a5bfc6bf23c11356d4b495d047c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000141", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000141", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [810, 510, 835, 526]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [345, 269, 358, 292]\nB: [711, 268, 722, 291]\nC: [406, 269, 411, 291]\nD: [721, 270, 735, 292]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c2b178c8ac7c4c4624b7eb47fa740c6d986fa1d53d0ecf98dc8aa8134aaf6ed4.jpg", "reference_inputs/images/19c0e00856338e2eb25d3e4fcbf3f9915e89330932d73e3988585d074b88b126.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000142", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000142", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [280, 161, 297, 172]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [910, 302, 929, 367]\nB: [0, 185, 28, 245]\nC: [329, 260, 381, 286]\nD: [242, 214, 270, 282]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4884734f4435bb37db842686b2f0da57786ff2d95dd3fab2f84a75a57e1a72e1.jpg", "reference_inputs/images/30e6bd80fec7eeb677a4b663c71af085dbdb8ae9f65814a556f247652e14e01b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000144", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000144", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [602, 244, 612, 262]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [54, 258, 84, 297]\nB: [851, 183, 910, 336]\nC: [938, 181, 1023, 344]\nD: [202, 267, 279, 309]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bc877a1e94d97c772c82e01761652fcb636ac32f087a441b9e822c775cb8b1d5.jpg", "reference_inputs/images/6f8268f4a1b4de318f12616aabc34ebe6f28ba8ef563703d6abe227cc94b8015.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000150", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000150", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [748, 237, 770, 248]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [447, 264, 474, 290]\nB: [295, 267, 306, 294]\nC: [774, 261, 782, 295]\nD: [26, 263, 38, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6e4af1e6834a23f481da3fb5481f4f060494ac2b58b1035dff9f3ad5ad3f2661.jpg", "reference_inputs/images/1f4e5bfc1dc07dd2fc6ab971db43666d4bee8f303a3194ecb836bf23599c8aa9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000154", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000154", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [555, 239, 570, 310]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [731, 425, 750, 438]\nB: [509, 218, 531, 227]\nC: [291, 242, 307, 252]\nD: [677, 403, 691, 414]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/425cf972cabd45e0522c7a870edb16f10b0dacffe2cf1406bcec77a1b396b7a9.jpg", "reference_inputs/images/53ad558c6a7324be4c3615e1313f2fcdd5099cfd3e3c380156235603895261b6.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000155", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000155", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [518, 243, 542, 310]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [257, 272, 274, 281]\nB: [282, 415, 302, 426]\nC: [693, 416, 714, 426]\nD: [678, 494, 690, 514]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/61ea2cfd8bcaf6fa9d4643bea76218762a46fbd3e21bf5b0b0150562730bbad4.jpg", "reference_inputs/images/83241a33f90d5353ade9d303ca3c190d92e2333f5f674957b328465ee8b42759.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000157", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000157", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [354, 251, 371, 302]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [673, 50, 691, 64]\nB: [750, 0, 766, 15]\nC: [252, 61, 270, 78]\nD: [822, 108, 848, 122]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/eb97a3246d862464d2b51b720f73d209b8d3a46bb24e54d9e5a2b762b5dfa240.jpg", "reference_inputs/images/d23def54bc38c6919f30f1cf05d51e32af4daf8d2413d5881837e1e5da81ad0b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000159", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000159", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [288, 430, 302, 446]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [2, 242, 33, 307]\nB: [98, 253, 125, 298]\nC: [297, 270, 302, 291]\nD: [645, 268, 652, 291]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fa77697dfc4085fced112433ab375a16fb84c526f445ccec94d4d1edbce39c59.jpg", "reference_inputs/images/5a50298408b5a44380b325f06a63c9e6f92a85c81613f17465ad2b8e5b4836a3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000161", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000161", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [249, 66, 269, 80]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [690, 245, 706, 306]\nB: [630, 247, 649, 306]\nC: [892, 246, 914, 304]\nD: [572, 252, 589, 304]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e4457e1a53e081868651ab4f5bdb7954e2e6197ba1a7b0805272ee9e12908fcd.jpg", "reference_inputs/images/fb65f7995f334114946a52d3172118daba99133ed34921956d2365e3e709e2a6.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000163", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000163", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [754, 401, 774, 411]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [506, 262, 519, 298]\nB: [906, 247, 934, 308]\nC: [556, 245, 572, 306]\nD: [446, 220, 478, 316]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7312f49d0b403f90f880885c10c3f564ca6b1d2fa5ce2b847dca574c103831d4.jpg", "reference_inputs/images/2506433c82e3e48ac5f1db08b9f7e3f477f766a7bfcdea59cc70da55f5e29026.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000166", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000166", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [459, 475, 471, 490]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [637, 282, 654, 296]\nB: [811, 279, 878, 302]\nC: [704, 243, 807, 303]\nD: [902, 274, 910, 298]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5f78e355e5e15fb436335e1dcade9259514f85cdeef39a550afcc8bcc9b80649.jpg", "reference_inputs/images/3dcc3b6ab37085f8ee320458b61a7cf7f2eab3f312e6f84ddc2d277d5a81c266.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000167", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000167", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [811, 279, 878, 302]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [475, 554, 494, 575]\nB: [506, 456, 522, 485]\nC: [459, 475, 471, 490]\nD: [510, 537, 526, 560]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3dcc3b6ab37085f8ee320458b61a7cf7f2eab3f312e6f84ddc2d277d5a81c266.jpg", "reference_inputs/images/5f78e355e5e15fb436335e1dcade9259514f85cdeef39a550afcc8bcc9b80649.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000169", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000169", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [218, 233, 246, 314]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [285, 63, 300, 77]\nB: [244, 234, 261, 241]\nC: [702, 511, 718, 531]\nD: [298, 210, 313, 220]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5c64e995683b1de21c283c671349dac82351d1b3a64375df3ccdb6bffaedb802.jpg", "reference_inputs/images/223cbee8041d6f1a55407a61b86a321c281765434511cd82b74d02a28bad5c15.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000173", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000173", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [849, 467, 922, 536]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [34, 245, 103, 297]\nB: [349, 272, 470, 317]\nC: [657, 253, 678, 302]\nD: [503, 258, 514, 298]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9a637268ccc66c4e35c4ed2e56daf521b49a57fefec9c91252cab14f54e236a3.jpg", "reference_inputs/images/6f0f1e07119f53abe8f08c28c43962d83b3debe85de2c64a6e89e21ebb95a02c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000174", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000174", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [787, 234, 812, 308]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [285, 45, 314, 60]\nB: [718, 246, 735, 254]\nC: [686, 346, 702, 357]\nD: [277, 68, 297, 86]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/73f3042a00b93c841d7c0bb1b50d101f51592807a9b7beca1af11b550ca2317b.jpg", "reference_inputs/images/15c5847d2bcd414f7eeb58fee91a029b0e12d4f80a9921c4543425250c494c5b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000175", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000175", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [624, 338, 634, 351]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [880, 270, 890, 301]\nB: [946, 271, 954, 291]\nC: [470, 267, 480, 303]\nD: [515, 274, 520, 299]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6bce99f6dd753d279a0d0644d9a053d9541c8135bb48b2a573fcfe67b36fa081.jpg", "reference_inputs/images/3e209e0da429c596a0bcf30b747b605a4d443dbf2a00e34eee9b83629e766597.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000176", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000176", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [230, 267, 295, 307]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [427, 489, 443, 510]\nB: [456, 395, 469, 414]\nC: [502, 497, 514, 514]\nD: [482, 438, 504, 448]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d12375d5e82700b64fab8fe32bb0439c4988961f226f86a61debff37c8aae019.jpg", "reference_inputs/images/872708a7c44b7a1d8d6ae30ce6249403dc1983e6cd71aa67668bc25bbec6a8a1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000177", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000177", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [238, 228, 286, 315]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [241, 324, 261, 333]\nB: [713, 88, 730, 103]\nC: [668, 450, 690, 462]\nD: [259, 186, 283, 197]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8d417a17c18faf40fc765c8b23d0f70cca61632b0e3f1a324e97c9d544bb2988.jpg", "reference_inputs/images/ea0b07874c421a30ce3bde9476a6b09e6eeb1f527b7870a89b564f5017a1476a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000180", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000180", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [466, 463, 480, 476]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [485, 233, 753, 396]\nB: [381, 274, 420, 298]\nC: [446, 282, 480, 290]\nD: [0, 257, 188, 352]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bec159013db8b15522ef914027d5b4939fa6aa1ec865aea29e1d74b66d73bbf3.jpg", "reference_inputs/images/22e82f3f5da3ca8e0a41507973a1054fc81a0436bd364afb7118546fe3ec6147.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000185", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000185", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [58, 245, 100, 314]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [289, 522, 307, 538]\nB: [686, 373, 702, 386]\nC: [930, 115, 954, 130]\nD: [257, 397, 274, 407]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6bde287cf09eccf13f1a7282d5f5bbecd74d66d92d3f8b16409fa292a063c87d.jpg", "reference_inputs/images/e5b411a68a7ef2e6b0baef87b94099e01703403ccf54d36ff3073d57e76ece61.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000186", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000186", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [173, 237, 215, 308]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [708, 20, 724, 38]\nB: [690, 162, 718, 172]\nC: [273, 123, 290, 138]\nD: [624, 389, 637, 399]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d6fbe93a4f5e999ad00e816341ccc35818ab5a6bd7ce3f2b31a5f3a913b357d5.jpg", "reference_inputs/images/cd60d6b1e87c07bad19308bec877e9c74bc0ee6bfc6dd24a1d8ee46e2d07f17d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000187", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000187", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [935, 244, 958, 306]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [666, 460, 690, 472]\nB: [256, 182, 274, 191]\nC: [287, 314, 309, 322]\nD: [275, 460, 298, 472]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e73ae81f2c52a4d3e7c1c26da6cd80d087797bd7073d6cf1fef9bfa935c7376f.jpg", "reference_inputs/images/0ce3f395dec63004673dd08739da1b4d01e42259f9328ea51089adf2b3c5c0db.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000191", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000191", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [510, 524, 524, 543]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [807, 249, 835, 302]\nB: [503, 270, 561, 310]\nC: [475, 278, 491, 290]\nD: [406, 280, 441, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/dd757b41e3a6615c95ab1be539de5dbc5ff534d425852717977a24e6142ba69e.jpg", "reference_inputs/images/a863c0425374d378735269f9a18b0e786addc41c4cfe1ec95b61f4e2c1feb0bd.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000192", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000192", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [644, 242, 668, 281]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [642, 139, 667, 150]\nB: [258, 416, 275, 433]\nC: [231, 480, 254, 496]\nD: [286, 139, 314, 150]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b7087cf8dfff20392c5cce16fb1cbc197f4b8a023fd6a1a14e127507670e1033.jpg", "reference_inputs/images/615f44861ee6ee580e557aa2ff2f6a04cff6e3258f6acafed97c60055df847c6.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000195", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000195", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [329, 242, 342, 252]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [425, 242, 454, 308]\nB: [489, 222, 520, 316]\nC: [730, 234, 758, 311]\nD: [406, 266, 417, 299]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ef9daea12006a2e51e5c982a5b280c67b50689e9b2b397d8028d0f4166e33386.jpg", "reference_inputs/images/82e49c092a52ba14ad4d6e963fee05e3b260013b8eddacfdcbdcbc606ff915ef.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000198", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000198", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [73, 254, 93, 308]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [241, 551, 258, 574]\nB: [247, 401, 270, 414]\nC: [214, 214, 237, 224]\nD: [278, 10, 298, 30]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ca81ce571e94b8df61dd054d0be0890873d8b7700b146926d53fe71f8e3c7405.jpg", "reference_inputs/images/d81274f557b0ca07c4521b0e223aee5057051066586632f9e207009e5fe30d47.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000201", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000201", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [379, 242, 402, 310]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [274, 215, 295, 225]\nB: [747, 449, 769, 464]\nC: [274, 362, 296, 374]\nD: [548, 182, 600, 313]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5e787882a87d24e352cab13c717c715f81757643a7d3e3a57a774b3fed9ac3ff.jpg", "reference_inputs/images/4d4be13a5256b9afde9a01a0399e95853f4c8c40a360410da68a3d90e6d97fa3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000203", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000203", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [146, 266, 166, 301]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [239, 158, 262, 170]\nB: [38, 148, 69, 166]\nC: [42, 540, 72, 564]\nD: [850, 178, 874, 194]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a6bbb87b287f5401a21be4253d6a9ba6d1b1fb7dd5a4446eec287e86133e81c1.jpg", "reference_inputs/images/0fef7bf43103cdecd4e2c7ba577149ffb74ee952c482f428f478d115ae197001.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000204", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000204", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [920, 172, 954, 184]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [515, 229, 547, 319]\nB: [871, 262, 885, 301]\nC: [432, 260, 443, 298]\nD: [454, 274, 466, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b68b6dbbc1bbfab91556c7af06bfa9a3d5ce616d1538f94c6a4c7b73338b2daf.jpg", "reference_inputs/images/fc4a796869a6465d6668f9ee17b73e88031099fc46ca0dc1820dd139d0835b3f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000207", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000207", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [275, 468, 290, 482]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [865, 248, 881, 299]\nB: [266, 260, 274, 294]\nC: [284, 266, 304, 294]\nD: [100, 250, 118, 302]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cec587f7f71d59f91723be411e06f634b2c38d94eef8cefcfe273dea4be0b7c4.jpg", "reference_inputs/images/214450fcb9ec8f693510b00eb2718a24d336b00136b94d6785fc63ff6db98864.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000208", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000208", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [635, 270, 641, 292]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [275, 468, 290, 482]\nB: [721, 167, 736, 177]\nC: [674, 78, 686, 93]\nD: [719, 456, 734, 468]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/214450fcb9ec8f693510b00eb2718a24d336b00136b94d6785fc63ff6db98864.jpg", "reference_inputs/images/cec587f7f71d59f91723be411e06f634b2c38d94eef8cefcfe273dea4be0b7c4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000213", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000213", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [694, 439, 706, 455]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [60, 250, 78, 301]\nB: [822, 250, 838, 301]\nC: [259, 263, 286, 294]\nD: [197, 256, 208, 296]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/aa0313535f6408ea007ddb45a0599098ed8f0188d9bf5d2fe8026c988710c120.jpg", "reference_inputs/images/fbdc39cc7686c6098f4439103a97584a6a002a045d6e03bb084cdca90a855d25.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000214", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000214", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [793, 274, 800, 298]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [648, 447, 659, 465]\nB: [244, 169, 261, 182]\nC: [422, 152, 459, 227]\nD: [288, 195, 303, 208]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4baf667324c9e2421a4e9669293e613fadbaf6dceb4a62dc6e78035cf572fcb6.jpg", "reference_inputs/images/65ab20dcbba598480d730324f3c9ceffe765f5fccfc0bf381aeb02abec648f0f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000216", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000216", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [427, 279, 436, 306]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [606, 140, 616, 152]\nB: [390, 6, 400, 29]\nC: [799, 130, 822, 142]\nD: [390, 462, 401, 482]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/02ee26196f052db1fd6094389ca8a203753d4dfdd0196d5aad127f63ffb772e3.jpg", "reference_inputs/images/8acf094b5d622819c031b510b0a37ae772303e14eb6cd58466b0dad41852fb44.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000220", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000220", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [0, 171, 350, 426]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [459, 475, 471, 490]\nB: [506, 456, 522, 485]\nC: [475, 554, 494, 575]\nD: [510, 537, 526, 560]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/423edee79517840c130357a8a1a08ab8664be9cec297a266fc992caea0152d92.jpg", "reference_inputs/images/5f78e355e5e15fb436335e1dcade9259514f85cdeef39a550afcc8bcc9b80649.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000221", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000221", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [315, 234, 350, 312]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [242, 33, 259, 48]\nB: [689, 439, 702, 455]\nC: [274, 242, 295, 252]\nD: [690, 242, 714, 252]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cf331cda1a5b2ad295fdcf40384ca6454fb53929613927098c1c408cd596a0b4.jpg", "reference_inputs/images/21ed4cefcbdc6fb4b7ebaacb0e26f1a0a7c235237e6407eb849c2b143f51a44a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000224", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000224", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [370, 324, 410, 398]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [789, 268, 819, 281]\nB: [624, 242, 638, 282]\nC: [834, 266, 855, 278]\nD: [932, 269, 986, 284]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5dbeddd1c3c7318355a2bef03ca4c4171d13aa2262ccc2af44a86fbae0d714a3.jpg", "reference_inputs/images/db28015dd827c7a87eca54c8bed03ee4194eed92a3b8f6f40966f1f3a8f4b38c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000226", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000226", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [408, 489, 418, 510]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [816, 269, 824, 294]\nB: [434, 271, 441, 296]\nC: [659, 282, 701, 297]\nD: [518, 269, 540, 293]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/38cbd1ff93865b744eb4349f981020bb46720241f0906414c243eb67eea9a2b1.jpg", "reference_inputs/images/9b83d6129aa61cdfbf7ff383b9a29ddbba8ea0418bbd5a27c7cb21866300128a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000227", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000227", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [434, 271, 441, 296]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [309, 190, 384, 232]\nB: [408, 489, 418, 510]\nC: [218, 94, 237, 107]\nD: [333, 39, 345, 55]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9b83d6129aa61cdfbf7ff383b9a29ddbba8ea0418bbd5a27c7cb21866300128a.jpg", "reference_inputs/images/38cbd1ff93865b744eb4349f981020bb46720241f0906414c243eb67eea9a2b1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000228", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000228", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [598, 485, 610, 501]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [554, 278, 579, 294]\nB: [500, 264, 544, 301]\nC: [331, 277, 393, 302]\nD: [411, 274, 441, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/982b4a63b2edbf1c4acd3dca191ce0fee7ef5a177a699d51862a19ce21f26430.jpg", "reference_inputs/images/876e52177c54d2f1a987ee7d6b60aef790c909d21871bccc19c29ad50f3249e4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000231", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000231", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [937, 258, 946, 297]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [288, 502, 311, 514]\nB: [10, 458, 44, 470]\nC: [76, 430, 110, 443]\nD: [741, 62, 764, 78]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6668b014712035c16c55c9d346ec9d65f9aea0505cf77f2f90be5d06ec528252.jpg", "reference_inputs/images/7ef8435af79d5039b6ea4551e5399f6861d16b59baa994c4a98a2e33c187dd61.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000232", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000232", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [400, 268, 456, 305]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [723, 475, 738, 490]\nB: [776, 477, 790, 492]\nC: [790, 392, 801, 407]\nD: [741, 525, 758, 547]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6c6c3ed7f711125d8ad4f91ad47c76a1099f952f0d52c2a0da03758335a0ec82.jpg", "reference_inputs/images/12ba627932e572d5a61af9cd22f432b21d00b9f8a92045064ca84ac488070f7c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000233", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000233", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [714, 397, 734, 406]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [712, 264, 721, 293]\nB: [986, 234, 1014, 305]\nC: [762, 256, 795, 298]\nD: [679, 262, 688, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/acba93faab2b2b94bd8f121833dcdc0686ddae44c5819cbb8c7ea4fc08b05082.jpg", "reference_inputs/images/51c4243b98d1a3e1b646c00d0570aefc32d478d1128648c52c2f84bc12b8f8ff.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000234", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000234", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [986, 234, 1014, 305]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [664, 130, 683, 142]\nB: [715, 82, 735, 94]\nC: [714, 397, 734, 406]\nD: [674, 291, 690, 299]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/51c4243b98d1a3e1b646c00d0570aefc32d478d1128648c52c2f84bc12b8f8ff.jpg", "reference_inputs/images/acba93faab2b2b94bd8f121833dcdc0686ddae44c5819cbb8c7ea4fc08b05082.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000239", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000239", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [594, 529, 609, 547]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [504, 273, 518, 288]\nB: [530, 276, 562, 298]\nC: [562, 278, 579, 292]\nD: [459, 276, 482, 291]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f836ca059921239eb42024ba8423dd2007a5475b13086e3b198e74a73afe9fb5.jpg", "reference_inputs/images/6735f98cc46d273f973863ea4bee7b2468bdd114361efff6508e8164bd3036ee.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000240", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000240", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [462, 470, 474, 487]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [252, 250, 407, 330]\nB: [500, 268, 538, 298]\nC: [422, 275, 449, 293]\nD: [549, 279, 571, 293]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/dd4850896eac74d7cb7eb6106ed750d6f3f7a800b678e68b687900ca4ba4251e.jpg", "reference_inputs/images/9ad8a50cccc69d4d19587748ca75a29e617773e0bff87ecdaec8ddeb22d583f1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000241", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000241", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [422, 275, 449, 293]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [446, 529, 466, 551]\nB: [521, 397, 531, 413]\nC: [508, 518, 523, 539]\nD: [462, 470, 474, 487]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9ad8a50cccc69d4d19587748ca75a29e617773e0bff87ecdaec8ddeb22d583f1.jpg", "reference_inputs/images/dd4850896eac74d7cb7eb6106ed750d6f3f7a800b678e68b687900ca4ba4251e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000245", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000245", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [65, 210, 144, 314]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [390, 500, 411, 510]\nB: [510, 274, 530, 283]\nC: [827, 410, 910, 449]\nD: [270, 475, 294, 485]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ce18c7291cac5b524f1a529a27f2a334f06d216fc6d67f3a8779cac876076b76.jpg", "reference_inputs/images/1a395da19feab8dcb1c6ef71df4563096778af9dbb7f71b0efcc8acea44a97fd.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000250", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000250", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [616, 250, 622, 302]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [754, 401, 774, 411]\nB: [278, 220, 300, 230]\nC: [677, 359, 690, 370]\nD: [209, 402, 233, 411]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/adb1110f0b44df30c1bee32ce544b2e18c85062beb4d45e1883fabb9dac746ba.jpg", "reference_inputs/images/7312f49d0b403f90f880885c10c3f564ca6b1d2fa5ce2b847dca574c103831d4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000253", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000253", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [830, 242, 879, 315]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [258, 212, 274, 226]\nB: [702, 294, 714, 309]\nC: [679, 532, 693, 552]\nD: [647, 410, 660, 421]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b61aa062beb20a503ef9818e5a8cd77c939d21027f93045c6adb69457f471c74.jpg", "reference_inputs/images/043861481c93694c02127ab7d41d2e19419e9b7d4039311d46ba9a2f73117327.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000256", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000256", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [234, 270, 254, 278]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [586, 266, 590, 298]\nB: [678, 258, 690, 300]\nC: [0, 238, 18, 314]\nD: [810, 244, 827, 305]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8c4f2dd84eebf0d99565b44297c8b9cfc106d37ee7773beba7cb3296109d5644.jpg", "reference_inputs/images/6d416a27ccd8e05a56094ad0d2585cf7065b4df2698ad5b976a0c76dcc4e9bd8.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000258", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000258", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [214, 221, 244, 313]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [282, 415, 302, 426]\nB: [693, 416, 714, 426]\nC: [678, 494, 690, 514]\nD: [257, 272, 274, 281]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4b3412f131eba9c58d6c44e15d8fa0e0172a7b388e35e4ae45cd42ceac6a5614.jpg", "reference_inputs/images/83241a33f90d5353ade9d303ca3c190d92e2333f5f674957b328465ee8b42759.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000260", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000260", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [610, 241, 632, 307]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [330, 448, 343, 463]\nB: [222, 90, 247, 104]\nC: [254, 479, 272, 495]\nD: [748, 409, 766, 426]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6de9c3585cf47e11898d5f77222fa22a23cf12d738d4e6bd2252109ea1039268.jpg", "reference_inputs/images/5503db5f5ae4225d2c6f3add36d2f86fa8ee0bd3d0e0cce90ad74ba26b5a2990.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000261", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000261", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [802, 406, 821, 419]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [703, 248, 717, 270]\nB: [994, 236, 1023, 311]\nC: [167, 243, 232, 298]\nD: [518, 266, 526, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9943189216579157cb5669ed05e0a556e83eb992a1b488f4986a075136f58655.jpg", "reference_inputs/images/a48a18e7653d72be7bef71d38f63063825f2392195ab9c9f2863e8d057499145.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000262", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000262", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [118, 215, 182, 318]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [646, 201, 668, 210]\nB: [700, 0, 714, 22]\nC: [244, 407, 265, 419]\nD: [287, 459, 314, 471]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/89cff0c1c699c614047b88b3f9603a7e7326a5b29bccf7abdef2a036bb6de68b.jpg", "reference_inputs/images/1432e33aabdc6f9551f47e1894e23338857b9c07f84cbe0f83f482c8a61515cd.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000266", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000266", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [827, 230, 857, 316]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [673, 77, 694, 90]\nB: [738, 484, 762, 500]\nC: [694, 358, 707, 369]\nD: [714, 186, 735, 194]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d29cac67f917b41bf253feba461d4cd1c2a80234455b6dfce3a8745ec2c770da.jpg", "reference_inputs/images/e02de4f49edc2fa3c357bdc47dfb2209aa693b1f328f22c5d25b242bed0a2c94.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000267", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000267", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [682, 430, 696, 446]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [703, 260, 710, 294]\nB: [260, 250, 278, 286]\nC: [628, 263, 636, 294]\nD: [510, 251, 530, 312]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cd60d6b1e87c07bad19308bec877e9c74bc0ee6bfc6dd24a1d8ee46e2d07f17d.jpg", "reference_inputs/images/0babe2b74ff3bef3cdac98263fe4524186cb1fd3c2bd927891a0e7a8de627389.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000269", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000269", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [530, 493, 543, 509]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [563, 272, 606, 300]\nB: [760, 258, 772, 298]\nC: [446, 281, 474, 294]\nD: [491, 258, 518, 292]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cfa18aef7c5f1fd61ebebff7dafc64d1918ac5e9d9bb3839c48b3dd33dba961b.jpg", "reference_inputs/images/34d027dbab11397f568f524e93a1358bb23b7502da274b0c04a1dba07548b837.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000270", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000270", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [664, 130, 683, 142]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [797, 231, 848, 318]\nB: [1003, 238, 1023, 312]\nC: [264, 234, 283, 307]\nD: [334, 218, 361, 318]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/acba93faab2b2b94bd8f121833dcdc0686ddae44c5819cbb8c7ea4fc08b05082.jpg", "reference_inputs/images/0434ee60b659113cb97e064670e8b02e9955eeff7f24ccedb8f970345927aa67.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000277", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000277", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [734, 150, 1023, 390]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [458, 490, 470, 508]\nB: [508, 481, 522, 493]\nC: [522, 402, 534, 420]\nD: [536, 526, 553, 547]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8d896736401234e10331aeea3ac1fabfe73c9ad39a25d19bf9916657bd895404.jpg", "reference_inputs/images/e243169b9506053af12d4cec254d3ac2558083486bf9158c6ce0ca07d9a10508.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000278", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000278", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [358, 231, 379, 310]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [255, 494, 277, 511]\nB: [283, 146, 300, 160]\nC: [688, 370, 703, 381]\nD: [711, 161, 729, 171]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b1eb07f97100b73a7ce897ebe7f850e4938e47b304369c99274b95ab5d9e2bd3.jpg", "reference_inputs/images/ca06006a99dff4ec29af714dc0e9667c80dcf9844d6e9455e7d8a194ee5f3ae5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000280", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000280", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [786, 227, 818, 320]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [690, 436, 704, 450]\nB: [672, 515, 683, 538]\nC: [650, 498, 662, 515]\nD: [239, 139, 257, 155]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7b76f625b0e713b81c3b2fd9005caf85291cd0af2b37d0152ae222104a3abd57.jpg", "reference_inputs/images/3c008430886d13cc06b3af1ca92b44323c9fdc224aa91af5bf1a822dea3e92b5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000281", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000281", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [902, 227, 950, 318]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [719, 456, 734, 468]\nB: [674, 78, 686, 93]\nC: [275, 468, 290, 482]\nD: [686, 553, 702, 574]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ee3ada7830f7db8c377bd2a67d16fa7bc563e9196fe06ce4de9abdabb106743b.jpg", "reference_inputs/images/cec587f7f71d59f91723be411e06f634b2c38d94eef8cefcfe273dea4be0b7c4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000282", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000282", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [757, 157, 776, 166]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [384, 244, 398, 305]\nB: [388, 229, 433, 318]\nC: [213, 261, 222, 297]\nD: [21, 242, 49, 302]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0730d69540a3f535f0c5f859670dcdc7d870edca0b255a3da3faf91c422c5c2a.jpg", "reference_inputs/images/f4b724f9882422540f8c4dfcc45653e3be4fab7dffab2344dfeb96c6b92544c1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000286", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000286", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [500, 458, 533, 534]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [558, 278, 619, 299]\nB: [487, 260, 558, 316]\nC: [156, 271, 316, 318]\nD: [448, 272, 478, 295]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/94afe022a1b3abd64effa4e05540b2eb9a3a7625b0e9302f064304f48566d76b.jpg", "reference_inputs/images/237069dce3d0317d10e096791b5d633b982014b0bd077bcd2d5a32473f3250ae.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000287", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000287", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [487, 260, 558, 316]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [383, 470, 421, 544]\nB: [542, 220, 614, 291]\nC: [500, 458, 533, 534]\nD: [438, 83, 472, 157]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/237069dce3d0317d10e096791b5d633b982014b0bd077bcd2d5a32473f3250ae.jpg", "reference_inputs/images/94afe022a1b3abd64effa4e05540b2eb9a3a7625b0e9302f064304f48566d76b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000289", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000289", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [693, 401, 714, 410]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [330, 270, 337, 293]\nB: [656, 273, 664, 295]\nC: [304, 233, 329, 313]\nD: [502, 266, 528, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/01bec1e6063f2abdaa5a2a82f0cc79873e8527ba96cf4d200ccc3916f0c1bc18.jpg", "reference_inputs/images/6d6df61bab2fdecf4a4ade212800131b7757e843304464e1832f01685f9cfb59.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000290", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000290", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [583, 274, 650, 313]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [506, 0, 536, 51]\nB: [386, 266, 418, 333]\nC: [0, 389, 131, 443]\nD: [765, 514, 834, 552]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6d6df61bab2fdecf4a4ade212800131b7757e843304464e1832f01685f9cfb59.jpg", "reference_inputs/images/01bec1e6063f2abdaa5a2a82f0cc79873e8527ba96cf4d200ccc3916f0c1bc18.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000295", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000295", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [688, 263, 699, 297]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [274, 149, 297, 159]\nB: [1, 180, 33, 191]\nC: [273, 0, 296, 15]\nD: [748, 237, 770, 248]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/388b410086023ba52a82e8be7cc76f44d7c7a20c19a0c5cf882a1a665d6dfb2c.jpg", "reference_inputs/images/6e4af1e6834a23f481da3fb5481f4f060494ac2b58b1035dff9f3ad5ad3f2661.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000296", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000296", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [286, 470, 315, 482]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [162, 255, 193, 300]\nB: [120, 246, 137, 301]\nC: [265, 267, 271, 292]\nD: [771, 258, 781, 295]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1a0ea3e0c43e1f8333c7b5ac26d86b0d4127ca6debff07ec7d82fe47ce562749.jpg", "reference_inputs/images/c6777d37805bc1e28ba259a064593108c7c27c287b70aceb3538e5e222e83c53.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000298", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000298", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [214, 203, 242, 212]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [380, 242, 402, 310]\nB: [243, 251, 259, 308]\nC: [532, 243, 554, 309]\nD: [304, 254, 322, 306]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5b432ec76c85bb5a57956a33adb0353060181098b08c679dfb6e3d264a7c7ff0.jpg", "reference_inputs/images/a71e45c547e3afc57b1c6f30cf4d9fc977114d99cb3e9ae0d66203b12df4ad69.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000301", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000301", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [94, 243, 119, 310]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [711, 161, 729, 171]\nB: [255, 494, 277, 511]\nC: [688, 370, 703, 381]\nD: [283, 146, 300, 160]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d9e468d7be22237679e26dc9df1cf204994c6b2c8d455e66a21f6693dca19f07.jpg", "reference_inputs/images/ca06006a99dff4ec29af714dc0e9667c80dcf9844d6e9455e7d8a194ee5f3ae5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000304", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000304", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [678, 146, 690, 159]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [741, 246, 761, 306]\nB: [197, 218, 230, 314]\nC: [385, 243, 401, 304]\nD: [918, 232, 971, 322]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/425ae2a6e520d6325a6ef5ca44e168b8600759bb84f76528dbe67ab98b445533.jpg", "reference_inputs/images/3b84b5d82519d128b30a19c0cdb9e0532e599f2540e6a0fb617747af66d36169.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000312", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000312", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [0, 193, 336, 389]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [485, 462, 497, 478]\nB: [525, 486, 552, 500]\nC: [506, 530, 522, 552]\nD: [446, 536, 466, 555]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b84ea74379a0dc35ed4efa117f79ffd3eeacf5644d4810946141066cea215ba7.jpg", "reference_inputs/images/ca513ffc9b1c99c0f5ed3236a7524c984df0176f05fa92e5b3fe95b58184d6e6.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000320", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000320", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [515, 466, 538, 478]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [415, 262, 452, 291]\nB: [320, 279, 331, 294]\nC: [550, 267, 562, 292]\nD: [190, 254, 321, 325]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a6d4189bd6cf025b8ad0727de720ce78c370467fdcb90a9a16b44d4dfbfc1062.jpg", "reference_inputs/images/ee79ccb6790b2978c939ac1dffc6c1fb7f5d18399cc47dc5d3f82e9224410408.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_30m/000326", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_30m/all_questions.json", "question_id": "000326", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [551, 241, 1023, 399]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [473, 30, 528, 162]\nB: [382, 0, 418, 33]\nC: [418, 179, 456, 254]\nD: [557, 479, 590, 555]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7c40dda52bb9866996ccd7d539527860ee133d45134812dab45eff791b1e2824.jpg", "reference_inputs/images/52099689106298857ec9115810cece30ac6c15664bf475beb76baf80766f357e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000001", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [507, 339, 527, 382]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [134, 254, 157, 299]\nB: [435, 278, 455, 292]\nC: [483, 218, 621, 337]\nD: [364, 278, 417, 299]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fa1285aa27505af151765e72343905c87820febea37d8825c4068020ee9ec089.jpg", "reference_inputs/images/389a42f43ad667172d6855c4182d3f7f74d9b73cea54b3b04b85012cce008507.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000004", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [311, 274, 377, 304]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [507, 0, 527, 34]\nB: [434, 492, 453, 532]\nC: [434, 12, 455, 58]\nD: [470, 190, 489, 234]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/326e309493b74656ab397a5c7fe4a7d782f7c238e5d6a88d3395323169ed9146.jpg", "reference_inputs/images/1336f727d14e0a43920ba0e565b7a3a66a5ef82e0946e2742228982971657d7b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000006", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [732, 280, 798, 305]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [62, 18, 108, 42]\nB: [935, 90, 978, 110]\nC: [752, 61, 794, 82]\nD: [541, 387, 561, 430]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9668b0e9de3420dc53f6742f0eaa1d9277e1041769fad82048ac9ad8b7f0e19a.jpg", "reference_inputs/images/ad6b90836b7b0755e703ef8597dab6bea6c084cb7d7d421bf247550bc0fa0b02.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000009", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [800, 546, 841, 568]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [122, 291, 157, 310]\nB: [190, 288, 218, 305]\nC: [94, 294, 125, 304]\nD: [273, 290, 321, 312]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c7e819587b9f47c42a1fb39bfbab51a2148d4a1db54e01a307a83ea1c0de29b9.jpg", "reference_inputs/images/02ee26196f052db1fd6094389ca8a203753d4dfdd0196d5aad127f63ffb772e3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000011", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [310, 292, 335, 308]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [502, 425, 522, 469]\nB: [508, 30, 526, 72]\nC: [511, 294, 532, 337]\nD: [570, 162, 589, 203]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/606dd5fea012af2756b373ee21b6a9f4a7ff6eca37a4eb356fb07f1cb3476f3e.jpg", "reference_inputs/images/6335503211ddf6b2368041e62c88cd8c440ca0c151d187f98a89d26e416d6c09.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000013", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [0, 18, 257, 336]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [507, 191, 527, 234]\nB: [437, 381, 464, 458]\nC: [539, 535, 566, 575]\nD: [509, 334, 527, 374]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/34591fc1fa22610d2519ee5814ed778a1c518ffb2f2dfb3ab15e3acf8d937d50.jpg", "reference_inputs/images/e4ff7a204c5127ccacc2cdbba88bc5213586c8fae62904c9c188f0a3e8929b50.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000017", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [0, 269, 39, 305]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [968, 71, 1014, 92]\nB: [114, 77, 158, 100]\nC: [398, 74, 438, 93]\nD: [673, 110, 718, 130]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6b7b77f4c99dfd37df994d2154d5b52d6924e2b46395660b62e6dc3ce09556c7.jpg", "reference_inputs/images/e42a5676f10283e3c9ffd2990f7493f1bd9bb9244ef187158759ee6bb03f2c45.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000023", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [406, 279, 446, 291]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [982, 81, 1023, 115]\nB: [778, 129, 830, 162]\nC: [410, 275, 453, 303]\nD: [986, 134, 1023, 167]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/35a9a25911a9f5b520854214e6ff4b226537fa7e88bee6e19c591f7b4b62e2ac.jpg", "reference_inputs/images/4f572f335c2a8fe2dd6d0e3ee4134e5947d4f62a9a958e2eb9d9812427b49b1c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000026", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [510, 374, 526, 414]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [490, 244, 680, 366]\nB: [945, 209, 983, 318]\nC: [467, 276, 488, 292]\nD: [365, 278, 415, 299]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5d2f1767d27bd298c77346a506fc1f34a7c27bed6120a4c3a93840e6e303f7cb.jpg", "reference_inputs/images/1c86cc5951e53a469d0dfa1d49811abb2200a75b723703e0b03968751a5c81cc.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000027", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [365, 278, 415, 299]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [505, 142, 524, 186]\nB: [530, 0, 573, 42]\nC: [435, 150, 457, 194]\nD: [510, 374, 526, 414]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1c86cc5951e53a469d0dfa1d49811abb2200a75b723703e0b03968751a5c81cc.jpg", "reference_inputs/images/5d2f1767d27bd298c77346a506fc1f34a7c27bed6120a4c3a93840e6e303f7cb.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000028", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [508, 470, 526, 511]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [485, 248, 644, 355]\nB: [467, 275, 491, 293]\nC: [390, 280, 431, 297]\nD: [229, 269, 338, 313]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ededbee325399e98678b611d225826acf13d11765882ee4f176ffa431c6d154d.jpg", "reference_inputs/images/64aea7386a93b8d411c31c3c476f955f4bb9d442dea86752dfdbfd3a90eec182.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000030", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [522, 457, 552, 498]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [458, 274, 520, 303]\nB: [831, 264, 849, 294]\nC: [621, 278, 642, 293]\nD: [678, 270, 698, 289]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/328da3101eea0b3ae63ad6574b7fbbc5314cfaf5506f98c4e67fa709b3f0c2d6.jpg", "reference_inputs/images/5962876e2b4cb80abcce4a9888010a061342e63f3d85987a6738be3e4f6b7af9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000032", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [500, 263, 586, 323]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [505, 142, 524, 186]\nB: [435, 150, 457, 194]\nC: [530, 0, 573, 42]\nD: [510, 374, 526, 414]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4bef85cf9114015576227a64dbbbb0fb2abfbf1cc0ae73d509da424e5b2b3f87.jpg", "reference_inputs/images/5d2f1767d27bd298c77346a506fc1f34a7c27bed6120a4c3a93840e6e303f7cb.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000036", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000036", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [481, 200, 870, 425]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [473, 453, 493, 496]\nB: [511, 402, 526, 446]\nC: [542, 522, 564, 567]\nD: [433, 209, 450, 253]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fc6f32bd6db38d02e41ac4f11d99c28059ead881cada91e34ed65f18dd1e8cf1.jpg", "reference_inputs/images/d78eac01e2212aeff0e4dc193dad35c2de808cf46bddb2f94cf15c2e06f99339.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000047", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000047", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [518, 491, 540, 536]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [288, 260, 374, 306]\nB: [518, 282, 542, 293]\nC: [432, 282, 455, 292]\nD: [664, 263, 691, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0682be99ca5f2461fd3a5acff9e0eedb97ec14d482c0bc7b88785ff88fc09d49.jpg", "reference_inputs/images/81046b24d8b0d3e578bdfbf2026c08137a814732d76eb9b298a0b6178fe200d5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000050", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000050", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [490, 506, 518, 575]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [486, 250, 522, 294]\nB: [974, 234, 1008, 310]\nC: [594, 262, 678, 312]\nD: [425, 280, 462, 297]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1670b21b278d683f28849b8b358df44eeecf5b31b892bfeee1b68df520facffa.jpg", "reference_inputs/images/37fdf1cfa1885a50dc1e3a45f4b4196e30f9e8011ea8b00b2aa25e97696ee22d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000052", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000052", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [427, 346, 458, 422]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [522, 279, 562, 298]\nB: [754, 254, 767, 299]\nC: [32, 167, 330, 324]\nD: [497, 266, 518, 290]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0834c11265656ac724ffc1aa245ff03a0d2c5a91325a8469463638eae5fb2714.jpg", "reference_inputs/images/135d5add10ec73a8f8dccd1efdf99084b9cf438829994e55973c598984e0abbc.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000053", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000053", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [507, 365, 527, 407]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [331, 277, 393, 302]\nB: [411, 274, 441, 294]\nC: [500, 264, 544, 301]\nD: [554, 278, 579, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5279243a189dac5df5abe787abc68a73bb0ecaf5b22deee4090ef8daf12ab644.jpg", "reference_inputs/images/876e52177c54d2f1a987ee7d6b60aef790c909d21871bccc19c29ad50f3249e4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000055", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000055", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [429, 430, 450, 474]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [188, 250, 329, 316]\nB: [700, 259, 728, 298]\nC: [424, 282, 450, 293]\nD: [518, 281, 546, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0682be99ca5f2461fd3a5acff9e0eedb97ec14d482c0bc7b88785ff88fc09d49.jpg", "reference_inputs/images/948eb1874b239d46c2fb2d23b88f7be52db81266aea190904fab7655e01e84d4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000061", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000061", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [534, 254, 571, 294]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [434, 492, 453, 532]\nB: [536, 289, 562, 367]\nC: [434, 12, 455, 58]\nD: [470, 190, 489, 234]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0eb73e9b59314fe88798044774dc7b2c640adb41215b0ea735d4ac8fa5c6b73d.jpg", "reference_inputs/images/1336f727d14e0a43920ba0e565b7a3a66a5ef82e0946e2742228982971657d7b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000063", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000063", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [462, 252, 595, 329]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [534, 402, 553, 446]\nB: [430, 398, 450, 441]\nC: [499, 139, 542, 160]\nD: [734, 33, 775, 53]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a4cb032f62d08649d215a14f7f4615f0350ca5a9e04fde3f92f65683205efc94.jpg", "reference_inputs/images/93edce0f073f265e71493e76d5e9bea18cd0070de557a1dc8f9748a53742b95e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000065", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000065", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [438, 282, 461, 292]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [542, 435, 562, 478]\nB: [542, 278, 564, 322]\nC: [473, 206, 493, 249]\nD: [435, 322, 457, 366]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0babe2b74ff3bef3cdac98263fe4524186cb1fd3c2bd927891a0e7a8de627389.jpg", "reference_inputs/images/8e830185c444506afc6719687ec1c17624b01db7c88652e33518d29d0f62951b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000070", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000070", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [433, 209, 450, 253]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [498, 278, 528, 300]\nB: [337, 241, 417, 302]\nC: [425, 281, 446, 294]\nD: [818, 245, 838, 310]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d78eac01e2212aeff0e4dc193dad35c2de808cf46bddb2f94cf15c2e06f99339.jpg", "reference_inputs/images/772ae829aab097e813edc986df6337aa53a4dfb46188b630cff56caf74228b46.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000071", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000071", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [498, 278, 528, 300]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [473, 453, 493, 496]\nB: [433, 209, 450, 253]\nC: [542, 522, 564, 567]\nD: [511, 402, 526, 446]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/772ae829aab097e813edc986df6337aa53a4dfb46188b630cff56caf74228b46.jpg", "reference_inputs/images/d78eac01e2212aeff0e4dc193dad35c2de808cf46bddb2f94cf15c2e06f99339.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000075", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000075", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [475, 278, 491, 290]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [434, 274, 457, 318]\nB: [506, 267, 522, 310]\nC: [468, 41, 487, 86]\nD: [509, 497, 527, 538]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a863c0425374d378735269f9a18b0e786addc41c4cfe1ec95b61f4e2c1feb0bd.jpg", "reference_inputs/images/d694827fa58b38fa5cc8c8a7d90659d6e5ce32234bd64db1dda9bec7d2c65c90.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000077", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000077", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [450, 275, 478, 296]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [536, 412, 565, 497]\nB: [434, 137, 456, 182]\nC: [471, 314, 487, 357]\nD: [506, 116, 527, 159]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/068981882776608705076f4c4c3e998b66611be3d5507c1608daae3a21fad37c.jpg", "reference_inputs/images/414068548b23bcfefea8c1f2c2a3e487b06a62f542faebda1598cf584ee6e015.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000079", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000079", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [734, 284, 771, 298]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [935, 90, 978, 110]\nB: [752, 61, 794, 82]\nC: [541, 387, 561, 430]\nD: [62, 18, 108, 42]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/57c3f70f35fc88b642076e72a90f518e3004de43237b3e468d97c6500ead1958.jpg", "reference_inputs/images/ad6b90836b7b0755e703ef8597dab6bea6c084cb7d7d421bf247550bc0fa0b02.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000081", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000081", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [707, 273, 841, 314]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [226, 388, 269, 407]\nB: [883, 392, 928, 414]\nC: [913, 318, 962, 338]\nD: [698, 346, 742, 368]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1920d160bf8c207cc2cfe03c2f54780add45804d245793d2333db80413b5314a.jpg", "reference_inputs/images/bf4e96f4b7d94ff33c01530a446d0874327259742aea607e7efd739a863ebae4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000083", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000083", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [607, 254, 656, 302]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [949, 463, 1023, 519]\nB: [651, 422, 697, 456]\nC: [166, 142, 214, 176]\nD: [400, 304, 442, 334]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0358db4168e42159a0bdd9d9139e684abca068fd5f6eb2a32df2a1a9289045e8.jpg", "reference_inputs/images/b962f92900ad30a5894942eb6439f696bd15162afcd0132235e430965c681111.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000086", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000086", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [638, 276, 672, 296]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [430, 317, 472, 359]\nB: [933, 459, 975, 481]\nC: [751, 432, 792, 451]\nD: [64, 391, 110, 412]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/526d042da138444008692c69fea74589f8f58bda0125d9daeb37be0c85723551.jpg", "reference_inputs/images/5febcd4fd7b05b603382688ef78ce174bec40c86327d2436495ad7d63bb32a23.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000089", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000089", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [439, 279, 477, 297]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [509, 322, 528, 366]\nB: [467, 213, 489, 257]\nC: [507, 90, 527, 134]\nD: [468, 537, 488, 575]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/67d3c4dfaf5334baa46fdb18857f233075e29e32dcbad3903bb91dcb5950e373.jpg", "reference_inputs/images/620fd84c54741d059022f3cef18c57fef7e1a6f5d4530e02050417f9182e68ba.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000098", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000098", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [532, 281, 559, 294]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [883, 392, 928, 414]\nB: [698, 346, 742, 368]\nC: [913, 318, 962, 338]\nD: [226, 388, 269, 407]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/912001ea44deb2e90793bbecd6ab0407711e1c2fc6b14c92a7fff3b001b742e5.jpg", "reference_inputs/images/bf4e96f4b7d94ff33c01530a446d0874327259742aea607e7efd739a863ebae4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000100", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000100", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [654, 257, 714, 296]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [198, 267, 243, 290]\nB: [426, 0, 457, 48]\nC: [14, 179, 114, 211]\nD: [499, 266, 542, 286]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d760e180b6d0fe7d5f5740e5ff281233dff08dc94c3c056a4d088a7a0203454d.jpg", "reference_inputs/images/7d8c6dc3e4048ac215463ba16b9bf009b62fb5cb1853919ed961264defdf375f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000102", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000102", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [490, 234, 592, 324]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [474, 426, 492, 466]\nB: [507, 439, 527, 482]\nC: [542, 251, 558, 294]\nD: [435, 250, 458, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e74ac6cd045c319af1517ac285faf98c10db9e03aab35bba1c6aae874a49105a.jpg", "reference_inputs/images/978982e89ee44383f5f276568b615973f4481d2be3a247f8cec0f555617eadff.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000103", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000103", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [638, 238, 677, 255]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [699, 374, 744, 406]\nB: [229, 70, 276, 106]\nC: [484, 207, 524, 238]\nD: [961, 479, 1003, 510]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5a3d23ade29fc32a74a40c56381bcee2a1f0227641011a213150ba8ecfd7e2f5.jpg", "reference_inputs/images/2dd9431a41f6dc6b4897b1d0bae97655b6dec83528c1669cde2402d072dcb172.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000105", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000105", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [500, 268, 538, 298]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [434, 261, 456, 305]\nB: [470, 436, 489, 481]\nC: [507, 241, 527, 283]\nD: [434, 77, 454, 120]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9ad8a50cccc69d4d19587748ca75a29e617773e0bff87ecdaec8ddeb22d583f1.jpg", "reference_inputs/images/93df5d76f25eb5b5a9e6ee01d8da4dc93df5407d5b90f8ac9ed2264eaa9b2d22.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000110", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000110", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [133, 271, 209, 306]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [665, 296, 705, 313]\nB: [50, 293, 98, 315]\nC: [904, 298, 950, 317]\nD: [848, 182, 894, 205]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6e3dc1f4e62b47d16d62d043bed5a418481ebc45f4896c9dc7e3614a0781bb42.jpg", "reference_inputs/images/878e4482608f331a650f587188930bd1c430697009c1fbe1fcf0812411ccb569.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000113", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000113", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [310, 278, 330, 291]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [499, 14, 542, 38]\nB: [534, 278, 553, 322]\nC: [510, 515, 529, 557]\nD: [430, 274, 450, 317]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/79eb13a479257c25e7f5b6897a0354841cbe1086dd11e973156ea1192b073fc2.jpg", "reference_inputs/images/c53616958facca5b5ba20f1d786e8f1c00a52102f781dd5f4c71c15b720f5a4f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000114", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000114", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [536, 412, 568, 498]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [434, 260, 467, 292]\nB: [322, 277, 428, 302]\nC: [542, 246, 590, 295]\nD: [874, 259, 885, 296]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7b151ccb6720a410136b56eaa274076a3a78d2dc385f54f1531fc6f92aaf3fae.jpg", "reference_inputs/images/4a9235ef98cd0be9d0efdcdc3660811c02b4055a0d190afdc8cf695abe4f7bee.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000116", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000116", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [409, 278, 454, 298]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [493, 260, 518, 337]\nB: [456, 344, 478, 388]\nC: [436, 215, 456, 258]\nD: [539, 523, 558, 569]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bb5fb5ea9e5337cd8439330bc758a7b429296eb13e3a974b49d53f06abdc07c5.jpg", "reference_inputs/images/435fb3438a5939da084e7fb0ad694d3be840746dfe99be4dbd160aa60bc58719.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000117", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000117", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [539, 523, 558, 569]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [463, 282, 483, 291]\nB: [496, 265, 517, 290]\nC: [545, 277, 570, 294]\nD: [0, 218, 316, 446]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/435fb3438a5939da084e7fb0ad694d3be840746dfe99be4dbd160aa60bc58719.jpg", "reference_inputs/images/bcdc1e69c0a36fdd6f81b6dc138bac8948640e1f401b454a6715a6941bd5a4a5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000122", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000122", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [505, 390, 524, 434]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [437, 270, 473, 298]\nB: [0, 262, 219, 346]\nC: [472, 245, 594, 338]\nD: [592, 277, 642, 302]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/88901f1b80fcc6a22add5b0413965bc44db5260cc9fff6217e99c9def73f08ba.jpg", "reference_inputs/images/360617ebd1ddf8668083527a4aeebbdda1341224f852b2297b1aa4e0edd6dbf3.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000123", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000123", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [472, 245, 594, 338]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [435, 397, 457, 442]\nB: [468, 167, 487, 210]\nC: [505, 390, 524, 434]\nD: [530, 247, 573, 290]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/360617ebd1ddf8668083527a4aeebbdda1341224f852b2297b1aa4e0edd6dbf3.jpg", "reference_inputs/images/88901f1b80fcc6a22add5b0413965bc44db5260cc9fff6217e99c9def73f08ba.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000124", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000124", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [373, 262, 401, 291]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [499, 510, 541, 533]\nB: [518, 55, 540, 99]\nC: [987, 431, 1023, 465]\nD: [427, 217, 454, 294]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9f8bd7afecd76acd880fa4f00ef12b38a4ea1e2d76f817d9fdbce48abab400e1.jpg", "reference_inputs/images/3e8d25f175bcbdde41ce4c6af50485ea88ee7c204b7bddd88503341235b27e2a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000126", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000126", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [534, 226, 565, 302]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [334, 265, 435, 320]\nB: [447, 264, 474, 290]\nC: [536, 250, 579, 294]\nD: [507, 282, 536, 290]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9825ac2cf22175ea4b1eccdc5ca6a7fbc946d54fd5eba1875ac2cbbcef46b1e9.jpg", "reference_inputs/images/1f4e5bfc1dc07dd2fc6ab971db43666d4bee8f303a3194ecb836bf23599c8aa9.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000127", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000127", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [545, 330, 564, 370]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [530, 260, 559, 291]\nB: [674, 258, 855, 332]\nC: [15, 238, 260, 330]\nD: [440, 276, 475, 299]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/981967d761992e7129c43905aa67c08a3b00f48842f047a8760372f783c45b72.jpg", "reference_inputs/images/43b0a3fe42f2acc793e99dea4dd8389f524b8f3d47f09236481fb1cc1befeb13.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000128", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000128", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [15, 238, 260, 330]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [434, 506, 456, 550]\nB: [507, 486, 527, 530]\nC: [434, 325, 454, 367]\nD: [545, 330, 564, 370]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/43b0a3fe42f2acc793e99dea4dd8389f524b8f3d47f09236481fb1cc1befeb13.jpg", "reference_inputs/images/981967d761992e7129c43905aa67c08a3b00f48842f047a8760372f783c45b72.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000131", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000131", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [741, 245, 1023, 405]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [540, 285, 562, 329]\nB: [474, 295, 490, 334]\nC: [436, 496, 458, 540]\nD: [505, 354, 527, 398]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8156fbafe831b86bac65b3a21863b8beeddbe10ef22d7a56993d4274368b123e.jpg", "reference_inputs/images/4471a64a77ea01d2f778c46a4df407fec00e0ad3aa7a02283dc971079df7e68c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000133", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000133", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [755, 246, 1023, 354]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [473, 540, 490, 575]\nB: [508, 101, 526, 141]\nC: [438, 18, 455, 59]\nD: [438, 270, 456, 310]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/59e4aa2a08930d64de52105fa7f3349672d21f92bf3adba38145d6c228691901.jpg", "reference_inputs/images/ab6d1bf6377517890052514fce9c2ff6f0574b4f6e6da70f73f7618d98f736b2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000134", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000134", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [473, 331, 493, 373]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [310, 273, 378, 305]\nB: [570, 266, 609, 299]\nC: [407, 280, 442, 294]\nD: [472, 276, 493, 292]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1741de326d767084dd4e5fd0fc6e3e2bad8b0f6b9c81fdd632056be66516870c.jpg", "reference_inputs/images/ab4684364fe6a38c2026ce521fb39e2527721bca2271eb4e42b0311ffe907e38.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000135", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000135", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [472, 276, 493, 292]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [509, 280, 528, 323]\nB: [473, 331, 493, 373]\nC: [542, 401, 564, 445]\nD: [435, 446, 457, 490]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ab4684364fe6a38c2026ce521fb39e2527721bca2271eb4e42b0311ffe907e38.jpg", "reference_inputs/images/1741de326d767084dd4e5fd0fc6e3e2bad8b0f6b9c81fdd632056be66516870c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000139", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000139", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [558, 204, 668, 306]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [536, 472, 565, 561]\nB: [506, 206, 528, 250]\nC: [434, 239, 464, 317]\nD: [509, 3, 527, 44]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/33205fcc688c463d26db6e85757732d996e12afdebcb5e4ff4d86f47f3b63602.jpg", "reference_inputs/images/92ff47316cb29caae88bdae25e58e335e04922e24f1cacf248db032a2d249108.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000143", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000143", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [498, 278, 548, 300]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [506, 452, 528, 496]\nB: [505, 18, 524, 64]\nC: [434, 27, 457, 71]\nD: [509, 252, 527, 292]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/42ee3d96da42feea79428db2a7aaa9f59bcc8eab925e72270f906ef0ce0d1823.jpg", "reference_inputs/images/e2d3d48257038542e95080942944bed57b0dd925a0658170946298c77bad1501.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000145", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000145", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [789, 279, 839, 305]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [499, 266, 542, 286]\nB: [198, 267, 243, 290]\nC: [989, 180, 1023, 209]\nD: [733, 158, 774, 177]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5f7d374ea1f609856cc0a326dbce75f0c14822bcb32ea1b54fe31767631b1b79.jpg", "reference_inputs/images/7d8c6dc3e4048ac215463ba16b9bf009b62fb5cb1853919ed961264defdf375f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000147", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000147", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [485, 233, 753, 396]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [438, 401, 459, 445]\nB: [541, 262, 562, 304]\nC: [509, 418, 527, 458]\nD: [436, 151, 452, 195]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/22e82f3f5da3ca8e0a41507973a1054fc81a0436bd364afb7118546fe3ec6147.jpg", "reference_inputs/images/5c0c04b288523e43c62b87496107f521b57697e3c84b671f772cdf6d5a80dbc2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000152", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000152", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [161, 250, 392, 351]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [475, 304, 493, 344]\nB: [542, 128, 560, 171]\nC: [506, 317, 527, 359]\nD: [509, 458, 527, 498]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1f14b718b819a632f51cf4ea6d5486cc08d1760dcb059aa48113fb4a35adc9bd.jpg", "reference_inputs/images/d1c99ae970acc5d983e9425d72c5654176ba542ae95b32114c33ffcf0ce51a4c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/City06_ground_aerial_v2_50m/000154", "task_name": "single_target_matching", "dataset_name": "City06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/City06_ground_aerial_v2_50m/all_questions.json", "question_id": "000154", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [651, 260, 887, 352]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [468, 414, 488, 458]\nB: [509, 199, 528, 243]\nC: [466, 90, 489, 134]\nD: [530, 493, 572, 536]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c33e1aeee82405ba47f657e32649e2e4e8d1e73b08d660d0fe811f4195c3ad2e.jpg", "reference_inputs/images/5f7449aa37f81ed4c60e2289ee7e1908e2c8a444221e73b2423da6704b840273.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000005", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [504, 542, 530, 575]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [627, 292, 662, 313]\nB: [588, 290, 607, 302]\nC: [534, 294, 584, 325]\nD: [725, 308, 804, 340]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6a0a48d01bc95f1f7fcd682cf5bf04d829e1ad86c258bf09a4e682a7c6e7a042.jpg", "reference_inputs/images/419c8e8c0256f002c06ea832dc2d2986c4c4ccb63acb18dc519acc5e537e18a2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000008", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [697, 549, 703, 574]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [942, 289, 966, 348]\nB: [726, 298, 790, 313]\nC: [375, 290, 386, 327]\nD: [545, 290, 557, 335]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9f86aa4244ba0a6f10005b92de656d75d1fff6c3bde4f6e41da74851b25dbef3.jpg", "reference_inputs/images/c472c7a537652977f5214ec7087946e5be959ccce589b4a953e2b9ccd03f25d0.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000014", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [728, 315, 828, 344]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [774, 556, 816, 575]\nB: [653, 429, 674, 440]\nC: [607, 455, 630, 473]\nD: [664, 518, 698, 551]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7a0bf23ec2ab312ed51cb9da293f10c6cf147291121dfec77e840170a2766309.jpg", "reference_inputs/images/ef914f0e8dd6ab8d097e08449220a058e4bf7d703d7db775bb65dc43e21a8c4c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000016", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [450, 430, 466, 444]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [498, 295, 523, 314]\nB: [410, 298, 439, 311]\nC: [451, 289, 479, 301]\nD: [774, 314, 930, 350]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/532a7bb18fb47c2c7a6afad29eee2e22fe5703751da3632aa370b49a67654041.jpg", "reference_inputs/images/bb1b3b3955a3f2a6491e5176b2d74a97d2bd6453e48df238d57ef60e4f23c296.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000019", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [338, 292, 353, 339]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [570, 90, 590, 110]\nB: [50, 290, 98, 306]\nC: [546, 505, 556, 533]\nD: [348, 418, 367, 437]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1ae237aad839b445a89cc0e8f1fdecf1e406c23485f9ed36443a2be1a3304a8c.jpg", "reference_inputs/images/faf5c975d2ec136c30cf96efcb52dbaabe98daa7b1c8995bf160f483c0b5d4bc.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000020", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [654, 159, 670, 174]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [850, 299, 861, 328]\nB: [141, 330, 194, 461]\nC: [674, 300, 743, 510]\nD: [34, 294, 48, 321]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9d4861e27c7af67f311d1114d48865240984cf9383313830a45ab9c71394e6ae.jpg", "reference_inputs/images/b28e4de0f5b59dbe34d9977ae7924f1d6f127b56999d871fee4297beb2c3d38a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000023", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [604, 211, 619, 227]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [623, 294, 653, 310]\nB: [78, 307, 142, 437]\nC: [401, 289, 410, 329]\nD: [748, 290, 756, 317]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/849cd22b6e40580812c81b67cf65b6fc71764024aab8e0eb8b5491978a1c08d6.jpg", "reference_inputs/images/d2a8bdb779ef00c832a03a9cf278cb6e861f8194b994ccea31a3c8cb1ddb27b1.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000025", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [822, 542, 831, 568]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [655, 287, 666, 320]\nB: [125, 290, 154, 338]\nC: [632, 289, 638, 310]\nD: [164, 290, 178, 318]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d5bfe9a4746a03c92cbeee7518807d18504fafcbc4319a64d05fe881f42e9aab.jpg", "reference_inputs/images/daafd56b7f9b0584e99977588f276d6d9f561cd5540ec6e91fc46c16fc19f642.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000028", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [488, 293, 507, 336]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [658, 302, 676, 322]\nB: [556, 0, 568, 22]\nC: [494, 277, 510, 289]\nD: [593, 176, 613, 200]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f49c4f52f85920e9e5dea3a1b3ae15bcd5c7ce1ad7abaff33f57fb4d0b32b31a.jpg", "reference_inputs/images/963c522f0c5074a9d318215cc863a414a03f91c1154003c1feec44249dcf3039.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000029", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [126, 522, 134, 542]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [937, 287, 949, 317]\nB: [923, 290, 930, 306]\nC: [878, 288, 886, 313]\nD: [404, 290, 415, 311]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7b64bdab408c95a6643d57103399d9c188327aad39abda79c85b69484ea97c84.jpg", "reference_inputs/images/5f9384bac607ee37e2b8a1e132423adf08623b83de9eaf6be2f0a31ece2f4bd8.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000031", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000031", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [574, 510, 589, 539]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [399, 262, 451, 314]\nB: [598, 290, 617, 355]\nC: [562, 294, 596, 338]\nD: [550, 287, 557, 314]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/98588555a9656d7cfc9f63266d2fcf2b6eb3064482278c14ddac2c450e16d2c2.jpg", "reference_inputs/images/c32b20010a9f85d7f75228c727ce166db866aa2ad664200bc62120323d137ad8.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000032", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [508, 286, 516, 317]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [574, 19, 589, 49]\nB: [574, 510, 589, 539]\nC: [507, 122, 518, 142]\nD: [13, 177, 26, 192]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c32b20010a9f85d7f75228c727ce166db866aa2ad664200bc62120323d137ad8.jpg", "reference_inputs/images/98588555a9656d7cfc9f63266d2fcf2b6eb3064482278c14ddac2c450e16d2c2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000033", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000033", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [555, 544, 568, 575]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [542, 288, 552, 319]\nB: [646, 288, 670, 321]\nC: [803, 297, 856, 364]\nD: [599, 292, 627, 330]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d7d943a4a54c062ace1a68a7b6609a9ceff742dd0079c3bd2182309ec298a983.jpg", "reference_inputs/images/209bd4735a24f658dec24cceca9171d144dbbabcfe4542d281051a67651e5bf0.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000039", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000039", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [366, 371, 383, 392]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [959, 295, 990, 350]\nB: [726, 289, 741, 331]\nC: [220, 294, 232, 322]\nD: [840, 290, 857, 340]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/04ae5b63297ef9faa39e0c51d2a61dcbc1fd82e72a57a0281f9362ae3f704f66.jpg", "reference_inputs/images/e6a930cc13c9435d48975c505cbff4ed5d6ffbc32bde3d87755e13d702c854ea.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000040", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000040", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [840, 290, 857, 340]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [903, 131, 948, 147]\nB: [289, 421, 310, 436]\nC: [186, 478, 218, 502]\nD: [366, 371, 383, 392]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e6a930cc13c9435d48975c505cbff4ed5d6ffbc32bde3d87755e13d702c854ea.jpg", "reference_inputs/images/04ae5b63297ef9faa39e0c51d2a61dcbc1fd82e72a57a0281f9362ae3f704f66.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000052", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000052", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [589, 307, 614, 319]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [124, 178, 155, 195]\nB: [171, 350, 284, 419]\nC: [971, 132, 1023, 190]\nD: [617, 378, 818, 506]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2de10aa040bcdf1ae947d053dc6f1971086bd5f033b6425794a494edb23eb40d.jpg", "reference_inputs/images/01886ce0440f5aca0e712fc42566fbbda12378686622992519e2b014b4fe3e5b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000054", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000054", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [938, 300, 985, 310]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [794, 387, 820, 397]\nB: [956, 406, 986, 415]\nC: [674, 406, 704, 414]\nD: [802, 405, 833, 414]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/356e298d4b7742b6b7c826b061b76cb8db7acd128030789aca33d2c107501fd6.jpg", "reference_inputs/images/46ca6a46d8e89f5f6d2288e2ff737e8f34222d083d61b4f925c0d2b9633da384.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000056", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000056", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [652, 388, 676, 406]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [319, 287, 326, 310]\nB: [102, 290, 122, 340]\nC: [597, 288, 611, 316]\nD: [829, 291, 860, 395]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1a01ed403e1d284e43fa05e564f7484c77989d75c0705a6d46e99f6f062bb7ce.jpg", "reference_inputs/images/4c030ac1994474c88f320c0a101f7b988a4282a18b3f7adb0be9b0a96bdc2f49.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000061", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000061", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [313, 293, 328, 353]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [894, 552, 927, 575]\nB: [900, 302, 946, 318]\nC: [939, 353, 990, 371]\nD: [360, 353, 385, 371]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/56fdec6518e6fd442d5bf50715cf842df5618b6e7125f80e58d99b5f95fb2643.jpg", "reference_inputs/images/5e4c2c0e430bd207efd500ecdc990d997bcaf0e0d27b3be7700bf9e63ded022e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000064", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000064", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [729, 507, 785, 526]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [581, 317, 653, 350]\nB: [560, 298, 586, 316]\nC: [47, 331, 253, 395]\nD: [406, 294, 432, 312]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/685f494fe9e8e65491aa89b5ade5db7f11ed63b16af3cd19c09d9acd7a59edcf.jpg", "reference_inputs/images/17677f743d8a6da8619151c529c0aeabda048dcabd2d71f1a689c00c52dca099.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000065", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000065", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [698, 340, 715, 354]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [840, 292, 852, 329]\nB: [0, 306, 86, 421]\nC: [249, 318, 343, 533]\nD: [990, 322, 1023, 418]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/eff8b82e21d3dfa9e0ad7b019d80bf17195a495a1a6241280a43cb597eba984e.jpg", "reference_inputs/images/0cff59f867fe7b86097ed51822007c1913efb6a96246b1483f058af263a31611.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000067", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000067", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [624, 470, 650, 491]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [677, 329, 868, 378]\nB: [448, 290, 458, 327]\nC: [47, 306, 357, 373]\nD: [863, 333, 923, 468]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c24189713c2885b33cdb7ec22d3ce8b0f26e20865d2a16e92c64822b1c8d6c4f.jpg", "reference_inputs/images/ae20cf585532b43c40ad46f616b7781893b01fe8238e42691c424f3f746caf4a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000068", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000068", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [677, 329, 868, 378]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [624, 470, 650, 491]\nB: [606, 434, 623, 448]\nC: [635, 417, 652, 429]\nD: [581, 389, 603, 416]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ae20cf585532b43c40ad46f616b7781893b01fe8238e42691c424f3f746caf4a.jpg", "reference_inputs/images/c24189713c2885b33cdb7ec22d3ce8b0f26e20865d2a16e92c64822b1c8d6c4f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000070", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000070", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [702, 266, 809, 327]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [684, 562, 713, 575]\nB: [471, 473, 503, 488]\nC: [563, 362, 576, 380]\nD: [720, 505, 759, 533]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b458fc1c79ce19d141010b566ef27d6d7d4f034b0f01d15dfc3ad0ca65fb4341.jpg", "reference_inputs/images/dc12abe33bc14f80ec2fcb9ff08c59844092337b36d418833dcffb48ac003407.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000071", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000071", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [438, 552, 444, 575]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [714, 294, 729, 350]\nB: [271, 290, 293, 374]\nC: [930, 287, 941, 314]\nD: [677, 290, 691, 342]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5decbba8daa7c5df5a7fff7c056b20766744bd23a17ecce6a98368f20a0ea9b9.jpg", "reference_inputs/images/b85a6b29f399b960fda0517b94c5f99319b190ff4a324961395b1f408498e61b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000073", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000073", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [314, 294, 334, 361]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [406, 459, 421, 481]\nB: [943, 94, 992, 122]\nC: [359, 94, 381, 122]\nD: [900, 534, 945, 559]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/09ba14ce8ded16be3daec4a82b919fc8ebbb4f1d8f69197aa587c4f85118470d.jpg", "reference_inputs/images/da04c191254855336a79e0a3f941e0ca2737508a8c214c4857aecc57b54a285c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000074", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000074", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [646, 422, 666, 432]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [807, 302, 866, 318]\nB: [625, 295, 650, 306]\nC: [582, 278, 602, 298]\nD: [392, 308, 493, 334]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9b77a6fdcbcd164e42d030afdba126226416d464d2bb4b234b73eec459bc4446.jpg", "reference_inputs/images/9454bfbf127a6ef586c8cbd21a88516a04c24bf9b4a4d2d70da2bdcf1d16e104.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000076", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000076", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [842, 422, 877, 433]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [2, 330, 157, 390]\nB: [267, 306, 327, 326]\nC: [0, 312, 23, 332]\nD: [385, 295, 410, 310]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8502f61f181f5d74610587b6ad6b6b5a8d2308a26aaf20c07336a3dfbc0b0ca5.jpg", "reference_inputs/images/efa5658689482f7351dfb2d3e76efc3166d2ac78bdd9cedd302f88a92851f48f.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000079", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000079", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [574, 362, 590, 375]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [878, 300, 897, 344]\nB: [636, 309, 662, 379]\nC: [590, 297, 614, 387]\nD: [793, 288, 807, 326]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1a8c244f9fd25465901629aefad10819dec40735eac73ba20e3c6519f9c110c0.jpg", "reference_inputs/images/99996be75264845d962cdba0a1ed1039ac5656471bd459aca071179e1ee05317.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000082", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000082", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [847, 306, 895, 326]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [530, 486, 571, 503]\nB: [408, 504, 456, 530]\nC: [537, 515, 588, 542]\nD: [383, 531, 440, 562]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/dce70d33d8f9602df8bca856a63b807f674d48cc5394e78f6422eb83339193aa.jpg", "reference_inputs/images/45161aa0655e931e7924983dd2f4d431231d378aad482002b7e8acb1c3c8ed8a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000090", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000090", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [434, 453, 450, 471]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [786, 290, 794, 318]\nB: [682, 295, 743, 327]\nC: [596, 320, 724, 402]\nD: [527, 295, 553, 314]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6e1496264eca10566efafa7606ae73487c5a8ed30b46085a4373d1994c6920e3.jpg", "reference_inputs/images/8a88af8aed0a862d09c57f4777a709d8920949e88250c84d2a84d9db1501ea1e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000091", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000091", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [829, 418, 889, 528]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [711, 297, 764, 321]\nB: [626, 296, 651, 307]\nC: [677, 294, 702, 303]\nD: [871, 300, 964, 326]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/46a9fb7ac10b3829f0e7f5ac450541aa9df5834fbab04c531ebea628c01549d1.jpg", "reference_inputs/images/90ded380178ee48995f1ae42f5000217afa8d28b717d39d22a2dea889bf2c026.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000093", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000093", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [563, 315, 604, 330]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [366, 414, 454, 512]\nB: [518, 311, 594, 404]\nC: [790, 162, 878, 270]\nD: [947, 58, 1023, 161]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/12612aa50c4da7e27c053652a3c4e25219e5ca62e5385dfabd28088fd37f4ec1.jpg", "reference_inputs/images/c1fe12231e3c6505f4561f7b59d589c5fc02488b59c32ec3725a2de26e313288.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000094", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000094", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [694, 450, 717, 468]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [649, 291, 669, 321]\nB: [1015, 290, 1023, 320]\nC: [915, 287, 926, 313]\nD: [567, 289, 582, 311]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f3f337c634a9cc8fb5161273331a95970b7f1312fe185c12bb1a6b550e321af6.jpg", "reference_inputs/images/cb10762cc79fb403ed3d060779fd177fdbc256d05c9dc19842a408604ffd0524.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000098", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000098", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [633, 526, 687, 552]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [273, 311, 351, 340]\nB: [421, 292, 442, 307]\nC: [653, 334, 856, 394]\nD: [549, 307, 592, 327]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6944ab270ef1de72c2ac192cacafe3ecee78bdff3b2179cbfc2ee7d0ffa87c47.jpg", "reference_inputs/images/65b7e6f323bd65c5d7669cb2b5b8e28d3141f449e05c138b7d4ca072cfd32073.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000103", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000103", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [367, 232, 383, 254]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [191, 288, 199, 319]\nB: [129, 287, 138, 322]\nC: [70, 290, 86, 342]\nD: [86, 294, 103, 347]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8a31b191f8c250bb6360bdd96a5e983847859947aa023305338aab9890ad4e4d.jpg", "reference_inputs/images/7aa8d2593ef92fb73eee2aadb88fb71465e0238902da12118e7801436cd57d36.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000104", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000104", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [86, 294, 103, 347]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [367, 232, 383, 254]\nB: [343, 190, 361, 212]\nC: [929, 175, 972, 190]\nD: [342, 476, 360, 506]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7aa8d2593ef92fb73eee2aadb88fb71465e0238902da12118e7801436cd57d36.jpg", "reference_inputs/images/8a31b191f8c250bb6360bdd96a5e983847859947aa023305338aab9890ad4e4d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000107", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000107", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [599, 534, 610, 556]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [925, 319, 1023, 464]\nB: [810, 306, 944, 575]\nC: [146, 288, 157, 325]\nD: [222, 308, 290, 440]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4b34f4cb71a2d71d19dd3baacb4efbffba03c9ad7a50932e6b3432ed3969757b.jpg", "reference_inputs/images/5c973ec6b63e4c1d2c149201ffa5cd3b1388a2ea9967a81b72eefd098be5952a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000111", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000111", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [151, 440, 192, 461]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [329, 290, 373, 306]\nB: [87, 289, 110, 331]\nC: [550, 287, 558, 314]\nD: [322, 295, 367, 309]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/62b21a2b843cd73f959a114c90ba7721502bba41770522608b3cf2779d5083c2.jpg", "reference_inputs/images/3a78ffcc0ed50e3d24ee4672dfc9222e7e836147cac5174430c65a014fb89efe.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000113", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000113", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [589, 406, 602, 424]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [0, 289, 10, 308]\nB: [564, 292, 571, 326]\nC: [553, 286, 558, 310]\nD: [518, 293, 530, 321]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b3179d4995386d7424b127065c3906520a6979f7d9e651da211736cdda95402e.jpg", "reference_inputs/images/2c4bc66211ccfe5085183b35111b64fa11a0bb4469dc650241a0294db4576f68.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000114", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000114", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [553, 286, 558, 310]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [541, 127, 564, 143]\nB: [829, 195, 862, 212]\nC: [589, 406, 602, 424]\nD: [575, 106, 590, 122]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2c4bc66211ccfe5085183b35111b64fa11a0bb4469dc650241a0294db4576f68.jpg", "reference_inputs/images/b3179d4995386d7424b127065c3906520a6979f7d9e651da211736cdda95402e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000115", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000115", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [931, 305, 966, 393]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [750, 346, 776, 359]\nB: [658, 34, 677, 68]\nC: [234, 110, 268, 150]\nD: [493, 9, 512, 33]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1f31fc5314bbe96377c8d32893358852fb5df5811d258ede43731ab443bd5a7c.jpg", "reference_inputs/images/27e2e55c2e7e50496595cc24dc88040224fdb2d575abf40b1320de508d4b939d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000116", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000116", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [711, 530, 766, 557]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [819, 292, 875, 310]\nB: [1006, 291, 1023, 307]\nC: [896, 293, 951, 303]\nD: [363, 294, 378, 345]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f459bf256a51aa8a9babcda85af13e41bdfe56751b2ca2d2cae1fc8e8bcc5e15.jpg", "reference_inputs/images/5343fb3594b5630345448d98043996f02515530e2eddc6c3b7565939086db183.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000117", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000117", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [819, 292, 875, 310]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [716, 497, 760, 517]\nB: [711, 530, 766, 557]\nC: [887, 494, 938, 522]\nD: [919, 538, 977, 567]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5343fb3594b5630345448d98043996f02515530e2eddc6c3b7565939086db183.jpg", "reference_inputs/images/f459bf256a51aa8a9babcda85af13e41bdfe56751b2ca2d2cae1fc8e8bcc5e15.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000118", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000118", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [410, 512, 416, 534]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [492, 306, 538, 334]\nB: [0, 313, 123, 575]\nC: [431, 298, 442, 332]\nD: [478, 295, 502, 345]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/affd2dcdaf8eb527c0164b784fcca03599307d983a8ea5f904659e5ebdc351f7.jpg", "reference_inputs/images/9abc00b4e40f897765fbf70bfcb30be772200c61adedab2028d69cc7c3131385.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000123", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000123", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [781, 534, 806, 560]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [306, 291, 311, 314]\nB: [172, 294, 180, 318]\nC: [192, 291, 198, 313]\nD: [340, 294, 349, 322]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/96282862df9fed1e54bdafbd1cefb080486775ae3b35c0d1cf0a80741f19ab67.jpg", "reference_inputs/images/6c19efe1630f6eca1709b8bf989300cf0ebc534fdb36b20b60aeb358ef605f59.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000124", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000124", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [493, 290, 498, 311]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [871, 98, 908, 118]\nB: [658, 364, 678, 379]\nC: [575, 0, 589, 22]\nD: [781, 534, 806, 560]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6c19efe1630f6eca1709b8bf989300cf0ebc534fdb36b20b60aeb358ef605f59.jpg", "reference_inputs/images/96282862df9fed1e54bdafbd1cefb080486775ae3b35c0d1cf0a80741f19ab67.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000125", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000125", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [448, 528, 454, 552]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [816, 287, 825, 302]\nB: [922, 289, 933, 310]\nC: [945, 288, 964, 321]\nD: [488, 297, 506, 344]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6a53cb19284675c6c5354bc6c6bd250a20bbdccc3923d22bef0751c01f751a7e.jpg", "reference_inputs/images/d077bbdaf768375de248571253a710a7fa0160cf0bd9752e6a23c3f637e67556.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000129", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000129", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [711, 550, 718, 575]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [251, 292, 258, 321]\nB: [166, 294, 174, 323]\nC: [70, 292, 78, 314]\nD: [22, 295, 31, 319]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6a6016d2229b73349092e87b4bb01534da951024528805cce457af4c643e780b.jpg", "reference_inputs/images/fdb606cadb3ed9dec989cac8afde68eac34da1d710796fc2e02eac839c823c29.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000132", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000132", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [860, 529, 866, 553]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [739, 292, 750, 329]\nB: [668, 309, 708, 419]\nC: [424, 290, 433, 318]\nD: [550, 293, 560, 324]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/49d6755483780222f4b2c746fdedcb4f0a357cfbfc7fee28d788d62ac63643b6.jpg", "reference_inputs/images/2bec413c323574edc50994e6be33fd04e6dd8462cd85f8440d3597eba943747a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000134", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000134", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [769, 289, 875, 368]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [759, 394, 1023, 552]\nB: [665, 377, 790, 423]\nC: [157, 335, 383, 410]\nD: [0, 360, 202, 465]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0671ae8ee169440e6627d92e84b2ba53644dd471a3dc9097fb5bbae57d101527.jpg", "reference_inputs/images/dc50adabea5c473de95da4c3e3bf5ddb25880f2dd43b3eb3b1e235dc6a75a193.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000135", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000135", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [157, 335, 383, 410]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [470, 373, 577, 454]\nB: [652, 297, 760, 382]\nC: [961, 496, 1023, 575]\nD: [769, 289, 875, 368]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/dc50adabea5c473de95da4c3e3bf5ddb25880f2dd43b3eb3b1e235dc6a75a193.jpg", "reference_inputs/images/0671ae8ee169440e6627d92e84b2ba53644dd471a3dc9097fb5bbae57d101527.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000136", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000136", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [568, 430, 586, 449]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [759, 293, 771, 329]\nB: [358, 296, 367, 325]\nC: [790, 308, 926, 575]\nD: [581, 294, 590, 325]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/267d540ee9b0ff0e106444fec0a4e8f20e611add97d7fec0044f1ab39f2b2569.jpg", "reference_inputs/images/30260f4557385cfaffdbff2134fbf3283819655eb5e46a4ff757dbb097a74244.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000137", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000137", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [590, 440, 607, 454]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [620, 302, 658, 323]\nB: [618, 293, 649, 302]\nC: [579, 297, 599, 308]\nD: [719, 294, 771, 319]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3fdf6aa1683d61f1361aa6904621ad2986d520365cb8109ac469f6c2d8fc6fad.jpg", "reference_inputs/images/c343894f686c9a6f9b4d32f34a3a2c8aabb3398fdb1bac343c44670e40d08d32.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000138", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000138", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [410, 519, 436, 553]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [526, 294, 548, 310]\nB: [572, 310, 646, 361]\nC: [947, 328, 1023, 390]\nD: [657, 294, 705, 320]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/eecc5fd6027c2498737cfd1fb08d36e7c6fe4d6927d160f242cd9a1edbb2dc0e.jpg", "reference_inputs/images/0270e6c49c8feef3c5b6875c7c99e75229547dceb165bee3d8a659b59ba4fe54.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000142", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000142", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [624, 297, 649, 346]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [241, 0, 262, 9]\nB: [265, 38, 295, 65]\nC: [936, 476, 976, 498]\nD: [575, 436, 589, 453]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1ad130f5de0e004fa42498c161a617cf14ec4a8d227e64fae1c07a2d2dd489a1.jpg", "reference_inputs/images/f19ea19b97516dfba6b6e801db04f01cf4b2e53346fce64b51c49c508a7e6444.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000144", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000144", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [969, 90, 1016, 114]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [593, 314, 619, 328]\nB: [221, 270, 254, 340]\nC: [421, 283, 427, 306]\nD: [732, 325, 739, 348]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2f5b9c5803ca4db366c311011221b5a28998dfc282cf3612aa0248beaeaaa7a9.jpg", "reference_inputs/images/2b0085602fd1cf33b6065a19d446355c48f9ee3e442de955cf7b10c5227f648e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000146", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000146", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [861, 530, 870, 552]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [877, 302, 891, 333]\nB: [299, 306, 311, 336]\nC: [949, 309, 959, 333]\nD: [0, 304, 31, 374]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/87d5cebc84b97e0be052460d2ca2b4ada12d3b9a1a0a465e6c4d5e9d259bbbbc.jpg", "reference_inputs/images/a6f0dc1acff8a9f1bda04ce3966e51b2cf5fe48703546cf8bfefc2aeb8f55b32.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000150", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000150", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [639, 418, 658, 430]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [666, 297, 689, 306]\nB: [427, 296, 458, 304]\nC: [514, 301, 541, 314]\nD: [578, 293, 611, 305]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4eb701386ccf91f87d8153ba6786d69689da1ca369180d1c62781fc6f8b641d2.jpg", "reference_inputs/images/d0eeb6b43019302b751b886be04ae065f56c73947ae476b6c614edaf1be9df61.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000151", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000151", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [578, 293, 611, 305]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [707, 486, 742, 506]\nB: [608, 438, 625, 452]\nC: [654, 539, 694, 575]\nD: [639, 418, 658, 430]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d0eeb6b43019302b751b886be04ae065f56c73947ae476b6c614edaf1be9df61.jpg", "reference_inputs/images/4eb701386ccf91f87d8153ba6786d69689da1ca369180d1c62781fc6f8b641d2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000153", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000153", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [666, 502, 682, 528]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [635, 290, 644, 324]\nB: [619, 288, 628, 321]\nC: [384, 288, 394, 332]\nD: [978, 288, 994, 321]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5fbcfe7065c52069fe8aa341d04ba4bc2dd6861a51449d3a1074f836c20015a1.jpg", "reference_inputs/images/5c67be9d3816bdac4d19867169d2ac85fa3ee42dc173288321467a7e9199fd18.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000154", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000154", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [544, 530, 550, 554]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [585, 290, 593, 310]\nB: [905, 288, 914, 309]\nC: [674, 290, 682, 318]\nD: [605, 292, 614, 312]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/724372a2fd2a537ed5a192a41177e90751b7e096a5fa9ca0de83f754ba02271f.jpg", "reference_inputs/images/1de7c9c6d5738af396b2ad9c14fc44a7e9c4bfe5311cc72c29fdbe77828a209b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000155", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000155", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [179, 535, 212, 566]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [0, 293, 17, 391]\nB: [446, 287, 451, 311]\nC: [199, 295, 238, 414]\nD: [381, 292, 392, 330]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/849cd22b6e40580812c81b67cf65b6fc71764024aab8e0eb8b5491978a1c08d6.jpg", "reference_inputs/images/fa7c54f3b909238f998f2c0cedd69494a8ae4325bbd507da601efa649d4e55b6.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000156", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000156", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [532, 472, 573, 492]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [679, 303, 732, 322]\nB: [594, 294, 634, 307]\nC: [507, 295, 546, 312]\nD: [633, 299, 665, 314]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0051393cd63d511e7658334ca57f89ab965af6108304cd2d8b82d0ed36c77437.jpg", "reference_inputs/images/6dd7fca2b6d17dd9c99effce28ce8aa2c7b8f274bbff06780d240f09f540dca2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000157", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000157", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [633, 299, 665, 314]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [527, 536, 578, 558]\nB: [409, 484, 454, 505]\nC: [399, 518, 450, 546]\nD: [532, 472, 573, 492]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6dd7fca2b6d17dd9c99effce28ce8aa2c7b8f274bbff06780d240f09f540dca2.jpg", "reference_inputs/images/0051393cd63d511e7658334ca57f89ab965af6108304cd2d8b82d0ed36c77437.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000159", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000159", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [393, 384, 415, 401]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [814, 290, 822, 317]\nB: [803, 287, 810, 314]\nC: [728, 293, 748, 337]\nD: [747, 289, 757, 326]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fb55ae6fe0a5e87f2a974b92788564c804ef2b673d2f3f9e41ec8f55f4f06477.jpg", "reference_inputs/images/8acb44882ad014a1644004c081aaf072d078f577d9277171e12f3dedf544739b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000160", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000160", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [541, 127, 564, 143]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [327, 286, 332, 305]\nB: [566, 287, 574, 318]\nC: [688, 286, 694, 311]\nD: [530, 289, 540, 309]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b3179d4995386d7424b127065c3906520a6979f7d9e651da211736cdda95402e.jpg", "reference_inputs/images/2b2d6b22125d98413e964783960ccf933791e9a38d0a3efda03636806edc9200.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000161", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000161", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [688, 286, 694, 311]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [829, 195, 862, 212]\nB: [541, 127, 564, 143]\nC: [575, 106, 590, 122]\nD: [589, 406, 602, 424]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2b2d6b22125d98413e964783960ccf933791e9a38d0a3efda03636806edc9200.jpg", "reference_inputs/images/b3179d4995386d7424b127065c3906520a6979f7d9e651da211736cdda95402e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000163", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000163", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [658, 195, 676, 215]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [13, 290, 35, 322]\nB: [243, 298, 289, 454]\nC: [166, 289, 178, 322]\nD: [449, 306, 480, 448]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3989f30fc712460e1b532e65a85eb8f559614bc7322ad4501dd3aceb0bfe3ef9.jpg", "reference_inputs/images/eb8d33c08eeb88818243035813fa10741bbd04cdaa0b5a3a0d96f183c4e4a9a5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000164", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000164", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [710, 520, 769, 542]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [624, 298, 665, 310]\nB: [796, 302, 868, 321]\nC: [118, 304, 286, 384]\nD: [546, 303, 587, 322]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/dabcddb2c5f31308ba63bf2401e624fa4873357724a5c22c4e346a03c97f202d.jpg", "reference_inputs/images/c871a2b6bab048eb4995aff931bf7060b4d846a70080dfeef714fc030b01d5d7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000165", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000165", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [624, 298, 665, 310]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [510, 532, 566, 558]\nB: [710, 520, 769, 542]\nC: [600, 470, 630, 485]\nD: [910, 464, 959, 483]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c871a2b6bab048eb4995aff931bf7060b4d846a70080dfeef714fc030b01d5d7.jpg", "reference_inputs/images/dabcddb2c5f31308ba63bf2401e624fa4873357724a5c22c4e346a03c97f202d.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000168", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000168", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [362, 516, 366, 538]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [306, 289, 311, 310]\nB: [254, 287, 262, 311]\nC: [179, 286, 187, 313]\nD: [598, 288, 605, 320]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/07ca824b99b314bd94d57eda78a95cebe85ce30bc8e7d8698d800d83e704a7c3.jpg", "reference_inputs/images/1075720d899aa88c5aa14cb7eeb20344e150199624b9bde5c85be648db3d26cb.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000170", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000170", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [155, 302, 186, 314]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [469, 292, 478, 325]\nB: [862, 286, 869, 303]\nC: [510, 290, 522, 318]\nD: [706, 291, 734, 299]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c97a9f8f7de6a094a1f743d885b8fb4e23676ff22a3149bfbf2818ff7fc8494b.jpg", "reference_inputs/images/1634592f4f73b92a835b205ab6e70cd9d32472082667dd9479c0f2f0c1eb1a71.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000171", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000171", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [584, 302, 641, 334]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [762, 393, 790, 414]\nB: [151, 436, 183, 457]\nC: [560, 78, 622, 189]\nD: [314, 361, 370, 470]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1634592f4f73b92a835b205ab6e70cd9d32472082667dd9479c0f2f0c1eb1a71.jpg", "reference_inputs/images/c97a9f8f7de6a094a1f743d885b8fb4e23676ff22a3149bfbf2818ff7fc8494b.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000172", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000172", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [130, 158, 165, 177]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [770, 290, 782, 320]\nB: [375, 290, 386, 325]\nC: [440, 287, 447, 314]\nD: [900, 286, 908, 308]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/41821988d9f59431e844993ef642182ec744c51695610e657e3aa61b90f80b96.jpg", "reference_inputs/images/5b5e53f763930b5c8ada4a79985b9f6114cb0b759d93796acd02cb2dc9a44557.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000173", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000173", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [770, 290, 782, 320]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [645, 227, 662, 259]\nB: [963, 306, 1006, 334]\nC: [130, 158, 165, 177]\nD: [118, 394, 155, 410]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5b5e53f763930b5c8ada4a79985b9f6114cb0b759d93796acd02cb2dc9a44557.jpg", "reference_inputs/images/41821988d9f59431e844993ef642182ec744c51695610e657e3aa61b90f80b96.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000181", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000181", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [371, 435, 393, 446]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [536, 307, 593, 324]\nB: [805, 308, 887, 325]\nC: [461, 334, 557, 391]\nD: [980, 298, 1023, 315]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d76465ef7fbd5baaf0fe865c9d44135d8dba217611268bbefb6586e6993147b2.jpg", "reference_inputs/images/c52d3ed15c8ad4699b4bc938831c5b06416414b65cdf6d09847efe920d0a1925.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000182", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000182", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [805, 308, 887, 325]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [624, 535, 671, 575]\nB: [510, 510, 534, 534]\nC: [621, 518, 659, 546]\nD: [439, 561, 487, 575]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c52d3ed15c8ad4699b4bc938831c5b06416414b65cdf6d09847efe920d0a1925.jpg", "reference_inputs/images/d76465ef7fbd5baaf0fe865c9d44135d8dba217611268bbefb6586e6993147b2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000183", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000183", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [649, 530, 682, 550]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [656, 293, 690, 335]\nB: [561, 287, 572, 327]\nC: [626, 290, 646, 319]\nD: [342, 293, 356, 342]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/05b6f20e4d9f1346db76bff920938e567e5ec7a0b5c20003a050d61e4e631ef7.jpg", "reference_inputs/images/153c42acc503b7f49f9717f56313e3f98e87cf0cd70405ca91151993eafe3310.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000184", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000184", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [561, 287, 572, 327]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [670, 298, 700, 311]\nB: [654, 116, 670, 134]\nC: [649, 530, 682, 550]\nD: [570, 431, 584, 451]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/153c42acc503b7f49f9717f56313e3f98e87cf0cd70405ca91151993eafe3310.jpg", "reference_inputs/images/05b6f20e4d9f1346db76bff920938e567e5ec7a0b5c20003a050d61e4e631ef7.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000188", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000188", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [830, 310, 897, 418]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [838, 86, 880, 110]\nB: [707, 266, 737, 279]\nC: [978, 513, 995, 520]\nD: [577, 276, 590, 288]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a460f7b7dbf8f43b48e8b5b279694f465bb12e5a397e8bfacd1f528cbfdcaa7e.jpg", "reference_inputs/images/38b14bfd106225a2e22f062250eb7d51bddb58c2ae67d4a7ab3f274540ea8a91.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000189", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000189", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [506, 446, 518, 468]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [507, 288, 516, 321]\nB: [674, 292, 698, 316]\nC: [726, 287, 738, 316]\nD: [565, 295, 576, 326]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/671c05560af838f8e7ce8cfb9c707b838fff0c85439a5c7e4b8e942e3e784aab.jpg", "reference_inputs/images/0e21b79909dbcad1e395e42f1bc899edc1a61134bc233babed9b8aeb536d73de.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000191", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000191", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [325, 287, 332, 310]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [575, 550, 590, 574]\nB: [618, 353, 634, 369]\nC: [435, 19, 456, 43]\nD: [242, 550, 271, 575]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f00be5f5431d5f216408980ab71683bd8405197facf8cec2f1e73396ba223dd5.jpg", "reference_inputs/images/40d0ce2fd4a56089ea8837e202fe35b54cc6d6a9d77031cfb4521ab135148a32.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000197", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000197", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [816, 104, 874, 203]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [898, 317, 1023, 366]\nB: [736, 298, 779, 310]\nC: [886, 296, 993, 327]\nD: [430, 302, 498, 318]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/eff8b82e21d3dfa9e0ad7b019d80bf17195a495a1a6241280a43cb597eba984e.jpg", "reference_inputs/images/d7773a87bfc65223ef136cddb83885c4a9ac2237e41809e937b5e43cb65c17c2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000198", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000198", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [608, 295, 647, 402]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [556, 304, 568, 319]\nB: [698, 340, 715, 354]\nC: [614, 471, 634, 490]\nD: [816, 104, 874, 203]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d7773a87bfc65223ef136cddb83885c4a9ac2237e41809e937b5e43cb65c17c2.jpg", "reference_inputs/images/eff8b82e21d3dfa9e0ad7b019d80bf17195a495a1a6241280a43cb597eba984e.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000200", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000200", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [598, 283, 621, 294]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [263, 294, 271, 312]\nB: [546, 292, 553, 318]\nC: [316, 294, 335, 358]\nD: [337, 290, 342, 313]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ad6fe422531a93b9485098c6c3e4bc860f8aeb03907c9f838ade835d139050c2.jpg", "reference_inputs/images/d2fdc4683ecf48ad9dd5cf6d1739017f93cacc73846c598a25716dd0416a30a8.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000203", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000203", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [884, 298, 923, 422]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [902, 37, 948, 63]\nB: [359, 94, 381, 122]\nC: [900, 534, 945, 559]\nD: [943, 94, 992, 122]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8034946f2608b55f1c1ec0013fdd86421442ba325dab5eb7aadbf96d5fd6d01c.jpg", "reference_inputs/images/da04c191254855336a79e0a3f941e0ca2737508a8c214c4857aecc57b54a285c.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000204", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000204", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [816, 130, 849, 149]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [133, 295, 141, 321]\nB: [159, 340, 183, 354]\nC: [774, 288, 783, 306]\nD: [723, 289, 730, 317]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/63315177cd2bff6a7058920f8a87748f28d0c6357e8e5a3199f35f040d74e698.jpg", "reference_inputs/images/3b12d533cb9b3083bc67fe032a90a315ea41b6adfe9c796d5a7e8b24805ca529.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000205", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000205", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [723, 289, 730, 317]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [559, 146, 586, 162]\nB: [779, 22, 808, 47]\nC: [816, 130, 849, 149]\nD: [859, 65, 896, 90]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3b12d533cb9b3083bc67fe032a90a315ea41b6adfe9c796d5a7e8b24805ca529.jpg", "reference_inputs/images/63315177cd2bff6a7058920f8a87748f28d0c6357e8e5a3199f35f040d74e698.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000209", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000209", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [279, 321, 436, 402]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [382, 126, 482, 225]\nB: [739, 0, 842, 76]\nC: [731, 166, 840, 269]\nD: [373, 0, 436, 35]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9d7dd89769e249a4eddd5467312146be0428ac4007260c0ea7d8f0137e5d7c62.jpg", "reference_inputs/images/1aedfd6862af453cdd58f1227e627eccd860b652ccd30b1767d50432e1f9439a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000211", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000211", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [554, 248, 568, 267]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [529, 306, 542, 361]\nB: [438, 317, 502, 515]\nC: [370, 271, 405, 293]\nD: [570, 290, 587, 348]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/82de66c33aa7a0590d806c77253770c7eab69ef7bfc0111dd4211431220f31f4.jpg", "reference_inputs/images/58b047aeb9b58dd0ab84674467ce22bda3cef16ba28ca4600778e7ba9afda82a.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000212", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000212", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [570, 290, 587, 348]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [554, 248, 568, 267]\nB: [491, 519, 512, 543]\nC: [265, 376, 283, 424]\nD: [526, 198, 536, 214]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/58b047aeb9b58dd0ab84674467ce22bda3cef16ba28ca4600778e7ba9afda82a.jpg", "reference_inputs/images/82de66c33aa7a0590d806c77253770c7eab69ef7bfc0111dd4211431220f31f4.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000213", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000213", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the AERIAL image and contains the query target.\nImage 2 is the GROUND image and contains the answer options.\n\nYou are Ground Agent, a ground-level agent observing the current scene from the ground view. Your aerial teammate has marked one target in the aerial view. Among the candidate targets in your own ground view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, local context, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (AERIAL): [630, 65, 646, 85]\n\nCandidate bboxes in Image 2 (GROUND):\nA: [548, 287, 555, 314]\nB: [812, 290, 823, 322]\nC: [620, 291, 631, 325]\nD: [749, 288, 760, 316]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c2217bb8e6aa009d3e3ca16374eb01c1883d2cc9c0bda56a8911c621037229e5.jpg", "reference_inputs/images/3d09ae4f750be8d742c79345cce31f91ef6007fac5bdd8f3d3c9caa054a123e2.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "single_target_matching/Town1_ground_aerial_v2/000218", "task_name": "single_target_matching", "dataset_name": "Town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/single_target_matching/Town1_ground_aerial_v2/all_questions.json", "question_id": "000218", "input_profile": "mcq_text_reference_v1", "prompt": "You will receive exactly two images from the same scene.\nImage 1 is the GROUND image and contains the query target.\nImage 2 is the AERIAL image and contains the answer options.\n\nYou are Aerial Agent, an aerial agent observing the current scene from the aerial view. Your ground teammate has marked one target in the ground view. Among the candidate targets in your own aerial view, select the one that refers to the same real-world target. Use cross-view cues such as relative position, surrounding structures, and target scale. Answer with only one option letter from A, B, C, D.\n\nBounding box rule:\n- Every bbox is in pixel-space XYXY format: [x1, y1, x2, y2].\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- [x1, y1] is the top-left corner of the box and [x2, y2] is the bottom-right corner.\n- Do not interpret the bbox as [x, y, width, height].\n\nQuery bbox in Image 1 (GROUND): [87, 299, 119, 378]\n\nCandidate bboxes in Image 2 (AERIAL):\nA: [781, 285, 806, 297]\nB: [858, 327, 894, 342]\nC: [815, 384, 847, 402]\nD: [658, 102, 678, 123]\n\nTask: choose the option in Image 2 that refers to the same real-world target as the query bbox in Image 1.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/367c209537f201822469726901355e9a9ac7b55b7871a02425f63afc00f44a20.jpg", "reference_inputs/images/c2217bb8e6aa009d3e3ca16374eb01c1883d2cc9c0bda56a8911c621037229e5.jpg"], "image_roles": ["source", "target"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_1/000001", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_1/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[390, 565], [398, 551], [406, 538], [413, 527], [418, 518]]\nB: [[357, 545], [365, 531], [373, 518], [379, 507], [385, 498]]\nC: [[364, 494], [358, 504], [352, 515], [344, 528], [336, 542]]\nD: [[306, 514], [343, 518], [377, 521], [403, 522], [426, 522]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/87681dd633f9bb7c74377ef7419019babece4306220a9288eb271cde960a7cdd.jpg", "reference_inputs/images/15245ba020cf6e783786ba757a33144baa2f9e8a74784a9974c8ad66089bb5ab.jpg", "reference_inputs/images/fb3c272b6dcf05cb557f791109d3f56dcff5fd38303dc5ace091da081bd90864.jpg", "reference_inputs/images/c5287c672a4f19e673363a8840ad91fcad4d7a5db443e566a639985ed4060112.jpg", "reference_inputs/images/71c1313d128d1ea725d216fb6734997a6b1d05aa4a59e923213ec925d890c1fe.jpg", "reference_inputs/images/2fe9db9c6a9419399f517a973fc6c6b5f82b5f8b422ff13af6334ae8df96745b.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_1/000002", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_1/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[366, 490], [361, 500], [354, 512], [347, 523], [340, 535]]\nB: [[361, 538], [368, 526], [375, 515], [382, 503], [387, 494]]\nC: [[310, 509], [346, 514], [379, 518], [406, 518], [428, 518]]\nD: [[394, 558], [402, 546], [409, 535], [415, 523], [421, 514]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/ad98916baafdcc0fffc690c04b0ae904e0c5ef41dbf322f267b2778f35d83cc7.jpg", "reference_inputs/images/de92f91f21fab86c2268edae8a488040bd0c31e710bc486b5341dbe22247a6b3.jpg", "reference_inputs/images/d12b964d3171aa974eb1d56025796191d9ed956b113fcecc9a44e09743ebfd2a.jpg", "reference_inputs/images/892f0eff86d7f7928015f85cc2dfcc646b3f7613faaf4137359938b627a951f3.jpg", "reference_inputs/images/830176b4180c4e34f35cc0a954bfc008a924b26ce328db90ff1125c3e775be13.jpg", "reference_inputs/images/0c0cfe050616909b424d3bc7da5259762732c1240421bb66415ae4b6fc7bc3e4.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_1/000003", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_1/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[406, 467], [440, 474], [470, 481], [495, 485], [516, 489]]\nB: [[490, 516], [495, 506], [500, 498], [505, 490], [509, 484]]\nC: [[457, 497], [462, 487], [466, 478], [471, 471], [475, 465]]\nD: [[454, 462], [450, 469], [446, 476], [441, 485], [436, 494]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/0bbf7809c194a64c36d9c8209f5f432b2747178f97a751700084415f8efc439e.jpg", "reference_inputs/images/01b4ca2f6870357b74b449e5b9d5185663a418a38b9f5a01ec5fa1f03c4abcf2.jpg", "reference_inputs/images/55ae4922c0eef6907f6ba8b5be744aa6bdc6710f261568135b4f25b42b45c008.jpg", "reference_inputs/images/f02410e0b1620b29227c3c7da6c00468d01671f5d951d3d62f82aa37a488328e.jpg", "reference_inputs/images/153eec0dbf2198a6389fac8783d9e0e78b99207804ff82741827f419638696c5.jpg", "reference_inputs/images/fb2ebc2e42b4fc69218f6b6f2b1eacdf45343abc4508af9340bf85b2461d95af.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_1/000004", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_1/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[301, 524], [339, 525], [372, 528], [399, 529], [422, 527]]\nB: [[352, 554], [361, 538], [368, 526], [375, 514], [382, 503]]\nC: [[318, 534], [327, 518], [334, 506], [342, 494], [348, 483]]\nD: [[361, 500], [354, 511], [347, 522], [340, 534], [331, 550]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/fcccfe923ee77df64bf0171d7cff12a75bca4f66b64d44b753d219b35e6ae039.jpg", "reference_inputs/images/6fc408b89d4a72bdeb270299db55ec0cc8a2d4eb16796ebcb33f368f4e6e6421.jpg", "reference_inputs/images/2f939dc6505586f57e67dd87b187494c7da981789a2763f241e26de7433efc95.jpg", "reference_inputs/images/f38d697f25180b8eead5545e9e951866ece635150d78351f188e02d527c7d4a6.jpg", "reference_inputs/images/37c621352526df543296e0de3beebe88604a0834a1b0b055b22dfbc9722c04f0.jpg", "reference_inputs/images/de92f91f21fab86c2268edae8a488040bd0c31e710bc486b5341dbe22247a6b3.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_1/000005", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_1/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[322, 488], [358, 495], [390, 499], [415, 502], [437, 505]]\nB: [[375, 477], [370, 485], [365, 494], [358, 505], [352, 515]]\nC: [[373, 518], [379, 508], [386, 497], [391, 488], [396, 480]]\nD: [[406, 538], [413, 528], [419, 517], [425, 508], [430, 500]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/c164b11a7a3c6533de7a6bcbeae60a9fb17915456ecdf5583066a2e1a5c5ebe6.jpg", "reference_inputs/images/25fa80aebb3e5ffa452b564944ac7af80c6a03a8ed1008bf2acfa606e3b216ba.jpg", "reference_inputs/images/20aa5be5de2561a490effa8bb071cb04043cdb2e90ae16d63391a00046ece418.jpg", "reference_inputs/images/8fb3812c58eee375dfde983f158e6393081dfdd0bdb17113206c2318ee8d5ad0.jpg", "reference_inputs/images/b58da46c7062dcee8c5c806cfc0576cb89a91b91168e849bf9afe7c5874e84e0.jpg", "reference_inputs/images/aba202adeae7c6672221ad49b5972bd2a47c37f022f40f0c1e8ad7efdb2ac55c.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_1/000006", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_1/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[366, 490], [361, 500], [354, 511], [347, 523], [340, 535]]\nB: [[361, 538], [368, 526], [375, 514], [382, 503], [387, 494]]\nC: [[327, 518], [334, 506], [342, 494], [348, 483], [354, 474]]\nD: [[394, 558], [402, 546], [409, 534], [415, 523], [421, 514]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/b19fd2d3e825354caa8c56a72bd83debd473402ec6090c7b169fb24fd8a999d1.jpg", "reference_inputs/images/f8a32a7b915f83a31e36352bb6259e45a48b5b4ac274fcdc95adbfa7be10d362.jpg", "reference_inputs/images/62feff420308996aa2246cdd6510a453031cd933add3cdb879e34f3d916fe3b9.jpg", "reference_inputs/images/0ff4adbe5d5ddb15816d3df560ae53d15f73cb08d9e3be8e4cbce312f9d090b1.jpg", "reference_inputs/images/4aa3914a6d5b995bb55d548a9d3c47d642d5b1131cdab5bf6ab6ad67ad64fdfb.jpg", "reference_inputs/images/86dbc598b881dc32dcbe96520dc80f5dbb02d8fa76da5a131f28300a54a5364f.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_1/000007", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_1/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[735, 134], [736, 140], [736, 98], [736, 66], [736, 31]]\nB: [[750, 40], [750, 74], [750, 107], [750, 149], [749, 143]]\nC: [[710, 40], [710, 74], [710, 107], [710, 149], [710, 143]]\nD: [[770, 41], [735, 74], [706, 107], [682, 149], [662, 143]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/3a7dd4bd7906840a62281421e357511f1b43f908726ec05b15286e2c99a09c0d.jpg", "reference_inputs/images/997586bd5040813fca98f320b2eb2d58904fdb6b7e7392d409b8fc3846f5e332.jpg", "reference_inputs/images/1cd30d863d46171e92d05945a25bb641fc747638be6fc628e09e6e30a1c9429b.jpg", "reference_inputs/images/57f62dc37e256e2288424960f82517358cedbedee0e1155684d0587f0f82cadd.jpg", "reference_inputs/images/fb141bf27b2ffd5b2e7873f94eebef684a990c8a972b981a86e43f39151386c6.jpg", "reference_inputs/images/03fa5e653ce78cd7e963ce681a4f6d27c98409b56097f634392cb3d6d1c5c584.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_1/000008", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_1/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[301, 524], [338, 526], [372, 529], [399, 528], [422, 527]]\nB: [[318, 534], [326, 519], [334, 506], [342, 494], [348, 483]]\nC: [[352, 554], [360, 539], [368, 526], [375, 514], [382, 503]]\nD: [[361, 500], [354, 510], [347, 523], [339, 536], [331, 550]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/b031a412d43c86a28cdc5a627432a1ad1990159b58df6b691b960a709ac755c9.jpg", "reference_inputs/images/4aa3914a6d5b995bb55d548a9d3c47d642d5b1131cdab5bf6ab6ad67ad64fdfb.jpg", "reference_inputs/images/86dbc598b881dc32dcbe96520dc80f5dbb02d8fa76da5a131f28300a54a5364f.jpg", "reference_inputs/images/bf404a778f138d8d04841fa3c11a5b4b5872ba69cc209c23244dcb74b0382560.jpg", "reference_inputs/images/962c9948a521049211be59352a9ab1816c6be2373097a7603cf13b4ef954a49d.jpg", "reference_inputs/images/6cde4c4f500966c707d61c84eb121e47ca29bb06d40cf7c382f8202a174a4b1f.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_1/000009", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_1/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[366, 490], [361, 500], [354, 510], [347, 523], [339, 536]]\nB: [[310, 509], [346, 514], [379, 516], [406, 518], [428, 518]]\nC: [[360, 539], [368, 526], [375, 514], [382, 503], [387, 494]]\nD: [[394, 559], [402, 546], [409, 534], [415, 523], [421, 514]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/2e85344c5feb3e4f663f61aed805594189d82177066b4f06289245c683ec2de3.jpg", "reference_inputs/images/bc14056ebbc165b711ae61a062ef2ba36e2114e063d945811f160f09bd143764.jpg", "reference_inputs/images/740b499effc4d51b0717f2813297c909b7c36e4d1dbbc3e605d0af9584a01660.jpg", "reference_inputs/images/a86a682aad11b8b3d4bd4478bb02889e988d5e8ee19d6acb0eceef4358a02c22.jpg", "reference_inputs/images/168d4981c74a6db104984a47213cb3f3a1a21258ed10eccb5c7c09766c5d3766.jpg", "reference_inputs/images/282fc906447f57ae0b4931bd129aab93ddc02563e1632c142bea9bc0baa31276.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_1/000010", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_1/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[742, 170], [742, 135], [742, 101], [742, 66], [742, 25]]\nB: [[710, 36], [710, 77], [710, 112], [710, 146], [710, 182]]\nC: [[758, 36], [757, 77], [757, 112], [757, 146], [757, 182]]\nD: [[663, 36], [662, 77], [662, 112], [662, 146], [662, 182]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/84790ae9735bef985f04e260abb7486d8b9c8491fb45fa15d35ecebad9672388.jpg", "reference_inputs/images/3d11aa6597f4294ecb543378970cbe7e1e1bb21a6db39e8e19b036fbc0d84e04.jpg", "reference_inputs/images/8701077606a2e68714992a9a364bd9e5b892ea65cdb2471e612c1ea21fdeb40e.jpg", "reference_inputs/images/8d49508d7daff6171fe14a6530da92602847bb2d2de6924f86f33e17aa24bd70.jpg", "reference_inputs/images/19cee4b454b837112cd9806f739d614de9e3e08c349aaf52882fabcc6f3e1db6.jpg", "reference_inputs/images/bc955cda9dca33db227da5e967f8607d309c64116b697e1ba9617f8133487dc9.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_1/000011", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_1/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[297, 530], [335, 533], [370, 533], [397, 533], [420, 532]]\nB: [[348, 561], [357, 546], [366, 530], [373, 518], [379, 508]]\nC: [[358, 505], [352, 515], [345, 527], [336, 542], [327, 558]]\nD: [[314, 541], [323, 526], [332, 510], [339, 498], [346, 488]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/c2a4726407c7989c1beb44359964dc567fd90edc1cd589c7db30fcc2d0bd6d65.jpg", "reference_inputs/images/25fa80aebb3e5ffa452b564944ac7af80c6a03a8ed1008bf2acfa606e3b216ba.jpg", "reference_inputs/images/20aa5be5de2561a490effa8bb071cb04043cdb2e90ae16d63391a00046ece418.jpg", "reference_inputs/images/8fb3812c58eee375dfde983f158e6393081dfdd0bdb17113206c2318ee8d5ad0.jpg", "reference_inputs/images/b58da46c7062dcee8c5c806cfc0576cb89a91b91168e849bf9afe7c5874e84e0.jpg", "reference_inputs/images/aba202adeae7c6672221ad49b5972bd2a47c37f022f40f0c1e8ad7efdb2ac55c.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_1/000012", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_1/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[421, 514], [426, 504], [431, 497], [435, 489], [440, 482]]\nB: [[354, 474], [359, 464], [364, 457], [368, 449], [373, 442]]\nC: [[387, 494], [393, 484], [398, 477], [402, 469], [406, 462]]\nD: [[386, 458], [381, 466], [377, 474], [372, 481], [366, 490]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/3750081222c7f82f8adf2fac3d243fd63827b9ef68e13334e569b229a97bc9bd.jpg", "reference_inputs/images/3ec0b8903aab93d648dc0f9c40a89d1463d40870e58825268a5039ceb4f9aac0.jpg", "reference_inputs/images/f8a32a7b915f83a31e36352bb6259e45a48b5b4ac274fcdc95adbfa7be10d362.jpg", "reference_inputs/images/62feff420308996aa2246cdd6510a453031cd933add3cdb879e34f3d916fe3b9.jpg", "reference_inputs/images/0ff4adbe5d5ddb15816d3df560ae53d15f73cb08d9e3be8e4cbce312f9d090b1.jpg", "reference_inputs/images/4aa3914a6d5b995bb55d548a9d3c47d642d5b1131cdab5bf6ab6ad67ad64fdfb.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_1/000013", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_1", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_1/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[742, 165], [742, 130], [742, 94], [743, 54], [743, 21]]\nB: [[711, 32], [711, 66], [710, 106], [710, 141], [710, 176]]\nC: [[665, 32], [665, 66], [664, 106], [664, 141], [664, 176]]\nD: [[758, 32], [758, 66], [757, 106], [757, 141], [757, 176]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/753b370abdc0efb1753db5ef029f446b9a79aefe5ddfe162d14e36479aced9a3.jpg", "reference_inputs/images/153eec0dbf2198a6389fac8783d9e0e78b99207804ff82741827f419638696c5.jpg", "reference_inputs/images/fb2ebc2e42b4fc69218f6b6f2b1eacdf45343abc4508af9340bf85b2461d95af.jpg", "reference_inputs/images/dd1eaf0bf612bcceceb210b8629ba96ceefa1486f0ebf55e66b94c2232646867.jpg", "reference_inputs/images/15a34577b3b52975518d0f19d73c3dbb9f8783605954e885220fb2c7274c8eca.jpg", "reference_inputs/images/f0c90bd95ea9a7a6c68a9e2d832ec35b1e32ba613ed94d35251dea430eec4560.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_2/000001", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[706, 506], [697, 494], [690, 483], [683, 473], [678, 464]]\nB: [[673, 528], [664, 515], [658, 505], [650, 494], [645, 486]]\nC: [[633, 503], [638, 512], [646, 522], [652, 533], [661, 546]]\nD: [[640, 550], [631, 537], [625, 526], [618, 516], [612, 507]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/8f26c76812a03036e894a34b4c58bbb37e9a63d25b5dccb6765c2c3fb9c4777f.jpg", "reference_inputs/images/ae6fe67732fff9ce8ec79bb0b80ec97acaffa2242f481078c7268be8bb8f4304.jpg", "reference_inputs/images/5c34941cfc7d630ed23e686d89f38d5317d535e31c2fb1c4877443d8f52eb113.jpg", "reference_inputs/images/6f8a1f306811996184346820c3d7a3bfafeb40cf3fa85cd000303e175112cc1c.jpg", "reference_inputs/images/5d950d23f92320bd380509b56d0068e1a0896fa6ddcff81075a3cf14cd8a8d0e.jpg", "reference_inputs/images/19066739318fcb491e6695b84b686138365a9a06461226f8defdd18fa3819eb4.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_2/000002", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[362, 61], [322, 95], [286, 134], [258, 168], [236, 206]]\nB: [[291, 62], [292, 95], [293, 134], [293, 168], [293, 206]]\nC: [[324, 194], [324, 156], [324, 122], [323, 83], [322, 50]]\nD: [[245, 62], [246, 96], [246, 134], [246, 169], [246, 206]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/2f4a3b581154554501f73e63b71fc70e80eede969d6a52bae0082be9feb113bf.jpg", "reference_inputs/images/999a34cf41fca92b5c02263ba1a62f3f9fae9586b8b350b4b138ab38f6ab458e.jpg", "reference_inputs/images/a59fe69eeccd1666a1b45c2f6b22038744071d2b684a89532f538e027bc66098.jpg", "reference_inputs/images/2d5b7510a3c39b8549155d1a422e79c5265f6e1766cb8d92409754a830caf950.jpg", "reference_inputs/images/47525823eeb4bf3886891e4b6eb94fada136dfc456a62c4d36de1902bacad226.jpg", "reference_inputs/images/625f30c67cc5a65b9ce210cee37e92c6327d76b24d621e2839d24754d92d20cb.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_2/000003", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[640, 549], [632, 538], [626, 527], [618, 517], [613, 508]]\nB: [[704, 506], [696, 494], [690, 484], [682, 474], [677, 465]]\nC: [[634, 504], [639, 513], [646, 523], [653, 534], [661, 545]]\nD: [[672, 527], [664, 516], [658, 506], [650, 495], [645, 486]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/9e0e371c40bb60914adbf20f9022644f8e82d6238aac8ad49e23984596df28ca.jpg", "reference_inputs/images/b18e1a5de34cf4e8435099dc7e97cdcb0821c9cb029b92e386961f42b35497ce.jpg", "reference_inputs/images/ca9a28917c96f7989c9192f4ac37d8eb1a5758da5ea09ea6f7bacefefb710105.jpg", "reference_inputs/images/7662db216052b047ccb8333c7bcc859fb7f5f436ff77f5231e349b1e3b820a40.jpg", "reference_inputs/images/6738f585633a59d9ccb7deb24c1f029e8cc82404e55d4be4607de1a73d1a85f8.jpg", "reference_inputs/images/89d7ae751303361d0c62543b148a65cb71eb639f2a12997d6c40ec79e3c754fa.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_2/000004", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[645, 487], [639, 478], [634, 470], [630, 463], [626, 457]]\nB: [[678, 466], [672, 457], [667, 450], [662, 442], [658, 436]]\nC: [[614, 474], [618, 480], [622, 487], [627, 494], [633, 504]]\nD: [[612, 508], [606, 498], [602, 491], [597, 484], [593, 478]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/45c6cab9e8728e4c9ed67e04ed8545520566ffdf23a217292fda38db1a85dbd6.jpg", "reference_inputs/images/f107fcdac3aa78cdb5eb81a827ad1517eee94ef79da4b8c373cd56a08d643f7e.jpg", "reference_inputs/images/6694797b6fb55bfaa4995bcbc4334c190d108892a547fde5afbe41521b520145.jpg", "reference_inputs/images/89b5be56db834eb9867e29b2036f0360bd0ecdcf15007e4c140f9c0e95516d89.jpg", "reference_inputs/images/338a1d771be7ee53ed0b65de66a71d473f4b0fe962f6d0524af5f084c88f0a36.jpg", "reference_inputs/images/737a63e72b7999630bf10b35af60689d7353c9db5fae8f693be7d15e75c5de44.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_2/000005", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[628, 495], [634, 504], [640, 513], [646, 522], [654, 534]]\nB: [[633, 538], [625, 526], [619, 517], [613, 508], [607, 499]]\nC: [[665, 516], [657, 505], [651, 495], [645, 486], [639, 478]]\nD: [[697, 494], [689, 483], [683, 474], [677, 465], [671, 456]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/87681dd633f9bb7c74377ef7419019babece4306220a9288eb271cde960a7cdd.jpg", "reference_inputs/images/d47a72488abc95f1d7574c77b6bdd1523b997b57a822816e1117f7e1d7003e89.jpg", "reference_inputs/images/7a70711dbc9397cce83bc33ab2574ba47789d652b001bd87f911400f74c31c46.jpg", "reference_inputs/images/f3385cc064ac2a92841da18f6e4d553192d311999c145d5604d326b2156dc33e.jpg", "reference_inputs/images/af49f2b7a71bfa09f55b0cdf0cd0288e63717c1fa2b86bb56dd17e237769ed41.jpg", "reference_inputs/images/79e5166eae395c1340e840097d09b3ef0ffae457ddeda1fa7e1e6fac787d201b.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_2/000006", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[664, 516], [658, 506], [650, 494], [645, 486], [639, 478]]\nB: [[631, 538], [625, 527], [618, 516], [612, 507], [606, 500]]\nC: [[697, 494], [690, 484], [683, 473], [678, 464], [672, 457]]\nD: [[627, 496], [633, 503], [638, 512], [646, 523], [652, 534]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/5ffe3f2d001d5cda4aa15d86a64b18fe6a57097fc0c159cea021f8859aade56a.jpg", "reference_inputs/images/85992d417dcada88ced81c65edcae69aa08102c450f710cef665b798d9f60d66.jpg", "reference_inputs/images/b647e56436eb022aab081d645d5866b00c6f07f13b706154af5e5871394c93b2.jpg", "reference_inputs/images/91f7066a660c931f97560f7a133936483e6006c74a4e013511b2747595152e33.jpg", "reference_inputs/images/fc20c684aa9d299743af5a630956c52ffa0aa74a409326a73a1bcdc76d858fa6.jpg", "reference_inputs/images/ddf024b546b7da45a87238c107894ad378056fa21596d1192862189b8cfb18a6.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_2/000007", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[639, 512], [646, 522], [653, 534], [661, 546], [669, 558]]\nB: [[648, 562], [640, 550], [632, 538], [625, 526], [618, 516]]\nC: [[681, 541], [673, 529], [665, 517], [658, 505], [651, 495]]\nD: [[714, 520], [706, 508], [698, 496], [690, 484], [684, 474]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/eef85de63d6edd963fd4fa502f471b163f82770c570bb07cc0870755cccdbe00.jpg", "reference_inputs/images/338a1d771be7ee53ed0b65de66a71d473f4b0fe962f6d0524af5f084c88f0a36.jpg", "reference_inputs/images/737a63e72b7999630bf10b35af60689d7353c9db5fae8f693be7d15e75c5de44.jpg", "reference_inputs/images/92b27f202cd13a39f8a9570c7e15e1aa33dacb0bdd70d9ebe4b09ae60d827168.jpg", "reference_inputs/images/6d9b5c8ef7e6f08ce8982702ca8f7533cc3dde75eaf9d3f84de4eac6a5786099.jpg", "reference_inputs/images/33e24c1c5fc14fb06153f587f216f328b3c8031e9ffeaad6fbb98f6e538955a0.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_2/000008", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[671, 456], [666, 448], [662, 442], [658, 435], [653, 429]]\nB: [[607, 499], [602, 491], [598, 485], [594, 478], [589, 472]]\nC: [[639, 478], [634, 470], [630, 463], [626, 457], [621, 450]]\nD: [[610, 468], [614, 474], [618, 481], [623, 487], [628, 495]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/b031a412d43c86a28cdc5a627432a1ad1990159b58df6b691b960a709ac755c9.jpg", "reference_inputs/images/9778738fb48db310e36235e307c82afb8cabe32983d36b7a0299736b2223c85b.jpg", "reference_inputs/images/9e4dd2c9721072abceadb6abd5ba6151a78f4738d49bd83a1cb9b143e4c3312f.jpg", "reference_inputs/images/02b16874d688f9691ac9276a2d1582c6f2224439d9be44c80e8282188ae12f6a.jpg", "reference_inputs/images/66a2e0cb2ff06d5dd08573b101c79ff741d97df076ac339909e79ffded7ea409.jpg", "reference_inputs/images/c1b3d5bc87f66bf2721a41e112cd6ee6e453f02e2d84414a6d86d2da7361672d.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_2/000009", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[245, 50], [245, 88], [245, 124], [245, 158], [245, 197]]\nB: [[325, 185], [325, 146], [325, 112], [325, 76], [325, 38]]\nC: [[341, 50], [341, 88], [341, 124], [341, 158], [341, 197]]\nD: [[293, 50], [293, 88], [293, 124], [293, 158], [293, 197]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/d0b658257fd1a29d040c9ad2b37cbda64ce850dcde52f6eb9728f6ae172e81d9.jpg", "reference_inputs/images/159d2db4f6b062bf78ff231667a090701cff41e81b781d61010fae0cc489b4d9.jpg", "reference_inputs/images/5f4228a2ff3ff0a75c8cc1c75000391739e4ceb684b52d72ad7f76d190f54670.jpg", "reference_inputs/images/24969c6de011e5f251b9c18d4043e0fd0a7f756d5ab96c7eacedcb582cf6d5f2.jpg", "reference_inputs/images/d220950f5a1dd23bd141ae1433ba0cbeede6d05183ea84e253ffcbb97e96d270.jpg", "reference_inputs/images/62ce59d72ead34f6304e5d216dee6d21dc5525c629a5a40f7f2ba696194aa4cc.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v1_2/000010", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v1_2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v1_2/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[246, 98], [246, 132], [246, 170], [246, 206], [246, 242]]\nB: [[364, 98], [322, 132], [286, 170], [258, 206], [235, 242]]\nC: [[325, 231], [325, 195], [325, 158], [325, 121], [325, 86]]\nD: [[293, 98], [293, 132], [293, 170], [293, 206], [293, 242]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/5ac6ddb283b9d716ffe2a3926a5d6362477c60be411a14d0bcc7d6df77a8a8d4.jpg", "reference_inputs/images/2d5b7510a3c39b8549155d1a422e79c5265f6e1766cb8d92409754a830caf950.jpg", "reference_inputs/images/47525823eeb4bf3886891e4b6eb94fada136dfc456a62c4d36de1902bacad226.jpg", "reference_inputs/images/625f30c67cc5a65b9ce210cee37e92c6327d76b24d621e2839d24754d92d20cb.jpg", "reference_inputs/images/cdd6e19718969940988af389132d68172bd4095cf74127899187f98529f48930.jpg", "reference_inputs/images/909ec630a28528cea464a7ca7b35b640e9f964dc4c1b278f86fa5215070ba167.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_20m/000001", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[551, 441], [551, 429], [551, 419], [551, 411], [551, 403]]\nB: [[493, 411], [493, 419], [493, 427], [493, 437], [493, 449]]\nC: [[512, 441], [512, 429], [512, 419], [512, 411], [512, 403]]\nD: [[453, 441], [486, 429], [516, 419], [540, 411], [559, 403]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/1c4a984fa871acdb7a4fcedb40fb895e0a1f4c3d65635ac5eb19c51fb3d8efe8.jpg", "reference_inputs/images/b800d7a40eea13b4f7e2a8f5023f5f5668249c4876513c6be9737791b96918ee.jpg", "reference_inputs/images/4361054b280cd9d73394edc6e371afc818b989aed9f9327cd8d6243c79133116.jpg", "reference_inputs/images/16a8e2767ddc2533679e06b8f287131d7eb896b588ca165bca382bb4df8e8d3c.jpg", "reference_inputs/images/564ec729fefa0968379ec0be485704b80be1d9bed12050c65437d7dbb5776981.jpg", "reference_inputs/images/52b26cca703c4d0c1888005e2d5b9176065956ebf1c03e25c0ac98e6513164a7.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_20m/000002", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[453, 549], [486, 518], [516, 498], [540, 474], [559, 458]]\nB: [[512, 549], [512, 518], [512, 498], [512, 474], [512, 458]]\nC: [[492, 466], [492, 482], [492, 506], [492, 526], [492, 557]]\nD: [[473, 549], [473, 518], [473, 498], [473, 474], [473, 458]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/bfbff7adbebc42ae595d39371347b63219639483ac4c74142b2bc61623dabbd5.jpg", "reference_inputs/images/f4a7098cbe94ae907eafc0f5997b8a099436c210eb839ecea4f7bac29d27b2a5.jpg", "reference_inputs/images/bcdc1e69c0a36fdd6f81b6dc138bac8948640e1f401b454a6715a6941bd5a4a5.jpg", "reference_inputs/images/c0d2a3e08cf989cab00b16bae02eabdb9d196be80eb7ba851f154ec5d4a68f11.jpg", "reference_inputs/images/ed2189f53b3f6d1c05bceb524c0005a7508a1ff0431f3713d530ae623c5b2642.jpg", "reference_inputs/images/34d027dbab11397f568f524e93a1358bb23b7502da274b0c04a1dba07548b837.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_20m/000003", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[614, 486], [656, 487], [699, 489], [744, 490], [788, 491]]\nB: [[772, 483], [730, 459], [686, 434], [644, 410], [602, 386]]\nC: [[775, 452], [731, 451], [686, 450], [643, 448], [601, 446]]\nD: [[774, 509], [730, 508], [685, 506], [642, 505], [599, 503]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/35d64d5db014b46a9cedcdabda23114867ba82e68fff96797de30425b2b1446f.jpg", "reference_inputs/images/67edc9f39ba4f2605ef24a73488e1d0298525c61eabf99e0c8b5b8008875bf37.jpg", "reference_inputs/images/006e4b97e40fb9fb2f043974ca480bf6e4fbbcfac6f82f579c3ab98b45c6c1f4.jpg", "reference_inputs/images/136f851f4ce6fede76cc50fdc3fc6048d01d9a5ec12763732777c13fbb606be9.jpg", "reference_inputs/images/fd1a741907a15ac959be34c50c789abef4ee72af9a0929d747d349ffbe6f6efb.jpg", "reference_inputs/images/0712923cb9ef7bc88dbc73ca5969dd9ffecc54a415e0f4d2e3a28e04b7f8a7a7.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_20m/000004", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[453, 544], [486, 514], [516, 491], [540, 471], [559, 455]]\nB: [[493, 463], [493, 479], [493, 499], [493, 522], [493, 552]]\nC: [[512, 544], [512, 514], [512, 491], [512, 471], [512, 455]]\nD: [[473, 544], [473, 514], [473, 491], [473, 471], [473, 455]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/54a45b98db6cf9dcae184341b11c3ada9efb63b5c8c0222a21ea269ead7746f1.jpg", "reference_inputs/images/fe8e110d6bb38794d1bebd55d8ac04077ab3b8a2e90fe6f56803526d46685019.jpg", "reference_inputs/images/8d896736401234e10331aeea3ac1fabfe73c9ad39a25d19bf9916657bd895404.jpg", "reference_inputs/images/b683797c51a05883b7e76e4ff54291fc1eb45bd00558f4d84c85fd38534d3947.jpg", "reference_inputs/images/d805d4a2c5ab72663e1777357194e745b5bdfd434697b3d4f6a88ba3b05e17b3.jpg", "reference_inputs/images/1c045cc9d56ac5597dd52c6857718f20cd7716c7605e6a518351ebcde2bc3b2d.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_20m/000005", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[246, 432], [246, 442], [245, 454], [244, 466], [242, 482]]\nB: [[285, 435], [284, 445], [283, 457], [282, 470], [281, 485]]\nC: [[301, 478], [302, 463], [303, 450], [304, 438], [305, 429]]\nD: [[323, 438], [322, 448], [322, 460], [321, 473], [319, 488]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/0afbd9401679e8a5f9f739e851e2d6565ca0847966b9c870ff25e6bd5f0227c3.jpg", "reference_inputs/images/a13aec45195756770182b27499d472f47b160ff507dee6ba042827d5c22fdc1c.jpg", "reference_inputs/images/0a637c542cd55efcd7db57209cb296cd5b3b8122a25a1081e21a75d2467f1eec.jpg", "reference_inputs/images/a37278d6f61a09212963b84767982a4ff7065abe6dae528d356c579d594d5b87.jpg", "reference_inputs/images/daa4685c900c7485f0b3274935690a92a3a0e3d43a08f6d210f5627453dcfb6d.jpg", "reference_inputs/images/f53561ce652da8ac3b7e5f0f6c6de9c351851774d487ef2584080a2549e351c1.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_20m/000006", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[551, 555], [551, 523], [551, 498], [551, 477], [551, 460]]\nB: [[512, 555], [512, 523], [512, 498], [512, 477], [512, 460]]\nC: [[491, 468], [491, 485], [491, 506], [491, 531], [491, 563]]\nD: [[473, 555], [473, 523], [473, 498], [473, 477], [473, 460]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/eb1a89bc11789e99b19e77585fd46e9e36fae0d3273a75e9ebfa9d805ddb8203.jpg", "reference_inputs/images/47f9e862659644a60fa5865a905e3616ae64a6438419abb7aae6d257478f5f7c.jpg", "reference_inputs/images/d6c7b56b517be5f9269a8f041fc150e7eb5f171fcdeee107c90b0307faf41b76.jpg", "reference_inputs/images/f288b97252aa8d63814994094c198f70d705ed1b8a7f816ac192773ad6f496dd.jpg", "reference_inputs/images/c2a6e42bede08dfc9058c2937c934922480236f09a08a93483f014a24a75068f.jpg", "reference_inputs/images/0cf563e80bf753791935a6b46f62138fe2d06824f6dfa4f844a21ad516ce6205.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_20m/000007", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[551, 546], [551, 517], [551, 492], [551, 473], [551, 457]]\nB: [[512, 546], [512, 517], [512, 492], [512, 473], [512, 457]]\nC: [[473, 546], [473, 517], [473, 492], [473, 473], [473, 457]]\nD: [[492, 465], [492, 481], [492, 500], [492, 525], [492, 554]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/be2f4479c1d90ca348b9c4b183bbb4721f60985a5f6d83c0b7f33fda89db8133.jpg", "reference_inputs/images/75bb9f3ac1921d7fed1cce6bc9bbfa9956c3988c4fcf366d0c757f6c48b81399.jpg", "reference_inputs/images/50a8652b77bf5cbed4b0cd0bb8bab9494f07e35a1602248d716e193cd683d51a.jpg", "reference_inputs/images/00a29879f7698025a5ffb5abd742219b52f130a6c1dbfc47820d2544470eddf8.jpg", "reference_inputs/images/f3d0a4d31d16c204ff9a76a17bcb947c6cb4b7bbb5cf7f59f6e1d7ad3080d41a.jpg", "reference_inputs/images/121ca0a666b27ee44d92209d4b2bcecda33df73f1f6332b6dff01271e6619a5e.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_20m/000008", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[512, 554], [512, 522], [512, 496], [512, 476], [512, 459]]\nB: [[491, 467], [491, 484], [491, 504], [491, 530], [491, 562]]\nC: [[473, 554], [473, 522], [473, 496], [473, 476], [473, 459]]\nD: [[551, 554], [551, 522], [551, 496], [551, 476], [551, 459]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/155fc91be7a97a5fdc159453201df0ee7518aad3bfa301d6c7114bad16e55843.jpg", "reference_inputs/images/bb1234fc75db8b2004f6e5dad03553677922756abe73a2bc76ac3db2608e9339.jpg", "reference_inputs/images/b01bcd72038777584cc4a4cc511b5af23e6bd0cf19996a6ed670e78ddfb4a5e1.jpg", "reference_inputs/images/f1cd24ff6d8c28314e77e586a6057bbb97f01a83e0555e13b0911369b7e423eb.jpg", "reference_inputs/images/861fd07722f02e7775960ebcbb39ddd697e58e8a2aa057c62f68cc8e41bceecf.jpg", "reference_inputs/images/dbdbafd1873f8374fc2da3b5dc8e39767781e722be8ed25fdcf2da41eac4a224.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_20m/000009", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[512, 467], [512, 454], [512, 439], [512, 428], [512, 418]]\nB: [[473, 467], [473, 454], [473, 439], [473, 428], [473, 418]]\nC: [[551, 467], [551, 454], [551, 439], [551, 428], [551, 418]]\nD: [[493, 426], [493, 436], [493, 447], [493, 462], [493, 475]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/3612c2cbf8424d2e8bf1f7b8bde5f9717a17e55bf38fa6195192aee8217a3baa.jpg", "reference_inputs/images/bcdc1e69c0a36fdd6f81b6dc138bac8948640e1f401b454a6715a6941bd5a4a5.jpg", "reference_inputs/images/c0d2a3e08cf989cab00b16bae02eabdb9d196be80eb7ba851f154ec5d4a68f11.jpg", "reference_inputs/images/ed2189f53b3f6d1c05bceb524c0005a7508a1ff0431f3713d530ae623c5b2642.jpg", "reference_inputs/images/34d027dbab11397f568f524e93a1358bb23b7502da274b0c04a1dba07548b837.jpg", "reference_inputs/images/0183cfd2866b057e06464418777bcf14316141f2c7c9027b08baf55273d48d83.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_20m/000010", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[512, 499], [512, 478], [512, 461], [512, 446], [512, 434]]\nB: [[551, 499], [551, 478], [551, 461], [551, 446], [551, 434]]\nC: [[473, 499], [473, 478], [473, 461], [473, 446], [473, 434]]\nD: [[493, 442], [493, 454], [493, 469], [493, 486], [493, 507]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/958716e20e8aa6ab2b8da68491a28dc338547810169b674e42f010812573df9d.jpg", "reference_inputs/images/502fe2b7bf5d1fe5c44e96178c57cdc8f6cda7fad2cf013c6853ede35a442838.jpg", "reference_inputs/images/07a0ad132afd99207a266320b96b576f8cf8b1518a143c3d17471bc368268383.jpg", "reference_inputs/images/9c4e551e68a3a6d6aa8c608c169cf03df9762fc2789534aeba9a6a5bf69ba888.jpg", "reference_inputs/images/0babe2b74ff3bef3cdac98263fe4524186cb1fd3c2bd927891a0e7a8de627389.jpg", "reference_inputs/images/bc9ecf1a01f58e317cba1408870ca8cd9224a3287a4cfcfb65985c9815fc847b.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_20m/000011", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[639, 550], [639, 418], [639, 290], [639, 158], [639, 23]]\nB: [[603, 519], [549, 417], [496, 318], [441, 214], [386, 110]]\nC: [[454, 478], [439, 390], [424, 305], [408, 216], [393, 126]]\nD: [[512, 550], [512, 418], [512, 290], [512, 158], [512, 23]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/0270d8b538279fc1cc2d42738a8e13de18ac9681608ba8b8b64e1d8f6abef7bd.jpg", "reference_inputs/images/ca8487ff33e0062cfeb68ed777e96a88e0c9431a893d7a94f47e9934bcac648e.jpg", "reference_inputs/images/59e4aa2a08930d64de52105fa7f3349672d21f92bf3adba38145d6c228691901.jpg", "reference_inputs/images/b0af3bb36620b6a360c476abd2224bdf5db62721179541bcc51903f4cb00b64f.jpg", "reference_inputs/images/7bb72823dd1ff767bcfa0af18ffdf886134eb00fa75108eafdef971bd5ab925b.jpg", "reference_inputs/images/3e7836c5acb2dcdcaa8ae2df4de0058e570a9f85222862e2d16221447f8c8257.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_20m/000012", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[473, 549], [473, 518], [473, 498], [473, 474], [473, 458]]\nB: [[492, 466], [492, 482], [492, 506], [492, 526], [492, 557]]\nC: [[512, 549], [512, 518], [512, 498], [512, 474], [512, 458]]\nD: [[551, 549], [551, 518], [551, 498], [551, 474], [551, 458]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/bfbff7adbebc42ae595d39371347b63219639483ac4c74142b2bc61623dabbd5.jpg", "reference_inputs/images/f4a7098cbe94ae907eafc0f5997b8a099436c210eb839ecea4f7bac29d27b2a5.jpg", "reference_inputs/images/bcdc1e69c0a36fdd6f81b6dc138bac8948640e1f401b454a6715a6941bd5a4a5.jpg", "reference_inputs/images/c0d2a3e08cf989cab00b16bae02eabdb9d196be80eb7ba851f154ec5d4a68f11.jpg", "reference_inputs/images/ed2189f53b3f6d1c05bceb524c0005a7508a1ff0431f3713d530ae623c5b2642.jpg", "reference_inputs/images/34d027dbab11397f568f524e93a1358bb23b7502da274b0c04a1dba07548b837.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_20m/000013", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[473, 432], [473, 422], [473, 413], [473, 406], [473, 398]]\nB: [[493, 406], [493, 414], [493, 421], [493, 430], [493, 440]]\nC: [[512, 432], [512, 422], [512, 413], [512, 406], [512, 398]]\nD: [[551, 432], [551, 422], [551, 413], [551, 406], [551, 398]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/efecc6ea655d2baf9c731021df4e0e26e6ddae097bd52949d1ab78c194d2649f.jpg", "reference_inputs/images/cbfef6891d3e74555360b427a6a7211fb069abf7f7d51a894e28baa461754599.jpg", "reference_inputs/images/e8f86a144ff07fab1b6ed4595caeffbf95d4782f342c62d61c804499bb0065bf.jpg", "reference_inputs/images/199b8fbe4a152e9d5ee803958dac2f67a88f2cf0659fdff516308206c27adca2.jpg", "reference_inputs/images/7f94190665dd70f7d3a511d556daa5af6ea4f449ea5a64eaf2788cf6931abf62.jpg", "reference_inputs/images/2c4d8a2bd0304a00d3fcdc28715ad25aedb7c02628a93f4a466fc02ebe7ae0fe.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_20m/000014", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[493, 450], [493, 465], [493, 481], [493, 500], [493, 525]]\nB: [[473, 517], [473, 492], [473, 473], [473, 457], [473, 442]]\nC: [[512, 517], [512, 492], [512, 473], [512, 457], [512, 442]]\nD: [[551, 517], [551, 492], [551, 473], [551, 457], [551, 442]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/be2f4479c1d90ca348b9c4b183bbb4721f60985a5f6d83c0b7f33fda89db8133.jpg", "reference_inputs/images/50a8652b77bf5cbed4b0cd0bb8bab9494f07e35a1602248d716e193cd683d51a.jpg", "reference_inputs/images/00a29879f7698025a5ffb5abd742219b52f130a6c1dbfc47820d2544470eddf8.jpg", "reference_inputs/images/f3d0a4d31d16c204ff9a76a17bcb947c6cb4b7bbb5cf7f59f6e1d7ad3080d41a.jpg", "reference_inputs/images/121ca0a666b27ee44d92209d4b2bcecda33df73f1f6332b6dff01271e6619a5e.jpg", "reference_inputs/images/81046b24d8b0d3e578bdfbf2026c08137a814732d76eb9b298a0b6178fe200d5.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_20m/000015", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[512, 498], [512, 477], [512, 460], [512, 446], [512, 434]]\nB: [[551, 498], [551, 477], [551, 460], [551, 446], [551, 434]]\nC: [[453, 498], [486, 477], [516, 460], [540, 446], [559, 434]]\nD: [[473, 498], [473, 477], [473, 460], [473, 446], [473, 434]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/eb1a89bc11789e99b19e77585fd46e9e36fae0d3273a75e9ebfa9d805ddb8203.jpg", "reference_inputs/images/f288b97252aa8d63814994094c198f70d705ed1b8a7f816ac192773ad6f496dd.jpg", "reference_inputs/images/c2a6e42bede08dfc9058c2937c934922480236f09a08a93483f014a24a75068f.jpg", "reference_inputs/images/0cf563e80bf753791935a6b46f62138fe2d06824f6dfa4f844a21ad516ce6205.jpg", "reference_inputs/images/2bd4d74ba8c6cb9ef4b662f6608f31292eb3651d57b98de1bdde0f1128d409c7.jpg", "reference_inputs/images/5fcd13c014fa1d1e4b7545218a236cf501122c460a36ef9cad0fc7c862f11600.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_20m/000016", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[659, 370], [635, 395], [615, 420], [602, 443], [591, 466]]\nB: [[608, 399], [614, 408], [619, 418], [626, 430], [633, 442]]\nC: [[642, 380], [647, 389], [653, 398], [659, 410], [666, 423]]\nD: [[646, 426], [638, 413], [632, 401], [626, 391], [621, 382]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/1a5875388e9777c9da6a86358adce81ccf7c461253ed1c37b5bfb0d1e954f77e.jpg", "reference_inputs/images/a13aec45195756770182b27499d472f47b160ff507dee6ba042827d5c22fdc1c.jpg", "reference_inputs/images/0a637c542cd55efcd7db57209cb296cd5b3b8122a25a1081e21a75d2467f1eec.jpg", "reference_inputs/images/a37278d6f61a09212963b84767982a4ff7065abe6dae528d356c579d594d5b87.jpg", "reference_inputs/images/daa4685c900c7485f0b3274935690a92a3a0e3d43a08f6d210f5627453dcfb6d.jpg", "reference_inputs/images/f53561ce652da8ac3b7e5f0f6c6de9c351851774d487ef2584080a2549e351c1.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_20m/000017", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[522, 542], [439, 417], [406, 286], [439, 153], [522, 21]]\nB: [[385, 542], [385, 417], [385, 286], [385, 153], [385, 21]]\nC: [[512, 542], [512, 417], [512, 286], [512, 153], [512, 21]]\nD: [[602, 512], [550, 415], [496, 313], [441, 210], [386, 107]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/5227f1514bc26367b9a685962255b280ec8c82f911ad7282b2912b36f5b3dc4c.jpg", "reference_inputs/images/9bd1be9135b490482662595471ecb13389cc21f0185e9afd81e9b1cf2152433c.jpg", "reference_inputs/images/91613324c1c5a055df08a78e89bc88cb93a4e51fd5b21ca12316859c761a71ab.jpg", "reference_inputs/images/1dd8da1c6e720f3f329f418fddf85c33c24dc0d7eddb073383e93040bd79026b.jpg", "reference_inputs/images/b4124de2733458d6b02f2a8d6824132a5a497241fa556c58ee1b573e171b45bc.jpg", "reference_inputs/images/008072f8938f8f1a63f207a5da458f8c1016122eb4926ea5e9575db7ace1132e.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_30m/000001", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[450, 430], [496, 343], [542, 255], [586, 174], [633, 86]]\nB: [[551, 494], [571, 382], [590, 270], [609, 166], [629, 55]]\nC: [[625, 434], [625, 346], [625, 258], [625, 174], [625, 86]]\nD: [[512, 434], [512, 346], [512, 258], [512, 174], [512, 86]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/3c008430886d13cc06b3af1ca92b44323c9fdc224aa91af5bf1a822dea3e92b5.jpg", "reference_inputs/images/b16a03557227e8f95de9c50c426c05f1c1350b29fff4a4bc8e35879fa14681f1.jpg", "reference_inputs/images/fb65f7995f334114946a52d3172118daba99133ed34921956d2365e3e709e2a6.jpg", "reference_inputs/images/9bd1be9135b490482662595471ecb13389cc21f0185e9afd81e9b1cf2152433c.jpg", "reference_inputs/images/91613324c1c5a055df08a78e89bc88cb93a4e51fd5b21ca12316859c761a71ab.jpg", "reference_inputs/images/1dd8da1c6e720f3f329f418fddf85c33c24dc0d7eddb073383e93040bd79026b.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_30m/000002", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[455, 342], [445, 283], [434, 225], [424, 168], [414, 110]]\nB: [[454, 62], [454, 149], [454, 234], [454, 320], [454, 407]]\nC: [[450, 381], [496, 294], [542, 209], [586, 126], [632, 39]]\nD: [[512, 385], [512, 298], [512, 211], [512, 126], [512, 39]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/d23def54bc38c6919f30f1cf05d51e32af4daf8d2413d5881837e1e5da81ad0b.jpg", "reference_inputs/images/ddd5cc7103e6c241522c5a84922eb8dbb23cc83dd72ec92bc371f4d0b684f4ce.jpg", "reference_inputs/images/d51ea870530d60397e31de4daf06a29f1f5c6919ef5f1ec66f0e45ab6e30faa2.jpg", "reference_inputs/images/eb97a3246d862464d2b51b720f73d209b8d3a46bb24e54d9e5a2b762b5dfa240.jpg", "reference_inputs/images/9668b0e9de3420dc53f6742f0eaa1d9277e1041769fad82048ac9ad8b7f0e19a.jpg", "reference_inputs/images/57c3f70f35fc88b642076e72a90f518e3004de43237b3e468d97c6500ead1958.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_30m/000003", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[512, 362], [512, 275], [512, 190], [512, 102], [512, 14]]\nB: [[399, 362], [399, 275], [399, 190], [399, 102], [399, 14]]\nC: [[569, 349], [533, 282], [498, 215], [462, 147], [426, 79]]\nD: [[341, 362], [440, 275], [526, 190], [594, 102], [650, 14]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/acba93faab2b2b94bd8f121833dcdc0686ddae44c5819cbb8c7ea4fc08b05082.jpg", "reference_inputs/images/c0316a436e666a9a4e034067935dde38dd44cd7999b44800d78aff318cd543b9.jpg", "reference_inputs/images/4ea6834c85dd9da4d0e393b1fbd27477d525a255f2c04252747fb6893e6cca2b.jpg", "reference_inputs/images/0434ee60b659113cb97e064670e8b02e9955eeff7f24ccedb8f970345927aa67.jpg", "reference_inputs/images/576d8be8f3dbf01571e62e0948abed9b4f3a1622dc4e9a7e3af95ef90daf400e.jpg", "reference_inputs/images/9f89d5cdd755643210dfe8c38a9f83fd2ae46616e34f41ba1e56266b6bf6ac4d.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_30m/000004", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[512, 414], [512, 327], [512, 239], [512, 153], [512, 67]]\nB: [[450, 410], [495, 324], [542, 237], [587, 152], [632, 67]]\nC: [[454, 90], [454, 175], [454, 262], [454, 350], [454, 436]]\nD: [[520, 414], [439, 327], [406, 239], [439, 153], [520, 67]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/0fef7bf43103cdecd4e2c7ba577149ffb74ee952c482f428f478d115ae197001.jpg", "reference_inputs/images/49f27fb1f76eac5e0f52afca875ebe856b4e9f60dcc74105da3782689cdaf5cf.jpg", "reference_inputs/images/55a985bde0abb3e5ad08f9b837111c168ca3381e9df6347ad27a6ef7782536a7.jpg", "reference_inputs/images/a6bbb87b287f5401a21be4253d6a9ba6d1b1fb7dd5a4446eec287e86133e81c1.jpg", "reference_inputs/images/4c59aef792f7b4cb068ed9ec42b7d976b072de55be5002948870efa2648c1b5f.jpg", "reference_inputs/images/010870be1ff3452306b89e08d9db3ca31baec1660cdaf57e6ddea918045c7365.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_30m/000005", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[625, 554], [625, 466], [625, 383], [625, 295], [625, 207]]\nB: [[341, 554], [440, 466], [526, 383], [594, 295], [650, 207]]\nC: [[512, 554], [512, 466], [512, 383], [512, 295], [512, 207]]\nD: [[569, 542], [533, 473], [498, 408], [462, 340], [426, 271]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/79637357a794d114c87fa20d24a320beb4983fb07bc013db84fecd8b57b8d6ee.jpg", "reference_inputs/images/c39e0e236839a3d83a8b6350bc0efe76a6fbc67aa9b6e4c2fc368e7de0a592cb.jpg", "reference_inputs/images/1e39378ac5766e78a1ddbeac3d2435169317ebdbb78727242c19741fb2c1f858.jpg", "reference_inputs/images/3eaadbc0c25a3b5713f45480a423aeb3ce28d83db7d841b55cc57d979135108f.jpg", "reference_inputs/images/1f2a00a0e56882a06b6b2ad1adfb16344a112cd9c178bc762cf53cbf501e6f03.jpg", "reference_inputs/images/17f870c3e80b4ea658ab9fad190bfd9efb07496fc221e0102bd1988432097401.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_30m/000006", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[841, 501], [841, 489], [841, 478], [841, 468], [841, 461]]\nB: [[880, 501], [880, 489], [880, 478], [880, 468], [880, 461]]\nC: [[802, 501], [802, 489], [802, 478], [802, 468], [802, 461]]\nD: [[822, 469], [822, 476], [822, 486], [822, 497], [822, 509]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/8965f22b2f72d587c9a9005be33de5ff477b593ae296956517dbaae4a1e4a5db.jpg", "reference_inputs/images/314d01bbfb4d254b346e7f6bc4fb62da813832c57ce8b979e13a5d0aae67e2bd.jpg", "reference_inputs/images/f2449548a2fb5afee173a59713a49688fa35c32ab2d5843906f2f9ba6baa0112.jpg", "reference_inputs/images/2a4ced4ecbda3f0ad9bfde9034be5e0a49ec9bab9c712884affd3ad08ecb94fb.jpg", "reference_inputs/images/1920d160bf8c207cc2cfe03c2f54780add45804d245793d2333db80413b5314a.jpg", "reference_inputs/images/326d0df537ddef5c1539b9ddbad502e4b450bb1d6abf8df538cadcd6be629d0a.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_30m/000007", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[450, 447], [496, 362], [540, 278], [587, 191], [633, 104]]\nB: [[625, 451], [625, 365], [625, 281], [625, 192], [625, 104]]\nC: [[569, 438], [533, 371], [498, 306], [462, 237], [425, 169]]\nD: [[512, 451], [512, 365], [512, 281], [512, 192], [512, 104]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/223cbee8041d6f1a55407a61b86a321c281765434511cd82b74d02a28bad5c15.jpg", "reference_inputs/images/3a78694109d6f4e2aa6c3201e224184d3c1c7c4736b8adac3b14a0233f1f3ea6.jpg", "reference_inputs/images/1fcb23b032daeca65a1a4f829f0e94b427e95c2ad50d71810101b247352557ba.jpg", "reference_inputs/images/d3a0110b6a67af346162179241e854b0f3917a8527d17d0238e525236b325fc3.jpg", "reference_inputs/images/861d640dd21fed339d57382a5df6359c6384cb724a93624887b95b2828f92d5b.jpg", "reference_inputs/images/5c64e995683b1de21c283c671349dac82351d1b3a64375df3ccdb6bffaedb802.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_30m/000008", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[625, 417], [625, 332], [625, 246], [625, 158], [625, 69]]\nB: [[569, 404], [534, 338], [498, 271], [462, 202], [425, 134]]\nC: [[520, 417], [439, 332], [406, 246], [439, 158], [520, 69]]\nD: [[512, 417], [512, 332], [512, 246], [512, 158], [512, 69]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/ef9daea12006a2e51e5c982a5b280c67b50689e9b2b397d8028d0f4166e33386.jpg", "reference_inputs/images/5050b13e3207376f5eaac13dd4c2b17aa7de047c93b4adb173daf9a227c6bf3b.jpg", "reference_inputs/images/6de9c3585cf47e11898d5f77222fa22a23cf12d738d4e6bd2252109ea1039268.jpg", "reference_inputs/images/82e49c092a52ba14ad4d6e963fee05e3b260013b8eddacfdcbdcbc606ff915ef.jpg", "reference_inputs/images/6fbf9e0872d4e16add6d278a01df5aa9f4b19360391d2d9887a86189ddce823b.jpg", "reference_inputs/images/ee43fb8141aa4576e3ebdf22ef02f663d6e11cac7f7e49ef76189736638c9d2a.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_30m/000009", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[473, 503], [473, 489], [473, 475], [473, 464], [473, 454]]\nB: [[512, 503], [512, 489], [512, 475], [512, 464], [512, 454]]\nC: [[493, 462], [493, 472], [493, 483], [493, 497], [493, 511]]\nD: [[551, 503], [551, 489], [551, 475], [551, 464], [551, 454]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/4bbaf7b875b927803278ab0e3bbe127dc986d7e632e71250f020a2395eb683e5.jpg", "reference_inputs/images/34d027dbab11397f568f524e93a1358bb23b7502da274b0c04a1dba07548b837.jpg", "reference_inputs/images/0183cfd2866b057e06464418777bcf14316141f2c7c9027b08baf55273d48d83.jpg", "reference_inputs/images/37fdf1cfa1885a50dc1e3a45f4b4196e30f9e8011ea8b00b2aa25e97696ee22d.jpg", "reference_inputs/images/bb5fb5ea9e5337cd8439330bc758a7b429296eb13e3a974b49d53f06abdc07c5.jpg", "reference_inputs/images/d58c9ae08eef29500a5bf2870e2b136cf317de0992e07d5be2a8fc5ab35b65a2.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_30m/000010", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[946, 427], [970, 280], [979, 191], [962, 162], [882, 130]]\nB: [[1006, 313], [995, 232], [974, 201], [933, 217], [834, 222]]\nC: [[826, 282], [925, 278], [966, 262], [987, 293], [998, 374]]\nD: [[979, 319], [1004, 238], [998, 198], [951, 189], [856, 138]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/38cbd1ff93865b744eb4349f981020bb46720241f0906414c243eb67eea9a2b1.jpg", "reference_inputs/images/9b83d6129aa61cdfbf7ff383b9a29ddbba8ea0418bbd5a27c7cb21866300128a.jpg", "reference_inputs/images/b829ddf4c587e79a67ffe16fe25873fe19448bfe353db00f3602571b6511c0a3.jpg", "reference_inputs/images/5a3d23ade29fc32a74a40c56381bcee2a1f0227641011a213150ba8ecfd7e2f5.jpg", "reference_inputs/images/a0ccf0f5e696281230600960e2603ce51a994ea07386f34ba1410ded71be1b4c.jpg", "reference_inputs/images/6e0075721026d6fd27359c56d0e9c4227a96e72271da50ba1dddeb99595ec730.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_30m/000011", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[454, 312], [444, 254], [434, 195], [424, 138], [414, 79]]\nB: [[454, 30], [454, 118], [454, 203], [454, 291], [454, 378]]\nC: [[569, 342], [533, 275], [497, 206], [462, 141], [426, 72]]\nD: [[512, 356], [512, 269], [512, 181], [512, 96], [512, 8]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/78c4d62d677825deee57612396c5c7a9551bb4faba10dc9c14391fb1ab84f3b3.jpg", "reference_inputs/images/53851ffaf7a1cd23d50ed89c31c148ef608b08af5c51d1c73461e56cac54c969.jpg", "reference_inputs/images/ca8487ff33e0062cfeb68ed777e96a88e0c9431a893d7a94f47e9934bcac648e.jpg", "reference_inputs/images/59e4aa2a08930d64de52105fa7f3349672d21f92bf3adba38145d6c228691901.jpg", "reference_inputs/images/b0af3bb36620b6a360c476abd2224bdf5db62721179541bcc51903f4cb00b64f.jpg", "reference_inputs/images/7bb72823dd1ff767bcfa0af18ffdf886134eb00fa75108eafdef971bd5ab925b.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_30m/000012", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[459, 442], [446, 385], [435, 327], [425, 268], [416, 213]]\nB: [[516, 485], [514, 399], [512, 312], [512, 224], [512, 142]]\nC: [[627, 483], [625, 398], [623, 310], [623, 222], [623, 141]]\nD: [[573, 470], [535, 405], [498, 338], [462, 270], [428, 206]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/4ab1ce6d791179a724a0abda241d57be67280c3d399e30581a826bced7e1b7b5.jpg", "reference_inputs/images/1c35f6d72b01ec6af2caa178bff92d13f3a4d9b4ec57b5e271a75bff2446db9a.jpg", "reference_inputs/images/aa07265962ce9129c110ee741915093469cd3f25c12624b6ce009dc9b39e385e.jpg", "reference_inputs/images/ff6c1dc8e7cddbfa32a9b40096be7e72aaf3a09ae2f0b40094b712370828563c.jpg", "reference_inputs/images/6f1921492b831a61cabf2999fbd5323f30262e211f6d8f33afff514562e0ae71.jpg", "reference_inputs/images/137841317eedd34bac33fb0fa9528cf8515675bc81c503205b05f1026a3b2852.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_30m/000013", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[264, 555], [326, 492], [388, 428], [448, 366], [511, 300]]\nB: [[454, 276], [391, 342], [331, 404], [269, 468], [207, 531]]\nC: [[182, 475], [243, 412], [306, 348], [366, 286], [429, 220]]\nD: [[253, 482], [286, 433], [321, 382], [354, 334], [388, 282]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/17e8acdee8139296e74054ad46f0b8e4ac3f31e20e58c6ad512a0f49c246c007.jpg", "reference_inputs/images/df47bcfa9e699d669e8d001dc1879ba67c1773063bf65c63a154588d61b651e2.jpg", "reference_inputs/images/b6aff2a8fef54cc5e165c564801057e32d68c9d91f8dcc1b31cb210563a545a0.jpg", "reference_inputs/images/8ea9525644723c20993995ad28a6ce3f93b6653ba9b3e1355c8ee62caa16f50d.jpg", "reference_inputs/images/eac4ef4f05689cc798bd3c9cd2a197d403c4ce41f66c184cab23e8b8adc70e69.jpg", "reference_inputs/images/0cca2cfe4b04a0ea15f1778800655229e8c0597740119369e5253f1ed89d2aba.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_30m/000014", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[624, 415], [624, 329], [624, 241], [624, 154], [624, 69]]\nB: [[512, 415], [512, 329], [512, 241], [512, 154], [512, 69]]\nC: [[450, 411], [495, 326], [542, 238], [587, 153], [632, 69]]\nD: [[454, 91], [454, 176], [454, 263], [454, 351], [454, 438]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/fbd220bfde2098e8a3e7cde089e706c09b0c83f795a2fa162d6194e1521da1fb.jpg", "reference_inputs/images/0f0ea43f2b6fd8241a5377bbd7cd403bd37949f2e803d8fe2a088645c622dd3b.jpg", "reference_inputs/images/c29d2ae0b8e4d26fa8e1494af668e571e75459fc2c61d548d1fdfe35fbebe69b.jpg", "reference_inputs/images/3a78694109d6f4e2aa6c3201e224184d3c1c7c4736b8adac3b14a0233f1f3ea6.jpg", "reference_inputs/images/1fcb23b032daeca65a1a4f829f0e94b427e95c2ad50d71810101b247352557ba.jpg", "reference_inputs/images/d3a0110b6a67af346162179241e854b0f3917a8527d17d0238e525236b325fc3.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_30m/000015", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[455, 521], [445, 463], [434, 405], [425, 348], [414, 290]]\nB: [[569, 551], [533, 484], [497, 416], [462, 350], [426, 283]]\nC: [[512, 564], [512, 478], [512, 390], [512, 306], [512, 219]]\nD: [[624, 564], [624, 478], [624, 390], [624, 306], [624, 219]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/01bec1e6063f2abdaa5a2a82f0cc79873e8527ba96cf4d200ccc3916f0c1bc18.jpg", "reference_inputs/images/9c4e551e68a3a6d6aa8c608c169cf03df9762fc2789534aeba9a6a5bf69ba888.jpg", "reference_inputs/images/0babe2b74ff3bef3cdac98263fe4524186cb1fd3c2bd927891a0e7a8de627389.jpg", "reference_inputs/images/bc9ecf1a01f58e317cba1408870ca8cd9224a3287a4cfcfb65985c9815fc847b.jpg", "reference_inputs/images/44166659bf064b7b7f17ebe1cd27d0ffcc7acc728691ab96f5c75bf2df465995.jpg", "reference_inputs/images/ab4684364fe6a38c2026ce521fb39e2527721bca2271eb4e42b0311ffe907e38.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_30m/000016", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[454, 42], [454, 129], [454, 228], [454, 302], [454, 388]]\nB: [[568, 353], [532, 286], [502, 229], [461, 152], [425, 84]]\nC: [[512, 366], [512, 280], [512, 206], [512, 106], [512, 19]]\nD: [[454, 323], [444, 266], [435, 216], [424, 150], [414, 91]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/54fe527ada2d1060e9abe51e500c229043388ec7321da0fcfb22ea61353b10b0.jpg", "reference_inputs/images/9dabade18b21c6aa67c58127d7a894f94f432a931037529764f20ab13e6c9cdc.jpg", "reference_inputs/images/5f7d374ea1f609856cc0a326dbce75f0c14822bcb32ea1b54fe31767631b1b79.jpg", "reference_inputs/images/794cfbcbe7aebf0fd07de502de177b39301e6da824fb87e6bb1e4aa716ca8342.jpg", "reference_inputs/images/16b98ca976b6de186d99c34f240152f162895e25eef46ec2f005aaf3034c6c1a.jpg", "reference_inputs/images/7b76f625b0e713b81c3b2fd9005caf85291cd0af2b37d0152ae222104a3abd57.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_30m/000017", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[822, 486], [822, 497], [822, 509], [822, 523], [822, 539]]\nB: [[880, 531], [880, 515], [880, 501], [880, 489], [880, 478]]\nC: [[782, 531], [815, 515], [845, 501], [869, 489], [888, 478]]\nD: [[841, 531], [841, 515], [841, 501], [841, 489], [841, 478]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/8965f22b2f72d587c9a9005be33de5ff477b593ae296956517dbaae4a1e4a5db.jpg", "reference_inputs/images/74e9269bc62e7cfe049687aacc67f3e47376e38cdd99138c4d65d99c4cbe82fa.jpg", "reference_inputs/images/c6ac3c4fb2e5c73c58f26744d3b42a68e4504e8ec1b0f59cd9733ed6e295c7be.jpg", "reference_inputs/images/314d01bbfb4d254b346e7f6bc4fb62da813832c57ce8b979e13a5d0aae67e2bd.jpg", "reference_inputs/images/f2449548a2fb5afee173a59713a49688fa35c32ab2d5843906f2f9ba6baa0112.jpg", "reference_inputs/images/2a4ced4ecbda3f0ad9bfde9034be5e0a49ec9bab9c712884affd3ad08ecb94fb.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_30m/000018", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[881, 216], [808, 207], [704, 194], [633, 185], [531, 172]]\nB: [[852, 443], [779, 434], [675, 422], [604, 412], [502, 399]]\nC: [[866, 330], [794, 321], [690, 308], [618, 298], [517, 286]]\nD: [[861, 273], [801, 296], [715, 330], [655, 351], [571, 383]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/9387fb07c1ea73e8176f603aa4bc4630b33ca3cd16a48c5c1865d623fa341a4d.jpg", "reference_inputs/images/5b6ac385e3afda8add29752aadc7568dcfefbad224d4b7851478badc13cf2acf.jpg", "reference_inputs/images/6f0f1e07119f53abe8f08c28c43962d83b3debe85de2c64a6e89e21ebb95a02c.jpg", "reference_inputs/images/a48a18e7653d72be7bef71d38f63063825f2392195ab9c9f2863e8d057499145.jpg", "reference_inputs/images/37f3b24621705c5ca5ca35f0206472c1147c1f78bba9a690554c9eed51564071.jpg", "reference_inputs/images/0954fbaf2620c6f5b47a8b30ce0b61391a5fa9e332bb66020514f148e81de546.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_30m/000019", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[484, 333], [458, 314], [426, 258], [387, 207], [346, 160]]\nB: [[587, 306], [550, 293], [486, 238], [418, 192], [349, 150]]\nC: [[458, 388], [429, 354], [393, 265], [348, 182], [302, 103]]\nD: [[550, 338], [522, 303], [486, 214], [441, 132], [394, 53]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/e4d2d3e636b4d3b3ff2b02938c80dc6fca66ffd777ec8927b57b34db548cdb62.jpg", "reference_inputs/images/0c12d48b25c8bc4d5b0ac7ed14b32318c9bab11bb184b1cdbb5f363e3f5c7fe2.jpg", "reference_inputs/images/a7242eaf131592bc02284e73891cbb27e9b17fd35a10e882525c8d20c2cdd198.jpg", "reference_inputs/images/33479d35b44aa18cd3869f33debd528efca86b2a06e5055948e2dd6a95cfdcf7.jpg", "reference_inputs/images/48f58850513c81ed977c5c8c574a23c6de9d7218bc84a9d8d11a36c15e3b65cf.jpg", "reference_inputs/images/ed72fde262725b00d0b3ff570a94c71361b2530ae61b40bdbcfadcc1433ba4ae.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_30m/000020", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[400, 518], [400, 430], [400, 350], [400, 262], [400, 172]]\nB: [[512, 518], [512, 430], [512, 350], [512, 262], [512, 172]]\nC: [[624, 518], [624, 430], [624, 350], [624, 262], [624, 172]]\nD: [[454, 194], [454, 284], [454, 372], [454, 453], [454, 540]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/4d54ca25c075d2117a23d1a492c6e787b27d6949726c7e86152b21bbed22fa3a.jpg", "reference_inputs/images/ff6c1dc8e7cddbfa32a9b40096be7e72aaf3a09ae2f0b40094b712370828563c.jpg", "reference_inputs/images/6f1921492b831a61cabf2999fbd5323f30262e211f6d8f33afff514562e0ae71.jpg", "reference_inputs/images/137841317eedd34bac33fb0fa9528cf8515675bc81c503205b05f1026a3b2852.jpg", "reference_inputs/images/7314b8f1b877457153c9f79fd61f048a1d3a6dda0ef1c25315ce50ed1f1c9712.jpg", "reference_inputs/images/d760e180b6d0fe7d5f5740e5ff281233dff08dc94c3c056a4d088a7a0203454d.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_50m/000001", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[449, 169], [476, 163], [484, 183], [484, 232], [484, 285]]\nB: [[441, 286], [482, 220], [518, 161], [538, 132], [534, 130]]\nC: [[558, 250], [558, 197], [558, 148], [550, 128], [523, 134]]\nD: [[512, 264], [512, 211], [512, 162], [504, 142], [477, 148]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/e42a5676f10283e3c9ffd2990f7493f1bd9bb9244ef187158759ee6bb03f2c45.jpg", "reference_inputs/images/93f8c8c3cf249ce2b1f1c01c90e4fd3362d4265bc084ad006f6dc4e0f549605f.jpg", "reference_inputs/images/9b83d6129aa61cdfbf7ff383b9a29ddbba8ea0418bbd5a27c7cb21866300128a.jpg", "reference_inputs/images/b829ddf4c587e79a67ffe16fe25873fe19448bfe353db00f3602571b6511c0a3.jpg", "reference_inputs/images/5a3d23ade29fc32a74a40c56381bcee2a1f0227641011a213150ba8ecfd7e2f5.jpg", "reference_inputs/images/a0ccf0f5e696281230600960e2603ce51a994ea07386f34ba1410ded71be1b4c.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_50m/000002", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[466, 272], [493, 188], [508, 138], [499, 118], [458, 92]]\nB: [[512, 211], [512, 162], [504, 142], [477, 148], [420, 141]]\nC: [[542, 171], [542, 122], [534, 102], [507, 108], [450, 101]]\nD: [[409, 175], [466, 182], [493, 177], [501, 197], [501, 246]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/e42a5676f10283e3c9ffd2990f7493f1bd9bb9244ef187158759ee6bb03f2c45.jpg", "reference_inputs/images/9b83d6129aa61cdfbf7ff383b9a29ddbba8ea0418bbd5a27c7cb21866300128a.jpg", "reference_inputs/images/b829ddf4c587e79a67ffe16fe25873fe19448bfe353db00f3602571b6511c0a3.jpg", "reference_inputs/images/5a3d23ade29fc32a74a40c56381bcee2a1f0227641011a213150ba8ecfd7e2f5.jpg", "reference_inputs/images/a0ccf0f5e696281230600960e2603ce51a994ea07386f34ba1410ded71be1b4c.jpg", "reference_inputs/images/6e0075721026d6fd27359c56d0e9c4227a96e72271da50ba1dddeb99595ec730.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/city6_ground_aerial_v2_50m/000003", "task_name": "trajectory_reconstruction", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[546, 346], [525, 306], [503, 264], [482, 224], [461, 185]]\nB: [[579, 354], [579, 302], [579, 249], [579, 197], [579, 146]]\nC: [[512, 354], [512, 302], [512, 249], [512, 197], [512, 146]]\nD: [[445, 354], [445, 302], [445, 249], [445, 197], [445, 146]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/334da4287271cddf47d6c5e5c4598fa1148a358472a4dead85ffcd9246528b37.jpg", "reference_inputs/images/1be0d45fc5eb88503ea3915f0b8fe008a69f91db0a9f5307220b605630c8732b.jpg", "reference_inputs/images/3e0a4ac2e88225b48e8c398702ff8943834354639cb6fce673aafc0cc14fd71e.jpg", "reference_inputs/images/ab664e3a1c1d9879f2ca1d50406f4e89dff8c209dd7abe815855c9c49daa0388.jpg", "reference_inputs/images/772ae829aab097e813edc986df6337aa53a4dfb46188b630cff56caf74228b46.jpg", "reference_inputs/images/c5b600c09a6d37cd46165ccf13dda6e2f27b7d463018639951817900fbdef907.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000001", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[486, 472], [438, 491], [394, 510], [346, 533], [299, 554]]\nB: [[485, 551], [430, 515], [379, 482], [322, 447], [266, 413]]\nC: [[275, 546], [334, 549], [394, 552], [448, 557], [506, 562]]\nD: [[491, 510], [433, 505], [378, 500], [318, 497], [260, 494]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/3cd45db824ae2b9344e05536329dd03abac72f8c28af459c4d35886b6b2aebb6.jpg", "reference_inputs/images/ac9a067321586244d3dae504d5e0938ffa4068e23f67bbdf00ea1667bfd1597d.jpg", "reference_inputs/images/b8a086918400b04805c6ab807ac91ad2108f05ac9d2379c90888138768a44312.jpg", "reference_inputs/images/3287c0cde9d06973f68e05609776faa670ad090c0dde10bcec003ba20bdbc5b3.jpg", "reference_inputs/images/5130830fdc667694537466c1b0917d92aa254c90c62b32fe6253654bcdf7e859.jpg", "reference_inputs/images/2e21fad0690d5f1b76873564a0c197839d27731f4ddc73083afac1751880b3db.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000002", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[551, 498], [551, 475], [551, 458], [551, 444], [551, 432]]\nB: [[493, 440], [493, 452], [493, 466], [493, 483], [493, 506]]\nC: [[512, 498], [512, 475], [512, 458], [512, 444], [512, 432]]\nD: [[453, 498], [486, 475], [516, 458], [540, 444], [559, 432]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/9b77a6fdcbcd164e42d030afdba126226416d464d2bb4b234b73eec459bc4446.jpg", "reference_inputs/images/f51cdc4fdab962d37e5494600db16c9ce68b3079d84ae214af152e6704b497ed.jpg", "reference_inputs/images/702dd7ba3d6815564455245f3199e57483d587890f24e75a7c679226bfe87823.jpg", "reference_inputs/images/8ed404085ac069432ca6885fa46e77d31c1a9cee698a07ec666853280f8b6fd0.jpg", "reference_inputs/images/153c42acc503b7f49f9717f56313e3f98e87cf0cd70405ca91151993eafe3310.jpg", "reference_inputs/images/b28aa8aaad129d53f280c819325ac5c973dbd9c05b862f43953b1c6b3a2077d7.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000003", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[493, 452], [493, 466], [493, 483], [493, 505], [493, 532]]\nB: [[512, 524], [512, 497], [512, 475], [512, 458], [512, 444]]\nC: [[453, 524], [486, 497], [516, 475], [540, 458], [559, 444]]\nD: [[551, 524], [551, 497], [551, 475], [551, 458], [551, 444]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/c24189713c2885b33cdb7ec22d3ce8b0f26e20865d2a16e92c64822b1c8d6c4f.jpg", "reference_inputs/images/696db644fc0e650348f4be016e8e5fd9c181eee2de39d8b0f2f71ba4a8fde652.jpg", "reference_inputs/images/4b42b02e39f44393275ac5e9a5490bd7d17f5f4b37b6848eb0fc0c58462fd46a.jpg", "reference_inputs/images/8f4940264acc56f4ce551a90fee909e68ecf5da8ccb78f8c6385b3774c7c049d.jpg", "reference_inputs/images/b458fc1c79ce19d141010b566ef27d6d7d4f034b0f01d15dfc3ad0ca65fb4341.jpg", "reference_inputs/images/15efa56c21d2bc6e235bb1a3ccf4483e7a48fdd00089930beb82826e589d1667.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000004", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[512, 454], [512, 441], [512, 429], [512, 418], [512, 410]]\nB: [[493, 418], [493, 426], [493, 437], [493, 449], [493, 462]]\nC: [[453, 454], [486, 441], [516, 429], [540, 418], [559, 410]]\nD: [[551, 454], [551, 441], [551, 429], [551, 418], [551, 410]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/00de2bf5235b40f382f6df163362e3b5d54ea51747f7f4f5ff7c89b03b95e023.jpg", "reference_inputs/images/be43d0fa9ccc98bfbb58151e006e7adb4030e7198ec36539f4b672f76ba29675.jpg", "reference_inputs/images/a28860f9ddaa250432484793a3d3aca1595d7b0ffe5d4f6f3cc5fbd29e8417ad.jpg", "reference_inputs/images/ef1c9e15657d46d314b74cd595a8bf5a6aa2686d5fedec13ec8318fc99ba7f6c.jpg", "reference_inputs/images/579e066ff20752ad4dba9dc815b3ebad186cbafdfdbecc0c13b0cd4a53f43534.jpg", "reference_inputs/images/700b82317f7b7799736a2401b76a6c47038f11fbd151a0c02396195b2dcfde77.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000005", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[386, 522], [400, 496], [411, 474], [420, 458], [428, 443]]\nB: [[407, 441], [399, 455], [390, 472], [379, 494], [366, 519]]\nC: [[421, 540], [434, 514], [446, 493], [454, 476], [462, 462]]\nD: [[352, 503], [366, 478], [377, 456], [386, 439], [394, 425]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/99ff36e6b012edd7fa9d4c7b1bc65614497e52512fc40b1cf1cc29158b33f267.jpg", "reference_inputs/images/457b3c328ed29d70688f513d171a75f1026ad2788f161a3bf4f6ff4a87768152.jpg", "reference_inputs/images/09c75aba58c003ce40bb681fe7b9397e8cc38c0610726570eef95c39e242f617.jpg", "reference_inputs/images/b3326e82ebd0776424f0ffd76f9ec5461e4bb8834505c636702c283f24589fac.jpg", "reference_inputs/images/61971f754619b632336ae31ac3f8a605ddfa8ef8f46e1d2f7529f8a352928ade.jpg", "reference_inputs/images/efa5658689482f7351dfb2d3e76efc3166d2ac78bdd9cedd302f88a92851f48f.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000006", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[551, 470], [551, 454], [551, 441], [551, 429], [551, 418]]\nB: [[473, 470], [473, 454], [473, 441], [473, 429], [473, 418]]\nC: [[453, 470], [486, 454], [516, 441], [540, 429], [559, 418]]\nD: [[512, 470], [512, 454], [512, 441], [512, 429], [512, 418]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/09fa584c9f04996ceb55e3ee99a4ce42752e25465beb6c4fb60b0b255d4cf56c.jpg", "reference_inputs/images/b7cce4d0c6e9e1f8e3bbdf264066a794ae08968e56c63157c9245e38f9a67787.jpg", "reference_inputs/images/c63e1ff48fec679249f03adce2581fb770f181347ffa8ef2f6f887c9599ab6f5.jpg", "reference_inputs/images/84fa91bee78795f7087343d63acb876160d6be00b2cf4dafc8f6d77db679e763.jpg", "reference_inputs/images/6b5df6351b777266ba97aa5ce4ccfc7c33b5fafa8409b33b3fc6151bfd4b616a.jpg", "reference_inputs/images/cbf42e21aa5e52d5b279acc05c3f5cee3c30e56eea52a6a51464d271a4771a84.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000007", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[453, 459], [486, 445], [516, 432], [540, 422], [559, 413]]\nB: [[551, 459], [551, 445], [551, 432], [551, 422], [551, 413]]\nC: [[512, 459], [512, 445], [512, 432], [512, 422], [512, 413]]\nD: [[493, 421], [493, 430], [493, 440], [493, 453], [493, 467]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/a05c6e1fc084f2c02448041fed5d2848f797979541b7617a3dcd203089276c02.jpg", "reference_inputs/images/2bec413c323574edc50994e6be33fd04e6dd8462cd85f8440d3597eba943747a.jpg", "reference_inputs/images/0cff59f867fe7b86097ed51822007c1913efb6a96246b1483f058af263a31611.jpg", "reference_inputs/images/92d0180982628f8a0f6c22980f4af4c90172b66fa59a007b5a04f60b2a1a9574.jpg", "reference_inputs/images/7a0bf23ec2ab312ed51cb9da293f10c6cf147291121dfec77e840170a2766309.jpg", "reference_inputs/images/29180798b6def462313b800f7cbd24174dba0446ce21c69b7e0ad3d3e8083450.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000008", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[453, 548], [486, 517], [516, 492], [540, 471], [559, 454]]\nB: [[491, 462], [491, 479], [491, 500], [491, 525], [491, 556]]\nC: [[473, 548], [473, 517], [473, 492], [473, 471], [473, 454]]\nD: [[512, 548], [512, 517], [512, 492], [512, 471], [512, 454]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/2a2b24ce924573f78bae10b41de1a97a9d7bbfda58f3f13b6de81b92595c916f.jpg", "reference_inputs/images/e7b7f59c7b63eefb38a9ac6bdee663ae7f5f7b8d38b61ca7f1812b245bb861f7.jpg", "reference_inputs/images/0e5ff0ab8d31164aded8f8513ac5bd67763baf668e9c2ee63493117ca582e2ce.jpg", "reference_inputs/images/d21d3c1a0a692a4aef11283757828e28aaa3d07faaceec7cdb53bce912de182e.jpg", "reference_inputs/images/c0404cac926fbf07bdbd5a17eaaf584ef3d5ae071613235c60478b9ec5232723.jpg", "reference_inputs/images/7e335e03c825200cce7a9e42c6bbfaa0e960ed5feec24233d18da53e6340c06c.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000009", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[624, 446], [662, 449], [698, 453], [739, 458], [754, 454]]\nB: [[742, 467], [727, 470], [686, 466], [650, 462], [612, 458]]\nC: [[745, 425], [730, 428], [690, 423], [653, 419], [615, 416]]\nD: [[741, 489], [729, 455], [690, 418], [655, 388], [618, 365]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/6a53cb19284675c6c5354bc6c6bd250a20bbdccc3923d22bef0751c01f751a7e.jpg", "reference_inputs/images/d0a52db1dbeb14c97e2b3db46d25cc05110379b875e07af221f7ad0ddc7a3a59.jpg", "reference_inputs/images/ac9a067321586244d3dae504d5e0938ffa4068e23f67bbdf00ea1667bfd1597d.jpg", "reference_inputs/images/b8a086918400b04805c6ab807ac91ad2108f05ac9d2379c90888138768a44312.jpg", "reference_inputs/images/3287c0cde9d06973f68e05609776faa670ad090c0dde10bcec003ba20bdbc5b3.jpg", "reference_inputs/images/5130830fdc667694537466c1b0917d92aa254c90c62b32fe6253654bcdf7e859.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000010", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[246, 517], [260, 492], [270, 474], [279, 458], [287, 443]]\nB: [[266, 441], [258, 455], [250, 471], [239, 490], [226, 514]]\nC: [[212, 498], [226, 473], [236, 454], [245, 438], [253, 424]]\nD: [[281, 536], [294, 511], [305, 493], [314, 477], [322, 462]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/9350f6e7776f5d668a325e2aa8b02c45ca142ce9e54713d403ee50178fde93ef.jpg", "reference_inputs/images/ef1c9e15657d46d314b74cd595a8bf5a6aa2686d5fedec13ec8318fc99ba7f6c.jpg", "reference_inputs/images/579e066ff20752ad4dba9dc815b3ebad186cbafdfdbecc0c13b0cd4a53f43534.jpg", "reference_inputs/images/700b82317f7b7799736a2401b76a6c47038f11fbd151a0c02396195b2dcfde77.jpg", "reference_inputs/images/eac1cd823bc72c2025536b3f288cdbf50e6e36912faf8c456d836dac352b77b0.jpg", "reference_inputs/images/b537d27ae67f2abe24e4fc75cba67f2e56273e13ed72686e305fec5a9130eb96.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000011", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[473, 558], [473, 526], [473, 498], [473, 476], [473, 459]]\nB: [[512, 558], [512, 526], [512, 498], [512, 476], [512, 459]]\nC: [[551, 558], [551, 526], [551, 498], [551, 476], [551, 459]]\nD: [[453, 558], [486, 526], [516, 498], [540, 476], [559, 459]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/a05c6e1fc084f2c02448041fed5d2848f797979541b7617a3dcd203089276c02.jpg", "reference_inputs/images/a64a4eeffc4e041cfee0750c7ca09cfaaa10ab0dd50029a956576eef779531a2.jpg", "reference_inputs/images/209bd4735a24f658dec24cceca9171d144dbbabcfe4542d281051a67651e5bf0.jpg", "reference_inputs/images/4ac750855081052d67777616342037ffb929eb82e8c5482dd32135476a09a272.jpg", "reference_inputs/images/368df42215c11cc923177df9dbd1e6976d129ac20bea2499186d182c89dabf62.jpg", "reference_inputs/images/991ccec01d07275f54ef0c23d1fec3855790da8a1c97e1745754d92509cfb8b9.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000012", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[581, 540], [598, 500], [614, 466], [626, 440], [635, 416]]\nB: [[667, 494], [654, 470], [644, 450], [636, 434], [628, 420]]\nC: [[598, 530], [586, 506], [575, 486], [567, 471], [559, 457]]\nD: [[633, 512], [620, 488], [610, 468], [602, 453], [594, 438]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/8c2e4db7b2a0a7c44e636216d8c794d5940ffd1167c0c8cef57ac48a1f425037.jpg", "reference_inputs/images/05c8df436da91bf7e804433e570a7e078d64c9b41c40e31b46ac71eb9d7ad0ff.jpg", "reference_inputs/images/1d84a9a0b9eae92903a90403b8e36ed93169b339aaf2332fd02ab0937159618e.jpg", "reference_inputs/images/ed2797fcf780f0d9265d28800e0c45d9f1ba10a1346d31bd5cf51b29afd10787.jpg", "reference_inputs/images/82352d0ec68cbcc276f89b6e3957823cbca981552fe094fbc81873bd4a89bdfe.jpg", "reference_inputs/images/8a27f6f27475bd1f64e42fd6b54dfb6a99b3e314601ab91db35ee291f67bcc1f.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000013", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[512, 472], [512, 455], [512, 442], [512, 430], [512, 419]]\nB: [[493, 427], [493, 438], [493, 450], [493, 463], [493, 480]]\nC: [[551, 472], [551, 455], [551, 442], [551, 430], [551, 419]]\nD: [[473, 472], [473, 455], [473, 442], [473, 430], [473, 419]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/31682db363b46092009c46310f7c94d910001b83360c2ae0fc04cee7908e7485.jpg", "reference_inputs/images/87c8c0fa878d99e3199f221c3218b81775105d0a3f4a531a399201e685b8bb72.jpg", "reference_inputs/images/d6e149454f9363282bab06cfbbca15cfbb01c5944ba3526f28baff43fa8ffcfb.jpg", "reference_inputs/images/05c8df436da91bf7e804433e570a7e078d64c9b41c40e31b46ac71eb9d7ad0ff.jpg", "reference_inputs/images/1d84a9a0b9eae92903a90403b8e36ed93169b339aaf2332fd02ab0937159618e.jpg", "reference_inputs/images/ed2797fcf780f0d9265d28800e0c45d9f1ba10a1346d31bd5cf51b29afd10787.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000014", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[551, 458], [551, 443], [551, 431], [551, 421], [551, 412]]\nB: [[512, 458], [512, 443], [512, 431], [512, 421], [512, 412]]\nC: [[493, 420], [493, 429], [493, 439], [493, 451], [493, 466]]\nD: [[453, 458], [486, 443], [516, 431], [540, 421], [559, 412]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/4b34f4cb71a2d71d19dd3baacb4efbffba03c9ad7a50932e6b3432ed3969757b.jpg", "reference_inputs/images/d8b90566b8858c821792263f232b94b97cfbbcdd36a917c9d73fc911fd7760da.jpg", "reference_inputs/images/5c813a42655a831c1e9be223a0ce85cc187b01a741ae67b1a073bee29034fea9.jpg", "reference_inputs/images/30260f4557385cfaffdbff2134fbf3283819655eb5e46a4ff757dbb097a74244.jpg", "reference_inputs/images/b9fd7c5ca9dc4373d3f3da922e6b43fb529998bc3bbd700ce86ad3c69a4b8540.jpg", "reference_inputs/images/7b93f66815b4158d5b0fc541b3bab05a50fe7688b77c59e8fdbcad8a16266806.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000015", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[473, 515], [473, 490], [473, 470], [473, 454], [473, 441]]\nB: [[512, 515], [512, 490], [512, 470], [512, 454], [512, 441]]\nC: [[551, 515], [551, 490], [551, 470], [551, 454], [551, 441]]\nD: [[453, 515], [486, 490], [516, 470], [540, 454], [559, 441]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/8c2e4db7b2a0a7c44e636216d8c794d5940ffd1167c0c8cef57ac48a1f425037.jpg", "reference_inputs/images/b7cce4d0c6e9e1f8e3bbdf264066a794ae08968e56c63157c9245e38f9a67787.jpg", "reference_inputs/images/c63e1ff48fec679249f03adce2581fb770f181347ffa8ef2f6f887c9599ab6f5.jpg", "reference_inputs/images/84fa91bee78795f7087343d63acb876160d6be00b2cf4dafc8f6d77db679e763.jpg", "reference_inputs/images/6b5df6351b777266ba97aa5ce4ccfc7c33b5fafa8409b33b3fc6151bfd4b616a.jpg", "reference_inputs/images/cbf42e21aa5e52d5b279acc05c3f5cee3c30e56eea52a6a51464d271a4771a84.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000016", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[453, 524], [486, 498], [516, 477], [540, 458], [559, 444]]\nB: [[512, 524], [512, 498], [512, 477], [512, 458], [512, 444]]\nC: [[493, 452], [493, 466], [493, 485], [493, 506], [493, 532]]\nD: [[551, 524], [551, 498], [551, 477], [551, 458], [551, 444]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/4eb701386ccf91f87d8153ba6786d69689da1ca369180d1c62781fc6f8b641d2.jpg", "reference_inputs/images/368df42215c11cc923177df9dbd1e6976d129ac20bea2499186d182c89dabf62.jpg", "reference_inputs/images/991ccec01d07275f54ef0c23d1fec3855790da8a1c97e1745754d92509cfb8b9.jpg", "reference_inputs/images/d7773a87bfc65223ef136cddb83885c4a9ac2237e41809e937b5e43cb65c17c2.jpg", "reference_inputs/images/8c822065e6a4a3137d25a6abc18bb3c2c267e21f7d088cad1a3a17dbee7b5c2a.jpg", "reference_inputs/images/9454bfbf127a6ef586c8cbd21a88516a04c24bf9b4a4d2d70da2bdcf1d16e104.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000017", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[453, 498], [486, 476], [516, 458], [540, 444], [559, 432]]\nB: [[493, 440], [493, 452], [493, 466], [493, 484], [493, 506]]\nC: [[512, 498], [512, 476], [512, 458], [512, 444], [512, 432]]\nD: [[551, 498], [551, 476], [551, 458], [551, 444], [551, 432]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/feaa3f8dbc2662c912c5f082c75476fe48efb114906e2ca0a613e8a534b79462.jpg", "reference_inputs/images/8d1edcc8b4f0ac34deaa785673c23e6570cfe67f401af03ad7373b196ee902b3.jpg", "reference_inputs/images/be43c524ae2c0a4f5cba49ce67e3a33f7ca881c5b79e6c1eb61e6f8372beb563.jpg", "reference_inputs/images/cb10762cc79fb403ed3d060779fd177fdbc256d05c9dc19842a408604ffd0524.jpg", "reference_inputs/images/d5086606c7e6c9862835a0ca1ce281e192ed5d712cdffd431b2c73a4c3e1e998.jpg", "reference_inputs/images/a64a4eeffc4e041cfee0750c7ca09cfaaa10ab0dd50029a956576eef779531a2.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000018", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[453, 498], [486, 475], [516, 458], [540, 444], [559, 432]]\nB: [[512, 498], [512, 475], [512, 458], [512, 444], [512, 432]]\nC: [[493, 440], [493, 452], [493, 466], [493, 483], [493, 506]]\nD: [[551, 498], [551, 475], [551, 458], [551, 444], [551, 432]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/9b77a6fdcbcd164e42d030afdba126226416d464d2bb4b234b73eec459bc4446.jpg", "reference_inputs/images/f51cdc4fdab962d37e5494600db16c9ce68b3079d84ae214af152e6704b497ed.jpg", "reference_inputs/images/702dd7ba3d6815564455245f3199e57483d587890f24e75a7c679226bfe87823.jpg", "reference_inputs/images/8ed404085ac069432ca6885fa46e77d31c1a9cee698a07ec666853280f8b6fd0.jpg", "reference_inputs/images/153c42acc503b7f49f9717f56313e3f98e87cf0cd70405ca91151993eafe3310.jpg", "reference_inputs/images/b28aa8aaad129d53f280c819325ac5c973dbd9c05b862f43953b1c6b3a2077d7.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000019", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[512, 557], [512, 523], [512, 496], [512, 475], [512, 458]]\nB: [[551, 557], [551, 523], [551, 496], [551, 475], [551, 458]]\nC: [[490, 466], [490, 483], [490, 504], [490, 531], [490, 565]]\nD: [[453, 557], [486, 523], [516, 496], [540, 475], [559, 458]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/f04de421039d64e117485e8dac560c6fdb0cec9c635852de6eb669771708748b.jpg", "reference_inputs/images/500edaaec00045b906687829a7605eacd5a56ab0f8af07328a64d37e94f2d3a0.jpg", "reference_inputs/images/c2a80ccfa48ade1ab639d3d891b2f0f29367a884e21c08d68ef75dfe0b481384.jpg", "reference_inputs/images/490f8a17d13d06fa9ca15c673e75dc505e5e539818c26c6cf294a356ceaa4a84.jpg", "reference_inputs/images/8e90901750b16872fdaa935ea78ac05d94b56c4f0c1df5921fade7be28e20261.jpg", "reference_inputs/images/4d8f43a0b660056232a02393f2057c66b821fe369a010fb53662c8cce8cbb763.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000020", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[473, 546], [473, 516], [473, 491], [473, 471], [473, 454]]\nB: [[492, 462], [492, 479], [492, 499], [492, 524], [492, 554]]\nC: [[512, 546], [512, 516], [512, 491], [512, 471], [512, 454]]\nD: [[551, 546], [551, 516], [551, 491], [551, 471], [551, 454]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/3af649bc1ded32eb4c23a0be47a0841d61cf1666558c6fd8369d70e79125362d.jpg", "reference_inputs/images/778acd66ce6b79ac291b2d9c2a18217f97c9cf2ed7c14781bd1bcf9246f8b465.jpg", "reference_inputs/images/eb9def764525c9bf1b1f3228bdf3015051c1b352d9a60cef5c6a46bbb4e73500.jpg", "reference_inputs/images/ee5f53244b68d895548d299b14a24319f0482734af72d321c9198ba5b830dcaf.jpg", "reference_inputs/images/462d3dff450dfd5c907f20d1d934b7066d1a5c8b3121c61e4aef9b8e13af0603.jpg", "reference_inputs/images/8f0e184da0c5396462d9803ef969a439f6cc3a75d2dd124ddf0e54e8cb05b6a0.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000021", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[335, 535], [352, 503], [366, 478], [377, 456], [386, 439]]\nB: [[395, 454], [386, 471], [375, 493], [362, 518], [345, 550]]\nC: [[318, 526], [364, 510], [404, 498], [436, 487], [462, 480]]\nD: [[370, 554], [386, 522], [400, 496], [411, 474], [420, 458]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/58535b94fbff0387d3d936599a30966b8283f1f79f82b2557b1e6cb15cb992b7.jpg", "reference_inputs/images/c63e1ff48fec679249f03adce2581fb770f181347ffa8ef2f6f887c9599ab6f5.jpg", "reference_inputs/images/84fa91bee78795f7087343d63acb876160d6be00b2cf4dafc8f6d77db679e763.jpg", "reference_inputs/images/6b5df6351b777266ba97aa5ce4ccfc7c33b5fafa8409b33b3fc6151bfd4b616a.jpg", "reference_inputs/images/cbf42e21aa5e52d5b279acc05c3f5cee3c30e56eea52a6a51464d271a4771a84.jpg", "reference_inputs/images/0270e6c49c8feef3c5b6875c7c99e75229547dceb165bee3d8a659b59ba4fe54.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000022", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[629, 421], [622, 409], [618, 399], [613, 390], [609, 383]]\nB: [[594, 439], [588, 427], [583, 418], [578, 409], [574, 402]]\nC: [[562, 418], [566, 425], [570, 434], [575, 443], [582, 455]]\nD: [[542, 467], [566, 439], [587, 415], [603, 396], [616, 379]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/3af649bc1ded32eb4c23a0be47a0841d61cf1666558c6fd8369d70e79125362d.jpg", "reference_inputs/images/49a8d354e7a01418072c7c697722506dd87cf39e7bd223bd9a07d8e47d694c7a.jpg", "reference_inputs/images/fdf2d889c2dffa46fe69ae36b3d498a5696e384b59b17da14b65c985f03e64a3.jpg", "reference_inputs/images/90ce09656a6a167da5f7fab284aee25ea2185e3afc014d9222e307ef4bb41141.jpg", "reference_inputs/images/570950fea3c5b9a9401fd164a88933b77a227eb1e8d7870537c987807cdfe0c0.jpg", "reference_inputs/images/a52813cb1bbfce665d080a5bd070465afd330ce7de389c78579a9e73d135e029.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000023", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[784, 519], [784, 497], [784, 478], [784, 462], [784, 449]]\nB: [[765, 457], [765, 470], [765, 486], [765, 505], [765, 527]]\nC: [[823, 519], [823, 497], [823, 478], [823, 462], [823, 449]]\nD: [[725, 519], [758, 497], [788, 478], [812, 462], [831, 449]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/f459bf256a51aa8a9babcda85af13e41bdfe56751b2ca2d2cae1fc8e8bcc5e15.jpg", "reference_inputs/images/08ee00872b2731bf28cb8dbb443a0330eddd3b81a22f85113e27ab6b98a52c7d.jpg", "reference_inputs/images/d67941aa72f5f3e082a7058c681bc793e0dcb1d9e747f8cc380d1b3aa1237006.jpg", "reference_inputs/images/5d207e5b65ca0222da15a2a10ef778a54ce9b33258b6c7575b95613a8cdd2fae.jpg", "reference_inputs/images/46059ac09bb5fd2bf734cc65eee12ca51f0666ff729ffedd50c5a90cfb0a5479.jpg", "reference_inputs/images/f8c7f0323e4255a4d38b8b20b491242f51b57176230d029df64e4db379c6a105.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000024", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000024", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[407, 442], [400, 455], [390, 473], [379, 494], [366, 519]]\nB: [[335, 494], [378, 483], [415, 477], [446, 470], [470, 466]]\nC: [[353, 502], [366, 477], [377, 456], [386, 438], [394, 425]]\nD: [[387, 521], [400, 495], [411, 474], [421, 457], [428, 443]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/532a7bb18fb47c2c7a6afad29eee2e22fe5703751da3632aa370b49a67654041.jpg", "reference_inputs/images/8495e0a6e823168148c752bc98323fbbb0310bc13ba7e889b83d5fa2332d8989.jpg", "reference_inputs/images/58ada6f3bd0b224329f107337fcf267ddbbfba0aa96a7e436680b6f7e41541d5.jpg", "reference_inputs/images/7cad57175d82d4f201ed3adc2e233f500c5b0f78d029ee33ac9fc179addfe37d.jpg", "reference_inputs/images/ffeabae965ca87a9a76fa6420a55de21d573a30f19c479c2e2f9f7fc91d2e546.jpg", "reference_inputs/images/a1551f96b3d25ba9ed2f87e863aae600d24c590493c03bc33b735a099462156f.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000025", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[512, 458], [512, 443], [512, 431], [512, 421], [512, 412]]\nB: [[551, 458], [551, 443], [551, 431], [551, 421], [551, 412]]\nC: [[493, 420], [493, 429], [493, 439], [493, 451], [493, 466]]\nD: [[453, 458], [486, 443], [516, 431], [540, 421], [559, 412]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/c9f7c7f6f0afa98fff06c666a9b198933b04106cbece6ddf36d410a8c11d1674.jpg", "reference_inputs/images/5da55d9a67c0398174a73b7c502ba42ae65f2b8981379fd7b7b6fb3d720d9bf2.jpg", "reference_inputs/images/90ded380178ee48995f1ae42f5000217afa8d28b717d39d22a2dea889bf2c026.jpg", "reference_inputs/images/e725185d137f557ca071a8e8d091ddcdd4b5a0ec63ab624c26e1d69739cc434e.jpg", "reference_inputs/images/af4b4b12b6990b9197b41e2fce0d52e8a3d2d083ff1a31cf5b3c6ed0e4e29190.jpg", "reference_inputs/images/8d1edcc8b4f0ac34deaa785673c23e6570cfe67f401af03ad7373b196ee902b3.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000026", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[512, 548], [512, 517], [512, 491], [512, 471], [512, 455]]\nB: [[453, 548], [486, 517], [516, 491], [540, 471], [559, 455]]\nC: [[491, 463], [491, 479], [491, 499], [491, 525], [491, 556]]\nD: [[473, 548], [473, 517], [473, 491], [473, 471], [473, 455]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/fc020544619452bceeb4c01d861680999eff0a57983a98dd5efea5e57ebef25d.jpg", "reference_inputs/images/852baac5526a6621171f2c610f1452b58769ff01607d0541ea0b62c5c826c2fe.jpg", "reference_inputs/images/317ab9b52b0f25ddefae5a487bf7d0fe2d07557b135018f3c808b09f864b4bca.jpg", "reference_inputs/images/30142c91b4a377fd8bb01708b286c070258fd5c4a84e7771a9bec6c4b1def4c8.jpg", "reference_inputs/images/7c4f0d9c44a8609aa39e413ab375dbd6030efec0819d7b0772e76f733400349a.jpg", "reference_inputs/images/2969ea4fa30a89a9dec3fd0ec29a8b7a41dd1fbfc3dfc6a0028bc779a0abcff1.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000027", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[398, 542], [380, 514], [365, 493], [352, 475], [342, 460]]\nB: [[329, 479], [338, 494], [351, 512], [366, 534], [385, 561]]\nC: [[430, 520], [412, 493], [397, 471], [384, 454], [374, 438]]\nD: [[366, 563], [348, 536], [333, 514], [320, 497], [310, 482]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/1b51b7b6b84ff1a9d8e198eaf5e87d3a7a4fa6656b294051d6264f679cc47f33.jpg", "reference_inputs/images/8d920b98306faae1dd570d511d985769d0571056126f4c467bc883416cf04161.jpg", "reference_inputs/images/cfd7002de6019ef0fb8c2c3841d17a36f430d07616395ad65086aea03ef16598.jpg", "reference_inputs/images/bc60445ea52425ada75540534387f46245b9f63fd8a70fd45360dfd726a0f883.jpg", "reference_inputs/images/c1fb896aae3912e6c5790383435d3e0aba6876c0dfa6ce3ef5a0d93ba3be3375.jpg", "reference_inputs/images/839e7480594c685cc455a57a7f79881f7053bafcfd91d1204a3a1244770b77d3.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000028", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[493, 440], [493, 453], [493, 467], [493, 485], [493, 506]]\nB: [[453, 498], [486, 477], [516, 459], [540, 445], [559, 432]]\nC: [[551, 498], [551, 477], [551, 459], [551, 445], [551, 432]]\nD: [[512, 498], [512, 477], [512, 459], [512, 445], [512, 432]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/c391ab20637d039794ce863c1b9db7c92b6d6325ce8db865c9cab7217923f7de.jpg", "reference_inputs/images/e6053d8c9723f6ff332562929bf195f0ddd6f261290aa0c325ea27bff1afcec7.jpg", "reference_inputs/images/78fe1bce8137d790e4412107e620d82758be4b98df99fea12ae0bcf5de56d8af.jpg", "reference_inputs/images/cbdd1a6ed8b5b494d10165aa541a404c08ab908f6e9e23b313eeebf05ec1f2b2.jpg", "reference_inputs/images/99844ffd0082a8c59921a0d522eed8757940efe1d74aca3a8fa296e7efc3337a.jpg", "reference_inputs/images/0f90ddd60f46164f884041a99a443c3b40d81a26059b3783928740213b5fc2f6.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000029", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[464, 450], [470, 439], [474, 430], [478, 422], [482, 415]]\nB: [[430, 431], [435, 421], [440, 411], [444, 403], [448, 397]]\nC: [[378, 403], [413, 409], [444, 414], [469, 417], [490, 419]]\nD: [[427, 394], [423, 401], [419, 409], [414, 418], [409, 429]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/ac8940cb28f0503c8fcc59639fb6547b5c33a53516bb8ee49d1fe78cf701053a.jpg", "reference_inputs/images/94ad25a01e79c6f2579f8ee6c6169e22e305ebca76a4a24e2a23c64949e9cb50.jpg", "reference_inputs/images/9aad48e0f370f5a3bbc727932acf204f3be58bf1e9515df50d9ed8abbe4eeebd.jpg", "reference_inputs/images/eacaf1a98a9dbc744a26038fdc19ad727532b4a998c7823ded6c5b202ae5e67a.jpg", "reference_inputs/images/bfc1fb5fe7cb11654b62cb5be6043c83f36ee5eba7f398484c2053fbf88b30c7.jpg", "reference_inputs/images/09877c19121d39e605beb372f1c53f06d994bf324b26e6c61a76ac2eaa76c408.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000030", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[548, 436], [544, 427], [540, 420], [536, 413], [533, 407]]\nB: [[617, 399], [613, 390], [609, 383], [605, 376], [602, 370]]\nC: [[530, 445], [556, 421], [578, 399], [595, 382], [610, 366]]\nD: [[582, 418], [578, 409], [574, 402], [570, 394], [567, 389]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/07ca824b99b314bd94d57eda78a95cebe85ce30bc8e7d8698d800d83e704a7c3.jpg", "reference_inputs/images/d3558a936a65613cbbec820d11b51cc2ac958af87155f8146abf0217b9691467.jpg", "reference_inputs/images/852baac5526a6621171f2c610f1452b58769ff01607d0541ea0b62c5c826c2fe.jpg", "reference_inputs/images/317ab9b52b0f25ddefae5a487bf7d0fe2d07557b135018f3c808b09f864b4bca.jpg", "reference_inputs/images/30142c91b4a377fd8bb01708b286c070258fd5c4a84e7771a9bec6c4b1def4c8.jpg", "reference_inputs/images/7c4f0d9c44a8609aa39e413ab375dbd6030efec0819d7b0772e76f733400349a.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000031", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000031", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[493, 410], [493, 418], [493, 426], [493, 437], [493, 448]]\nB: [[512, 440], [512, 429], [512, 418], [512, 410], [512, 402]]\nC: [[551, 440], [551, 429], [551, 418], [551, 410], [551, 402]]\nD: [[453, 440], [486, 429], [516, 418], [540, 410], [559, 402]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/685f494fe9e8e65491aa89b5ade5db7f11ed63b16af3cd19c09d9acd7a59edcf.jpg", "reference_inputs/images/a1551f96b3d25ba9ed2f87e863aae600d24c590493c03bc33b735a099462156f.jpg", "reference_inputs/images/419c8e8c0256f002c06ea832dc2d2986c4c4ccb63acb18dc519acc5e537e18a2.jpg", "reference_inputs/images/b7cce4d0c6e9e1f8e3bbdf264066a794ae08968e56c63157c9245e38f9a67787.jpg", "reference_inputs/images/c63e1ff48fec679249f03adce2581fb770f181347ffa8ef2f6f887c9599ab6f5.jpg", "reference_inputs/images/84fa91bee78795f7087343d63acb876160d6be00b2cf4dafc8f6d77db679e763.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000032", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[607, 504], [607, 481], [607, 463], [607, 447], [601, 434]]\nB: [[582, 444], [589, 457], [589, 473], [589, 490], [589, 514]]\nC: [[646, 501], [646, 478], [646, 460], [646, 444], [639, 431]]\nD: [[549, 510], [582, 483], [611, 463], [635, 445], [648, 430]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/900db2df02c613ebaadf403651e63b99611b05b7ead8dd4a1a13b9aa5b1dc6ee.jpg", "reference_inputs/images/9261f3e23e13073afd8113384d260f6108fd5fdc99a21920c3122c549b87b0a3.jpg", "reference_inputs/images/79371f57f6869ad105d39af8e3ff76c5e4b73c2aedcc49bfe7aa2411d7ca6f69.jpg", "reference_inputs/images/f5e8542e9a2697a38ec46cf5fa5e9693e2b8983e5891bf504f56fd207bf1a334.jpg", "reference_inputs/images/e8614e06c5dc07c20662802cbb553ebfbdc10ce46bf9dc4f555dba7a3b4e696d.jpg", "reference_inputs/images/dc50adabea5c473de95da4c3e3bf5ddb25880f2dd43b3eb3b1e235dc6a75a193.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000033", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000033", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[551, 523], [551, 498], [551, 475], [551, 458], [551, 444]]\nB: [[473, 523], [473, 498], [473, 475], [473, 458], [473, 444]]\nC: [[512, 523], [512, 498], [512, 475], [512, 458], [512, 444]]\nD: [[453, 523], [486, 498], [516, 475], [540, 458], [559, 444]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/09e4d63e55bd5f22dd2cec7666efaac4ee4d83856b562dc1c9726bfdc0600797.jpg", "reference_inputs/images/153c42acc503b7f49f9717f56313e3f98e87cf0cd70405ca91151993eafe3310.jpg", "reference_inputs/images/b28aa8aaad129d53f280c819325ac5c973dbd9c05b862f43953b1c6b3a2077d7.jpg", "reference_inputs/images/71c1d97e3384db44f94c215ca7826cfdc15574bf16d173d6c6f24d4872a16d89.jpg", "reference_inputs/images/f732fef1023b152fb2d8586c31d138f9774e49c0669823a575ceb8d45f09768a.jpg", "reference_inputs/images/696db644fc0e650348f4be016e8e5fd9c181eee2de39d8b0f2f71ba4a8fde652.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000034", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000034", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[453, 458], [486, 444], [516, 432], [540, 421], [559, 412]]\nB: [[493, 420], [493, 429], [493, 440], [493, 452], [493, 466]]\nC: [[551, 458], [551, 444], [551, 432], [551, 421], [551, 412]]\nD: [[512, 458], [512, 444], [512, 432], [512, 421], [512, 412]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/4eb701386ccf91f87d8153ba6786d69689da1ca369180d1c62781fc6f8b641d2.jpg", "reference_inputs/images/92d0180982628f8a0f6c22980f4af4c90172b66fa59a007b5a04f60b2a1a9574.jpg", "reference_inputs/images/7a0bf23ec2ab312ed51cb9da293f10c6cf147291121dfec77e840170a2766309.jpg", "reference_inputs/images/29180798b6def462313b800f7cbd24174dba0446ce21c69b7e0ad3d3e8083450.jpg", "reference_inputs/images/0c8a9b920fb752d17854e01eaeabd648352c592e688d3546ef3830f74c8df359.jpg", "reference_inputs/images/cd76bcd55385041f93231333c73afa5c7080a078cd83c505f954450c93282456.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000035", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000035", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[458, 539], [489, 501], [516, 468], [530, 442], [533, 442]]\nB: [[476, 531], [476, 506], [476, 485], [469, 469], [453, 476]]\nC: [[512, 516], [512, 491], [512, 470], [505, 454], [489, 461]]\nD: [[474, 476], [490, 469], [498, 485], [498, 506], [498, 531]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/17b62a74f74c1ea060246e106301a4bc1105d82173d30d693bf2f9d1dd4a700a.jpg", "reference_inputs/images/717c6d354a7dd0b8acb74a631f7252328a333be184595a537f119fed42b30c57.jpg", "reference_inputs/images/955206a46841fe968052b35a8d73e9cfbccd99cf3f7df44d6930acea822d82bf.jpg", "reference_inputs/images/b83fb520d683368c4f8ac19874880aea135c6bab093c88367e8eabaae6585642.jpg", "reference_inputs/images/8a95beb66f899026d6ba71640e1e83d091440b50694db3dc6c1c48145d9f2ab6.jpg", "reference_inputs/images/f7a4b9bc95096798a69dc1dd32f1d6ccf8dadca03f0a88b331bfc9a7ae275be7.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000036", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000036", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[646, 501], [646, 478], [646, 460], [646, 444], [639, 431]]\nB: [[582, 444], [589, 457], [589, 473], [589, 490], [589, 514]]\nC: [[569, 507], [569, 484], [569, 466], [569, 450], [562, 438]]\nD: [[607, 504], [607, 481], [607, 463], [607, 447], [601, 434]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/900db2df02c613ebaadf403651e63b99611b05b7ead8dd4a1a13b9aa5b1dc6ee.jpg", "reference_inputs/images/9261f3e23e13073afd8113384d260f6108fd5fdc99a21920c3122c549b87b0a3.jpg", "reference_inputs/images/79371f57f6869ad105d39af8e3ff76c5e4b73c2aedcc49bfe7aa2411d7ca6f69.jpg", "reference_inputs/images/f5e8542e9a2697a38ec46cf5fa5e9693e2b8983e5891bf504f56fd207bf1a334.jpg", "reference_inputs/images/e8614e06c5dc07c20662802cbb553ebfbdc10ce46bf9dc4f555dba7a3b4e696d.jpg", "reference_inputs/images/dc50adabea5c473de95da4c3e3bf5ddb25880f2dd43b3eb3b1e235dc6a75a193.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "trajectory_reconstruction/town1_ground_aerial_v2/000037", "task_name": "trajectory_reconstruction", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/trajectory_reconstruction/town1_ground_aerial_v2/all_questions.json", "question_id": "000037", "input_profile": "mcq_text_reference_v1", "prompt": "Task: Trajectory Reconstruction\nYou are Aerial Agent, observing the scene from the aerial view.\nImage 1 is your raw aerial view.\nThe remaining ground-sequence images are provided individually in time order, one image per frame.\nThe ground sequence frames are labeled G1, G2, G3, G4, G5.\n- Image 2: frame G1\n- Image 3: frame G2\n- Image 4: frame G3\n- Image 5: frame G4\n- Image 6: frame G5\nSelect the candidate trajectory that best matches the true path of the ground-view sequence in your aerial view.\nUse scene layout, motion continuity, and the turning/progression trend across the ground sequence.\nThe candidate trajectories are given below as ordered point sequences on Image 1.\nEach point is an XY pixel coordinate on Image 1 after resizing, formatted as [x, y].\nThe order of points represents temporal order from the first ground frame to the last ground frame.\nCandidate trajectories:\nA: [[444, 403], [448, 397], [451, 390], [454, 385], [457, 380]]\nB: [[478, 422], [482, 415], [486, 409], [488, 403], [491, 398]]\nC: [[436, 378], [433, 382], [430, 388], [427, 394], [423, 401]]\nD: [[392, 374], [426, 385], [455, 393], [478, 398], [498, 403]]\nBenchmark mode instruction: respond with only one option letter from A, B, C, D.", "images": ["reference_inputs/images/ac8940cb28f0503c8fcc59639fb6547b5c33a53516bb8ee49d1fe78cf701053a.jpg", "reference_inputs/images/bfc1fb5fe7cb11654b62cb5be6043c83f36ee5eba7f398484c2053fbf88b30c7.jpg", "reference_inputs/images/09877c19121d39e605beb372f1c53f06d994bf324b26e6c61a76ac2eaa76c408.jpg", "reference_inputs/images/02fa92474473ebe9f1a1fda3ba193c9d869369b4b68692e22ea9a78bfa93953f.jpg", "reference_inputs/images/c4b4361d5a1279566dac8a9d342747059a39a8d9a68a07eb976dc2ee90f536c2.jpg", "reference_inputs/images/1e7b3515bbc18e19e874005bb680a85b86b524dbafd95cdbb9a85308de0ac895.jpg"], "image_roles": ["raw_aerial", "G1", "G2", "G3", "G4", "G5"], "image_sizes": [[1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000001", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fd68c43d41916a19c3434a0ed5d3049462d88713e192b18e8f51febc390cdc51.jpg", "reference_inputs/images/dbf22de41642b41b301aa73c4e3c022d051887a622da1c2f105e40cf171c91a8.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000003", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b0382be12d2c85e51697193d9193414c662b366062cf8c11b124a8efee5b4bf9.jpg", "reference_inputs/images/ebfa3e4e0b1e369bd197de8803340f89b0e019810fdd19183074acb8416c2de1.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000004", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c1c4c16c5c4f7375ec69c2f13d00eb7c2ce0de9fe9e8b6fabc682275254caa50.jpg", "reference_inputs/images/369d8c5a94fad57873a397e17382502d73f237070752a22f9e7d59767de6f1e1.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000011", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d6c7b56b517be5f9269a8f041fc150e7eb5f171fcdeee107c90b0307faf41b76.jpg", "reference_inputs/images/34de71c24a498b7940b9ebdbae28007cb1accc0e64a50685a45ce4c311e90779.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000012", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b84ea74379a0dc35ed4efa117f79ffd3eeacf5644d4810946141066cea215ba7.jpg", "reference_inputs/images/dd5c2d674513490b3a0ea625463703ba785dc554758dafe6efaa0c320a669a73.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000014", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/75bb9f3ac1921d7fed1cce6bc9bbfa9956c3988c4fcf366d0c757f6c48b81399.jpg", "reference_inputs/images/be2f4479c1d90ca348b9c4b183bbb4721f60985a5f6d83c0b7f33fda89db8133.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000015", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e6f52202e903210d25243e3735afe40fa075f6ed21c211af9dc0feaa58a7cb34.jpg", "reference_inputs/images/65bc64984b2063a131f48fe526cf6a7b07da1d78c38097687558bd599f26a609.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000025", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/16652bd8a0749e2b8903fd22753cff94fa7d4fe455aeb70555cc2ffe81a55955.jpg", "reference_inputs/images/12a5c1693e965c644508b6abab0f5fd0314a4bf1eadf2ef3d5612281e96bcd61.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000026", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6735f98cc46d273f973863ea4bee7b2468bdd114361efff6508e8164bd3036ee.jpg", "reference_inputs/images/cfd3ae4956d0d912e42e8b069a600698afb87d5f906b5079bc3f973feb58a839.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000028", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5b6ac385e3afda8add29752aadc7568dcfefbad224d4b7851478badc13cf2acf.jpg", "reference_inputs/images/4437282e019a268754e5fc5cc041febd6eafe5e9d08535e50d8a4d4d4a7a58c1.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000033", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000033", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6ff5e1786c8836850cecd1c27039e58dd261b509459acb33e3f5b3a7324081ee.jpg", "reference_inputs/images/37193519644ede3cffe2c34e62b72d4eca36138768a5b768ea8675d7056bb134.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000036", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000036", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fc2f455a3a134b67bdb86f8967c4e888682308e77817c4a1c0ede8e7b7ffb7cb.jpg", "reference_inputs/images/c1bd0271eb6687fc02d0d636a157471d657f873c3affd73cdb9593443aab7b57.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000038", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000038", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/170de95388c1beacd301d8b4deddad64ffae0d5a318f315340f7ecf49a4a5c07.jpg", "reference_inputs/images/c39ffc71392900cb722b2de9e5fc7e0bfa3f44928702031c7009afbfa7b4d8e9.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000039", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000039", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/37f3b24621705c5ca5ca35f0206472c1147c1f78bba9a690554c9eed51564071.jpg", "reference_inputs/images/7d3719bc63e8a194e6b5bade4d3c3bfbb3b34170c480ec70ad99fbee5cfcb1c4.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000042", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000042", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f2449548a2fb5afee173a59713a49688fa35c32ab2d5843906f2f9ba6baa0112.jpg", "reference_inputs/images/38e299805f5e8bc1db99be3cec22cb6a8251cd5209cea9da7e027f956e4fdc69.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000045", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000045", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4183a0b6de334d0e851eef8ba64a5005f51b443781882f4bdd69e0315019eef6.jpg", "reference_inputs/images/ee076bd854493700b9fab044dd65fda55b8ff5f246f0fb232ded9339d9b3fc6c.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000048", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000048", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/135d5add10ec73a8f8dccd1efdf99084b9cf438829994e55973c598984e0abbc.jpg", "reference_inputs/images/35cd9180f440e70c746afd7565e001662570370fa734d09ad27e609693f282e9.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000053", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000053", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0486102d6cf270083dfd4851360d3816fd05edcbea0624eaf6fd0c00659c604f.jpg", "reference_inputs/images/307a2d80ed2c534e0f1851bf19b95efdbdf842be63534ec427534bd9c7dba2b8.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000056", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000056", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cc432e0bbc20b1fa31132acb84df4fcdde4420d63cdec092f34992c12a82e331.jpg", "reference_inputs/images/ee568e58e8260c2a49a21cd116e1e6065a266da834f0bc842f073d0ee0e5da43.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000065", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000065", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0babe2b74ff3bef3cdac98263fe4524186cb1fd3c2bd927891a0e7a8de627389.jpg", "reference_inputs/images/efecc6ea655d2baf9c731021df4e0e26e6ddae097bd52949d1ab78c194d2649f.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000068", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000068", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/58b688227dea5c6ecb2553323fb0db535a0319fe0d4a271ebe130044271e3cc2.jpg", "reference_inputs/images/334b4e4d96aa2cc5e2ef412bf4ca996a78d03fc955f3ccf5b8987a427f0f5fb7.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000071", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000071", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bb3abf7b2abea3d7a9ec1ab8e3aeb474abd9a7c4128812c31c699c7d7397f122.jpg", "reference_inputs/images/dedfb06ec513842c3556edc6180d1cbc0ae0754bcd127e769bfb1729ff588f32.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000073", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000073", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b829ddf4c587e79a67ffe16fe25873fe19448bfe353db00f3602571b6511c0a3.jpg", "reference_inputs/images/85fcde214a5f739a3704430b0913bc6d212a709b9de608a20af5a8d063ae6d6c.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000078", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000078", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9e191a9c833a6d0191d5f681a95a3241003ac2ef139ea07eb111a9690c44cf99.jpg", "reference_inputs/images/38ccc557c2c7f36d3afb0b4ec64f79eadd0fd34a1ffbaab0a18d60eaf685afb4.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000080", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000080", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/37f3b24621705c5ca5ca35f0206472c1147c1f78bba9a690554c9eed51564071.jpg", "reference_inputs/images/9ed5bb3c857f6555223bb966ea696124ceaa9cef69ac9fb5efea95ff067acf2d.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000088", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000088", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cab1e1175c71d6cc3c6da6ded15543fe45d2443dcddeb393de29fd85636ab273.jpg", "reference_inputs/images/8fd402abf4edbff749a00e1626fa8df9dd2e0f2e68c3b40d39f362cc7f7a842a.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000091", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000091", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1b5c7d35d99b4d90fc9acbf22dad919f289e379585dda663df500ef2c2dbccf2.jpg", "reference_inputs/images/7abe8bf0e8316b5c50cf7a1af93b7def29fa22d68678887ccb5eace7bbf814b0.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000092", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000092", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/75270f18876035c28fbee5b811c3377171fab0e7f2d61776291fcbb87bae0a0e.jpg", "reference_inputs/images/f0ae3459f8df85f62392dfd5641283026356ce81720448a506ed43a090ec860d.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000093", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000093", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6f0f1e07119f53abe8f08c28c43962d83b3debe85de2c64a6e89e21ebb95a02c.jpg", "reference_inputs/images/90f04a6a2238a7fdcc466fd00f9d19ae60103ef575e4a7d3a7eff0223c6367b0.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000095", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000095", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/62bc1d89c2ab42acf8235ee12e229525f31983d1e1f0d596665befbb9a4745c8.jpg", "reference_inputs/images/7ca6fea9636e4e86dfe839da263dd1709a3a0380123f4d2fb0b156b04383084a.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000109", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000109", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f67eec63d5c0c61458bbae0766c70b18ecdbfa0cb0affc46531d9447a2b74645.jpg", "reference_inputs/images/4c4ce9a2a452eb6f170a53a1fc9424444112c57777f71b6b9064acc133562968.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000110", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000110", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1f14b718b819a632f51cf4ea6d5486cc08d1760dcb059aa48113fb4a35adc9bd.jpg", "reference_inputs/images/07ce94b08f3eb198a9dff4281ea74631308b0765bb341ac4c720a3b229986266.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000113", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000113", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f715261c31086ae963af6c95167737fcc81cedadfdf0f8460aaa5262d2a8969f.jpg", "reference_inputs/images/a60a62748a3f2e935195f4fb999c581d2c0846a7a84952173733ba07423749b3.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000114", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000114", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/175853c64e42bceeece3255191689fb55134b6455e8d10cacc3499e57298ce54.jpg", "reference_inputs/images/5a1fcce3da65f5c5d01ae776a7c8c7e6e18ee91e4769388a451c9516a32b16a7.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000118", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000118", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f54aa2a4519e2ad65a9712796f114abfbe2311e5d26c72e03401a6216aaee1ef.jpg", "reference_inputs/images/e0b524644f1d95620780c68536d47b596d7a83e60233ae80b64447070211558d.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000123", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000123", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ada7625d028d5cd99b2117a2492a1c060ccdf998585fddbbb5c1176c9319ee5f.jpg", "reference_inputs/images/a2c94439294592ac645c2eebdc4815dacd145ee16cf456363ed94e9886ac961c.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000124", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000124", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3339fa2656e9030dacd6cb1596214a8954c2f4a38264a3f937f2226a0cdfc610.jpg", "reference_inputs/images/e0381b03e29afa40e6d18828c0f36bc0d89bb5a74d329b1a9516c1a79e4b51a4.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000126", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000126", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6563f897b37584639b1afafefd1626392246384180788bd24bb9c2d4b1b198bf.jpg", "reference_inputs/images/797828bf03278f7cf8e37902680e04413d1fe417eed251e50e129b408b24664f.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000127", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000127", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f6c484dcb9b6020582b2c92fbdb472b4796aeaa6ee3b54cbc4c7fb823701544f.jpg", "reference_inputs/images/387eee192cc021d94a93d3ba27331b505992d6b714b2291e6d3f1b53665fad27.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000131", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000131", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b367fff5f8df9c6f1e989c908e780fda5357be69271af95101aedda35adbed4c.jpg", "reference_inputs/images/99c0c832d1351270dd4d4b11a136b2e7c88bc5c4b173804985164dde364a5a7d.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_20m/000136", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_20m/all_questions.json", "question_id": "000136", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5eacdfdf76480a1e03a3c33ef1ebad5f46356b2004a522db7938309e7211ccf7.jpg", "reference_inputs/images/ea61647a779ebee95a853b01b7e4f7fe6774eea3190000185f3ba72e9be63f7c.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000007", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/16652bd8a0749e2b8903fd22753cff94fa7d4fe455aeb70555cc2ffe81a55955.jpg", "reference_inputs/images/9387fb07c1ea73e8176f603aa4bc4630b33ca3cd16a48c5c1865d623fa341a4d.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000010", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c5b7c80344326eadf236c095b74caab02fec7f079c683b63408b21f0a4fb239d.jpg", "reference_inputs/images/2ebaf8e09a7ea4e210037cf5e4f92b83803cc3ee2ba5c2eb6ca9a8f58923c755.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000019", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cc1dde71a68532425702ecfea7de2b7bdd525f12e46722c7d019f842281e8165.jpg", "reference_inputs/images/e5cdc2ffbe31b0bf6dc8b74abe10ba04fd31c30e5a07a413559f73a37a56dd6e.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000021", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bcdc1e69c0a36fdd6f81b6dc138bac8948640e1f401b454a6715a6941bd5a4a5.jpg", "reference_inputs/images/4bbaf7b875b927803278ab0e3bbe127dc986d7e632e71250f020a2395eb683e5.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000022", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/aac69c69ff5f9d453bf6bb6074e31ff4cb816d0b707e3d1072b815e801adaf7e.jpg", "reference_inputs/images/4998766b8db5674941077b47c7b16cb6e42a6052f5423fc1280ac164d2f67fbc.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000027", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5323a1810ba62510b86632f6d1ab527f3c2b53f59ca191528180e760ddc72b2b.jpg", "reference_inputs/images/4bd2a4c71978d5a712b997487d72145417d2d58cc4dc3a1d87079cf74e6d3024.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000029", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6bde287cf09eccf13f1a7282d5f5bbecd74d66d92d3f8b16409fa292a063c87d.jpg", "reference_inputs/images/ca513ffc9b1c99c0f5ed3236a7524c984df0176f05fa92e5b3fe95b58184d6e6.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000030", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3e7a7afd976f26eb7daae205a58442e3eabfcc25a0f36e6a9a8e95d4a5c902b3.jpg", "reference_inputs/images/ab28713cac5e6c36e3c2b220a44bc443a3ded844c4683e71d7dad7d2fbca87d3.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000046", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000046", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/34d027dbab11397f568f524e93a1358bb23b7502da274b0c04a1dba07548b837.jpg", "reference_inputs/images/b93a5eea86d8ed6ee483ca1841644a860eed2f66bd6ad287b7ff0aa36b80dc9f.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000048", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000048", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d8cd18a659c88b8227d5914487e01aed64e57235ebab82de7adffed392de971e.jpg", "reference_inputs/images/090d77c98d567ec0e852cd40b7b62b5fa363743a0cb6c3e924bc7c6646849be3.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000054", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000054", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/29fc872b749e944ac734776f6892fe8e4600cf609cbd0a71f1ff2dda5f3dc4e4.jpg", "reference_inputs/images/e4d2d3e636b4d3b3ff2b02938c80dc6fca66ffd777ec8927b57b34db548cdb62.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000062", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000062", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/eec11c6f6b1d2709050048da2041d35cd94265ab26d33a33a26396132f37f44a.jpg", "reference_inputs/images/2d80bc60677dd8b2a63c877219f78968fdfbf756f5998945b085e9e6ae8f4dab.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000067", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000067", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0894d098956895b83d6bb6a0a385f9e9ae317e9b7b8e55a59a7025b77164db49.jpg", "reference_inputs/images/9943189216579157cb5669ed05e0a556e83eb992a1b488f4986a075136f58655.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000069", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000069", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/912001ea44deb2e90793bbecd6ab0407711e1c2fc6b14c92a7fff3b001b742e5.jpg", "reference_inputs/images/f193bb00ef247ef9d88a96146d29e7f7a175c10da3e12d859889d7018f720d73.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000075", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000075", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0b04d2c048d9ec3220836158a321c7d88654a550351b69a7abad236dadc214fa.jpg", "reference_inputs/images/1a395da19feab8dcb1c6ef71df4563096778af9dbb7f71b0efcc8acea44a97fd.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000078", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000078", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ee79ccb6790b2978c939ac1dffc6c1fb7f5d18399cc47dc5d3f82e9224410408.jpg", "reference_inputs/images/0e9350de4cca0d2038837e11399dedb6a9b8f68d57044155fcfa082cdea0672e.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000081", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000081", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0954fbaf2620c6f5b47a8b30ce0b61391a5fa9e332bb66020514f148e81de546.jpg", "reference_inputs/images/9a637268ccc66c4e35c4ed2e56daf521b49a57fefec9c91252cab14f54e236a3.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000085", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000085", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/37552d687aa2831c0c5d04c6d1b8d2a990c2a4233686be5406484478e674dfa2.jpg", "reference_inputs/images/4f8eaac7af7f13cb761612f8be0c0f453bd97567a8e5124a80eb708e992df585.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000087", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000087", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5323a1810ba62510b86632f6d1ab527f3c2b53f59ca191528180e760ddc72b2b.jpg", "reference_inputs/images/19a1b73f12288dcce7ae3ca0cbe6db93634d4b933c6efb82423b324d681afd14.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000096", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000096", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/75270f18876035c28fbee5b811c3377171fab0e7f2d61776291fcbb87bae0a0e.jpg", "reference_inputs/images/6478b0df46a3771a666dc8f7e2e1c2b6014023ef1ea6b3231153cf64f92614b3.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000097", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000097", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6f0f1e07119f53abe8f08c28c43962d83b3debe85de2c64a6e89e21ebb95a02c.jpg", "reference_inputs/images/43d5746cf60013f61d81966c0c6e49fe26429e21c2249a6021ac2a039deb0240.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000098", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000098", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0b04d2c048d9ec3220836158a321c7d88654a550351b69a7abad236dadc214fa.jpg", "reference_inputs/images/4884734f4435bb37db842686b2f0da57786ff2d95dd3fab2f84a75a57e1a72e1.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000099", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000099", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/29fc872b749e944ac734776f6892fe8e4600cf609cbd0a71f1ff2dda5f3dc4e4.jpg", "reference_inputs/images/0dc5e9dd4c1857374a67d4565eaa428149d3d5ceb0c6e075ca173641439171ec.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000103", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000103", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0d248986a749fc13bbbbd7f16f564db05fc709e1a398f8b7f179f2517a96214c.jpg", "reference_inputs/images/2d2ef6df3da25801c3ed7b3d82b7a254ad3494a2ffe5a16ce06e0580edce6853.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000111", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000111", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2a4ced4ecbda3f0ad9bfde9034be5e0a49ec9bab9c712884affd3ad08ecb94fb.jpg", "reference_inputs/images/7f2c138cee390f61dd6156dc869309b4f6ec85f8bd45cf4b301361e124a961be.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000120", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000120", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bb3abf7b2abea3d7a9ec1ab8e3aeb474abd9a7c4128812c31c699c7d7397f122.jpg", "reference_inputs/images/59a6d775da7d604eb34fbd25fb53bd037512667d4e39c4c6bdc48da38b7dd8ba.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000122", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000122", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/59e9ef9660a543204e89f2dcd6b0a303812122a15cfe2f208507ab32393d960b.jpg", "reference_inputs/images/7c4ebde7c865ca73b1dc3e4ffab10bf3f26db0315b2ab3af95c13b72d1e06e58.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_30m/000129", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_30m/all_questions.json", "question_id": "000129", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d12375d5e82700b64fab8fe32bb0439c4988961f226f86a61debff37c8aae019.jpg", "reference_inputs/images/f836ca059921239eb42024ba8423dd2007a5475b13086e3b198e74a73afe9fb5.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000001", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7ca2725add12eef4f5e680ff3811e70b33b603449326e30b171a499282aeaa45.jpg", "reference_inputs/images/d74e78d148ddc207f187b8c65684309c6422848dd1386141f62e3e8c2923de7d.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000002", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3968266896f044d15bad291d7598cf1d78a7370419cdf14edbc6a4a4ab52fa2c.jpg", "reference_inputs/images/c48c8f3dd8f09173d5255e08434d0646d30829d71101af174f6bcfe76d7c0499.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000005", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9b83d6129aa61cdfbf7ff383b9a29ddbba8ea0418bbd5a27c7cb21866300128a.jpg", "reference_inputs/images/da15a917ccedf6ae8c07672a212ea6dd2051a4fa963b240d071b49a08f112a12.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000006", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/93f8c8c3cf249ce2b1f1c01c90e4fd3362d4265bc084ad006f6dc4e0f549605f.jpg", "reference_inputs/images/b962f92900ad30a5894942eb6439f696bd15162afcd0132235e430965c681111.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000008", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b829ddf4c587e79a67ffe16fe25873fe19448bfe353db00f3602571b6511c0a3.jpg", "reference_inputs/images/e016680aefb4975fd585156c0476df68609701f820f03430ddd12cabb6f12fc3.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000010", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5940e2d6e51e07fc71ca2c19e8d33a5a3ee00569378163517bc52e26913430ab.jpg", "reference_inputs/images/e6ac568847fed700544b66b14c167167d1eee1a1f47aa9f545ae15d69b8fa152.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000011", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d8cd18a659c88b8227d5914487e01aed64e57235ebab82de7adffed392de971e.jpg", "reference_inputs/images/7bcf502d4516a23309f8c75d6165d8aec30a508806b688b2e0ca79943264fb03.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000013", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/776afb1738377b6edf284ddab5a114760509f3b1d0007907cadcebd6d89f34e5.jpg", "reference_inputs/images/882e7f6afe44c6c73a1c021aa9776d97e0d0ade90f3fb086b175cb64a95e66a7.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000014", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7c735bd6482ec417d7194d29f57872c7f9dfcdde111b1c37df18ad611819e036.jpg", "reference_inputs/images/68ca831a375dd3794996c6f4cbd058259ea3eedd7872e0312cbeea3a2ac6d907.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000015", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e8f1b9d03fe3b3b60044e05053b10d6c5713685735425b089f2fc508c67c0527.jpg", "reference_inputs/images/328da3101eea0b3ae63ad6574b7fbbc5314cfaf5506f98c4e67fa709b3f0c2d6.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000016", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0954fbaf2620c6f5b47a8b30ce0b61391a5fa9e332bb66020514f148e81de546.jpg", "reference_inputs/images/6d4b69271babc4b16c2b15393536cc0da931dde391770a9cf055420b674ec17f.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000017", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5b6ac385e3afda8add29752aadc7568dcfefbad224d4b7851478badc13cf2acf.jpg", "reference_inputs/images/048b340f01c30e27e6544c1eaced1f0f55fa86a9eac9bb35bb55b6a990851eaf.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000018", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/29fc872b749e944ac734776f6892fe8e4600cf609cbd0a71f1ff2dda5f3dc4e4.jpg", "reference_inputs/images/1c5457950384d1d8f40f4509497d99dfa6f81e9a2dbba70aa46cd97481aad06a.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000019", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9dda419c3bfb342e9f766eac02b43633a9db8704e47d2067503de3aebe5e004a.jpg", "reference_inputs/images/7e21132157647b3b4c12ed0ea965b2a977e1abf4def101ddb54755d0bc8ac83b.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000020", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/47f9e862659644a60fa5865a905e3616ae64a6438419abb7aae6d257478f5f7c.jpg", "reference_inputs/images/49ac4b66a522bc561a7a204f29c911c5be509340702cdfc96d04b553d517def1.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000023", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f17f0642177cfd4dabd289ea57ad315d5726afa8866c914631a2578bdb2e5167.jpg", "reference_inputs/images/fb21cd2b6d948a708e87f10abc4753f5d0cabdef37ac3f3047ed1a96d8e703a9.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000024", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000024", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1d5f9dec7756bf1faa672e44115577f7b1d0a59014a9a88e1d52979f20ad2a82.jpg", "reference_inputs/images/c7e819587b9f47c42a1fb39bfbab51a2148d4a1db54e01a307a83ea1c0de29b9.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000025", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c5b7c80344326eadf236c095b74caab02fec7f079c683b63408b21f0a4fb239d.jpg", "reference_inputs/images/74b31aee53bec746cfe75b95269b54832945cc93e1e41e332bd12f1614f13f1e.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000026", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6f0f1e07119f53abe8f08c28c43962d83b3debe85de2c64a6e89e21ebb95a02c.jpg", "reference_inputs/images/ec9ee0259f373dca2413d97a227e137f2436e9424f801502cc802d3aedfa473e.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000027", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bd083d00ac139e843cc29c66bcbca0f3a6b9e3304a4d97fdaee74e6fd9f212e8.jpg", "reference_inputs/images/f31555b0167352e8f2e0a952a39401692a59d6e9f8a476d52cc62814504573b1.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000028", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7f4ea77e613070389d83ea65a680648d61804ab4f4649a3a46415e73bc51a18a.jpg", "reference_inputs/images/4f572f335c2a8fe2dd6d0e3ee4134e5947d4f62a9a958e2eb9d9812427b49b1c.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/city6_ground_aerial_v2_50m/000029", "task_name": "view_localization", "dataset_name": "city6_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/city6_ground_aerial_v2_50m/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1d5f9dec7756bf1faa672e44115577f7b1d0a59014a9a88e1d52979f20ad2a82.jpg", "reference_inputs/images/aa5639329e5e4797037644436aafd5986c14d7cddc9e935fbd6e86c97590d097.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000002", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/317ab9b52b0f25ddefae5a487bf7d0fe2d07557b135018f3c808b09f864b4bca.jpg", "reference_inputs/images/32c5f93737b6221c27cddf6f96ed46ea4fbe2cb59fe06807ecf52eafeac8d3c0.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000004", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/08ad072f4beec4f109ee7613067d420465a71994083955ca5b6c3f3dce6308e6.jpg", "reference_inputs/images/532a7bb18fb47c2c7a6afad29eee2e22fe5703751da3632aa370b49a67654041.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000006", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2c4bc66211ccfe5085183b35111b64fa11a0bb4469dc650241a0294db4576f68.jpg", "reference_inputs/images/9f86aa4244ba0a6f10005b92de656d75d1fff6c3bde4f6e41da74851b25dbef3.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/town1_ground_aerial_v2/000007", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/af00de34a0a8156b25fa9cd1398e7cfd58c1674afe33b86509ac463390783e8b.jpg", "reference_inputs/images/fc9f3ab184f9d2360b7db17d33b2212cbc734213614595ef3be4095076e63226.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/town1_ground_aerial_v2/000008", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cd09321292f0788f78323c26abbf46a241c86ce743f184ba8aece6a49f533751.jpg", "reference_inputs/images/900db2df02c613ebaadf403651e63b99611b05b7ead8dd4a1a13b9aa5b1dc6ee.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_localization/town1_ground_aerial_v2/000009", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000009", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d11a038bedfedc7ff9ed235572e8f29187035b0e81e4c5a624212d405c977e5f.jpg", "reference_inputs/images/92668046913805a01a3557f79887ac3249b5bbcb0420b9195f26313580e9f7c9.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/town1_ground_aerial_v2/000013", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/490f8a17d13d06fa9ca15c673e75dc505e5e539818c26c6cf294a356ceaa4a84.jpg", "reference_inputs/images/4edc6aafa7d35d67daa9f41e6f1f66a25806c272325f634343c0e542173fe5f7.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/town1_ground_aerial_v2/000014", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4c030ac1994474c88f320c0a101f7b988a4282a18b3f7adb0be9b0a96bdc2f49.jpg", "reference_inputs/images/ac8940cb28f0503c8fcc59639fb6547b5c33a53516bb8ee49d1fe78cf701053a.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/town1_ground_aerial_v2/000026", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000026", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/17c36cd999f47770c35ff1f5ec06bd0b9c32857125716a6c51d4fed960fea06e.jpg", "reference_inputs/images/d9984a287ff5447a4720a69174962625e9d679ce734ad6a6ccd16e9ae8fa2a56.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/town1_ground_aerial_v2/000027", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d9fc7939013ad6c56897acd46ae1062ef8670cea62d52876078e1ad30ec58c43.jpg", "reference_inputs/images/31aecb76edc4859a56ef7861758fd7369720bf45ec14e2df2b9d49d05015cacb.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/town1_ground_aerial_v2/000028", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000028", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/52f75b3baf552b2d21ded19fbe97e159b0362e15c0030dcac1725e1841405e8b.jpg", "reference_inputs/images/538338aa2b161d435ae2192a2bacd363590d3e8137ac0c96fd5f92142d6493db.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/town1_ground_aerial_v2/000029", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/232b55bf4fd70fc275e40c17aa5786c94e4cdb3ba6adf0f17550b78704a74eda.jpg", "reference_inputs/images/cf5b61828dd09b354e09c239c919322d551543e0d0501ad256e236fe64e116ec.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000031", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000031", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cfd7002de6019ef0fb8c2c3841d17a36f430d07616395ad65086aea03ef16598.jpg", "reference_inputs/images/3d91c45a797db089579e59426df5d2694d63f16a52b3e2a02525e4ff3e3856ec.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000032", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/23d606339b44ef900e4f4b8fd734f19f809273819df9d7e7aac8b402b5896432.jpg", "reference_inputs/images/456aa37300709b21428b8ad663cd9319beebed0434d7ff1f2552726ef275f559.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/town1_ground_aerial_v2/000035", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000035", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/820be4106d714dd210aa72fc0d64056ca42bf12125609d506ce24ed90bf348af.jpg", "reference_inputs/images/c1fe12231e3c6505f4561f7b59d589c5fc02488b59c32ec3725a2de26e313288.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/town1_ground_aerial_v2/000036", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000036", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c1af4763b62bb1c4e18c6f2872b14ca427b73e8e71ef4b9b97e5add7b49f9e43.jpg", "reference_inputs/images/587a0fa2190c5ccaabe4fd2b59c4141418e3cfe67fdef9efdc88d01ed187075d.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/town1_ground_aerial_v2/000045", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000045", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/dc50adabea5c473de95da4c3e3bf5ddb25880f2dd43b3eb3b1e235dc6a75a193.jpg", "reference_inputs/images/45161aa0655e931e7924983dd2f4d431231d378aad482002b7e8acb1c3c8ed8a.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_localization/town1_ground_aerial_v2/000046", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000046", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/711ce8e2496c393beb47a7df2abb51a802dda0a39b8039b2f0be4ae01c465d8a.jpg", "reference_inputs/images/ecc4acc96adf7c3979c31ffb91951a9a19ba78253b731c04299d813779925993.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_localization/town1_ground_aerial_v2/000049", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000049", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/19a560adf2a1d5d370d42237b6bcfb0185312f1660c833c0bc1e4893af2f8cc7.jpg", "reference_inputs/images/342a8f6b2e4e2e3da2b004e4d40976ddc79752e75595f7d2723a8475621c2975.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000050", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000050", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/401c56d273fafd2810d07698e3dcd970bfac1b1b5a8c7b407198007ba0d7d67a.jpg", "reference_inputs/images/3c7c8c12910aa07db95a7b466fc31da7edf091f3d89c531448b5cd9b25826513.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/town1_ground_aerial_v2/000051", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000051", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f23bf32195a08f6f7a10706be401255d7b8be8ee1901532061d640a211f0ebd8.jpg", "reference_inputs/images/14ea2bc6950721a50cbe799744bfeb8e25310267ee07f2c07dcf584490493184.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000056", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000056", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4c030ac1994474c88f320c0a101f7b988a4282a18b3f7adb0be9b0a96bdc2f49.jpg", "reference_inputs/images/fb34b53b679d1237735335f190243505fad3e06e30ee7b8151f738312973ddd5.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/town1_ground_aerial_v2/000057", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000057", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/82352d0ec68cbcc276f89b6e3957823cbca981552fe094fbc81873bd4a89bdfe.jpg", "reference_inputs/images/7ff9a063ea7e900809d71200db9a76b02c6c06d8d46e1304fb4b04a2199158b9.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000059", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000059", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2c777fd51ad25bf1b32b22cce92ada8ec6c7e7dea1b045657058c317db0756f0.jpg", "reference_inputs/images/0671ae8ee169440e6627d92e84b2ba53644dd471a3dc9097fb5bbae57d101527.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000060", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000060", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0b1a37a656d3fb6c2b0b1eedc6c8bb0527fef397451d75847ff913fc5e5dcd54.jpg", "reference_inputs/images/60bd10c12f3b7523c7935fbe4cf64c148a3a898097e880c43e43d5e4d741ac41.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000068", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000068", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ae4d1bcde5555a2743e98a6971bd3ff58fc27085dd56b5a570a70204a4430c3f.jpg", "reference_inputs/images/4461bc21821aa5f114eb02886c31e811ba08f1d524f50e25467bb55c8ba3a808.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_localization/town1_ground_aerial_v2/000069", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000069", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/82352d0ec68cbcc276f89b6e3957823cbca981552fe094fbc81873bd4a89bdfe.jpg", "reference_inputs/images/0d72fb719c98a09f795afcf8941afcfbf1fb13fa4f2c2af3ef0c77f6ab2bda9e.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/town1_ground_aerial_v2/000077", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000077", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2de10aa040bcdf1ae947d053dc6f1971086bd5f033b6425794a494edb23eb40d.jpg", "reference_inputs/images/254a8008d5c1831bfe499fa77d542fe0723aa9c7f84c00870e1229494845aa3a.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000079", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000079", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1f7ce81c46828baad3113bd8ef4a0eaac9e9573ea01bcb9cadd31c1dafcae86b.jpg", "reference_inputs/images/f52074d663dd3124ec4f524062fa3051f9168c71a7cd5f8baab31bd55101c2d6.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/town1_ground_aerial_v2/000080", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000080", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1f7ce81c46828baad3113bd8ef4a0eaac9e9573ea01bcb9cadd31c1dafcae86b.jpg", "reference_inputs/images/141981d1304600d2b7202d32a0cceeb7ababc6271a8ddc2f8a4c0d2ef862a981.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/town1_ground_aerial_v2/000083", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000083", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/45524dbb516fc865ba0b9581a3965608ca58581fcf187acf0322a8e4b0663b1c.jpg", "reference_inputs/images/41821988d9f59431e844993ef642182ec744c51695610e657e3aa61b90f80b96.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/town1_ground_aerial_v2/000086", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000086", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/570aa2e1dcd552d6dd226263519545311ebc7a1e09b440e10a71bb754cf35d99.jpg", "reference_inputs/images/6a53cb19284675c6c5354bc6c6bd250a20bbdccc3923d22bef0751c01f751a7e.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_localization/town1_ground_aerial_v2/000087", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000087", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1ae237aad839b445a89cc0e8f1fdecf1e406c23485f9ed36443a2be1a3304a8c.jpg", "reference_inputs/images/f6ee8fcebd766adf15d2a51cb919078bd990a44b54169fcf0b1995cb5e9e9706.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/town1_ground_aerial_v2/000093", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000093", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/99996be75264845d962cdba0a1ed1039ac5656471bd459aca071179e1ee05317.jpg", "reference_inputs/images/28237e8c293ed50e766846b66ba45d969c4b97f8a36662c2ca6e532b338c989b.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/town1_ground_aerial_v2/000100", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000100", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5c67be9d3816bdac4d19867169d2ac85fa3ee42dc173288321467a7e9199fd18.jpg", "reference_inputs/images/ba9849bf6d9b1ccb29b7751c1299bb365c5c6df0cc7f3ff744dee12a076a7255.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/town1_ground_aerial_v2/000104", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000104", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2b9ebe127de558219779e74edbe2262829c28ac14fa229591b8245e5b1109e5c.jpg", "reference_inputs/images/d8e886ef8017e79d51b8e28b086972d80c36acca30155554ef89f001b4210f04.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000110", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000110", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b53fb9c0796915773d625f68391841a4fa0af07f4530e0c9b3769fb61d9089ae.jpg", "reference_inputs/images/3814847d25158f5ccaa4511249134621b0b048607f59da5f14c624ca40ff168a.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_localization/town1_ground_aerial_v2/000112", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000112", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7fc67fd1699d1ffea8cc7ec540bf7b5d6c4ce8d4e90cd08280e08ec29d956f95.jpg", "reference_inputs/images/2f77831cc86c65fef413762aff99a2d59078e856fc64823fb778e579485d5864.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000113", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000113", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a8825c3ea8aefe3e6a1da5a55aa7242f4c037b15dacd0ba6d173469aaf064d16.jpg", "reference_inputs/images/913b6e587a94082eaf302612259df56c4680a9c9280c99c8866de0038c2d9730.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_localization/town1_ground_aerial_v2/000118", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000118", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/90f44cd8773a291f27a102d5881790e7e81d55d55a2d817a0bdfbe890d1a4036.jpg", "reference_inputs/images/525c13b6e514a8ab386a16ba15cf0faaa60413d2eb6b298a03017301a4385643.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000119", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000119", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a613f0e862743d96f788c7b41c2e4abb6c48def73ada6194e7a8a319fa81b08e.jpg", "reference_inputs/images/676c851854c7cf7d16738d6c27c175fcb0eda69d01fbb83ca329815fa80119cd.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/town1_ground_aerial_v2/000129", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000129", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a75e375d0eb6d4151b86fe3e505b0c8d1603b45502213337387d22a1a023ad95.jpg", "reference_inputs/images/f0ba7d7e7df8b29ab831e513363cc3fe6a14ab9c3fbc17865adcf9798007292b.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000130", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000130", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2c57f3e448d39cce4b09504a4ae1620a90cfaa78e110b78269db0fbd1d03ddb7.jpg", "reference_inputs/images/282a126ad09da66758c5b345540b7a451cc3f0816e95fe4193922ca93a00dfc7.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/town1_ground_aerial_v2/000136", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000136", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9e940adea7c32759f0c95212bcc83902fac8ef5f8cc4fef74e0d153479f1ec3c.jpg", "reference_inputs/images/affd2dcdaf8eb527c0164b784fcca03599307d983a8ea5f904659e5ebdc351f7.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/town1_ground_aerial_v2/000137", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000137", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ae001cd42d76e6f87147dd08d6efa1f9ee182e555926abab696efe15c5f62a73.jpg", "reference_inputs/images/8d73b72cfd9d23656a556327bbce1ad9e8202251671c9b5d92f488eed380d689.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000138", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000138", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/614bea61ac9a60f84dd3186ff090ac89c59c5f19c19daa82db3fc2053a6cd2cb.jpg", "reference_inputs/images/0238f2f24037ed86b818cbce1240597128732f2ff9a788e59d4209d2b8c6fcfc.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000139", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000139", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d364b0e3541250a939545692adbf2b7f5dd4eda4eed7a82b53526764b5620522.jpg", "reference_inputs/images/c8bd03e2a0a3aee37bcbe453b1f52a0c3ec44fdc6daf9221c6fa84a0a6f5da82.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000140", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000140", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ee619f07d01e40018c525e41f160c6184809b171a852607efc945dc96e411053.jpg", "reference_inputs/images/879e112e284022f2692d0d7d519fe264dcfb695bd5472fce5af5345ec094d295.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/town1_ground_aerial_v2/000142", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000142", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2c57f3e448d39cce4b09504a4ae1620a90cfaa78e110b78269db0fbd1d03ddb7.jpg", "reference_inputs/images/0b80405f2c9cea1a3987fe9a1a8fb0322542d6131d4f4412afc3804ff1adc151.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_localization/town1_ground_aerial_v2/000155", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000155", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cd09321292f0788f78323c26abbf46a241c86ce743f184ba8aece6a49f533751.jpg", "reference_inputs/images/d172a9974ab6890dcb9d9bd0f4b65fb7dee6bc554912eb8ea5b2a08e43ec7e16.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/town1_ground_aerial_v2/000174", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000174", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/56fdec6518e6fd442d5bf50715cf842df5618b6e7125f80e58d99b5f95fb2643.jpg", "reference_inputs/images/b4299a14a7952da519e752a8614758e26ffd40e03f55f4082f9bfad7635732dd.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000176", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000176", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a8825c3ea8aefe3e6a1da5a55aa7242f4c037b15dacd0ba6d173469aaf064d16.jpg", "reference_inputs/images/73b0591df8bd42450b092703a7faf5032294f338d53e2fc9e85ddb18f4a0ef0f.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000177", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000177", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0991f603f641e6d628915793b4b36c83a5e4fc8369ca430f93a31e37fb53216f.jpg", "reference_inputs/images/4402e7cfe83543709449b72d6280666cd26854a0afc3cb9155307a83259697f9.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/town1_ground_aerial_v2/000178", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000178", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1d84a9a0b9eae92903a90403b8e36ed93169b339aaf2332fd02ab0937159618e.jpg", "reference_inputs/images/58535b94fbff0387d3d936599a30966b8283f1f79f82b2557b1e6cb15cb992b7.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/town1_ground_aerial_v2/000180", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000180", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/352d4dc632d358cc2ec1d87429f646632fdfc7c2672fe382b967ec52a5473e41.jpg", "reference_inputs/images/b1eaace9e8c9c750f4158d3ffc2793b2110b47bb8da21300d3824573f59668c6.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_localization/town1_ground_aerial_v2/000181", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000181", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/570950fea3c5b9a9401fd164a88933b77a227eb1e8d7870537c987807cdfe0c0.jpg", "reference_inputs/images/9350f6e7776f5d668a325e2aa8b02c45ca142ce9e54713d403ee50178fde93ef.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_localization/town1_ground_aerial_v2/000186", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000186", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/52f75b3baf552b2d21ded19fbe97e159b0362e15c0030dcac1725e1841405e8b.jpg", "reference_inputs/images/a642ca7e6bb82570e9d59cf1856e4d369d8fef09df3e4fb490ebc3ec0f196f8a.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_localization/town1_ground_aerial_v2/000187", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000187", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw aerial image without pre-drawn grid labels or boundary lines.\n- The grid regions are defined only by the text below; mentally split Image 2 by those definitions.\n- Each option refers to one full grid region, not to a single pixel point.\n\nGrid-region rule:\n- Use the full visible extent of Image 2 as the grid canvas.\n- Image 2 is partitioned into 2 rows and 2 columns of grid regions.\n- Rows are counted from top to bottom; columns are counted from left to right.\n- Pixel coordinates in the option list refer to Image 2 as sent to the VLM after any evaluation resizing.\n- Each option corresponds to one entire grid region in Image 2.\n- The exact street-level position may be hidden in Image 2 by foreground buildings or structures.\n- Even if the exact point is not directly visible, choose the grid region that contains that true location.\n\nSelect the grid region in your aerial image that contains the ground teammate's street-level position when capturing the query ground view. The exact position may be partially or fully occluded in the aerial image by buildings or other foreground structures, so infer the containing region from road layout, nearby structures, viewing direction, and visible landmarks.\n\nCandidate grid regions in Image 2:\nA: Top-Left region (row 1, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 0, 512, 288]; center [256, 144])\nB: Top-Right region (row 1, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 0, 1024, 288]; center [768, 144])\nC: Bottom-Left region (row 2, column 1; pixel bbox [x_min, y_min, x_max, y_max] = [0, 288, 512, 576]; center [256, 432])\nD: Bottom-Right region (row 2, column 2; pixel bbox [x_min, y_min, x_max, y_max] = [512, 288, 1024, 576]; center [768, 432])\n\nTask: based on the query ground image and the text-defined grid regions in the aerial image, select the option whose grid region contains the ground teammate's true street-level location.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c050fcbf6a2fbb7ae73cec88f53be285420ba47b9cda22afa8774910d8ad59db.jpg", "reference_inputs/images/a9d1dbd8325bc3c204aef035b36eed8ee745e3fb66204e5f9be015754f8582fc.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_localization/town1_ground_aerial_v2/000188", "task_name": "view_localization", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_localization/town1_ground_aerial_v2/all_questions.json", "question_id": "000188", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Localization.\nYou will receive exactly two images from the same scene.\nYou are Aerial Agent, observing the scene from the aerial view.\nYour ground teammate provides the query ground view captured at its own position.\nImage 1 is the query GROUND image (ground view).\nImage 2 is your answer-option AERIAL image (aerial view).\n\nImage usage note:\n- Image 1 shows the ground-level visual evidence captured by your ground teammate.\n- Image 2 is the raw target image without pre-drawn option markers.\n- Candidate locations are specified in the prompt as pixel-space points in Image 2.\n\nPoint-coordinate rule:\n- Every option corresponds to one pixel-space point [x, y] in Image 2.\n- x is the horizontal axis and y is the vertical axis.\n- (0, 0) is the top-left corner of the image.\n- If the exact real-world camera position is outside the current aerial image, choose the candidate location that is the nearest plausible in-image correspondence.\n\nSelect the option that is most likely to correspond to the ground teammate's position when capturing the query ground view, or the nearest plausible in-image location if the exact position is outside your current aerial image. Use cross-view scene layout, road geometry, nearby structures, and visible landmarks.\n\nCandidate locations in Image 2:\nA: Candidate point A\nB: Candidate point B\nC: Candidate point C\nD: Candidate point D\n\nTask: based on the query image and the candidate locations in the aerial image, select the option that best matches the ground teammate's location, or the nearest plausible in-image location if the exact position is outside the image.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/165fbcae599cf7a3a388132616855423a6c4a51f382b94071614dded67ca684d.jpg", "reference_inputs/images/6db787a38fdfe5093fdf8136c518aa1863546af9b840485e248b051b33982271.jpg"], "image_roles": ["ground_query", "aerial_candidates"], "image_sizes": [[1024, 576], [1024, 576]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000001", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093148_right_194; view right\nB: frame 1776093148_front_194; view front\nC: frame 1776093141_right_192; view right\nD: frame 1776093158_front_197; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1f9a676dfadd5f130a478760d9424fb255475ea365551dc571e073d8683d47ad.jpg", "reference_inputs/images/f14428af07d4e98d02371f5990cd221dc705a66a4c3a678235d8f637470c17ad.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000003", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094272_right_361; view right\nB: frame 1776094290_front_367; view front\nC: frame 1776094284_front_365; view front\nD: frame 1776094284_right_365; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/55b6ac9f4b4f02575516420af6cc027ab87f367df89241b4a215f3bd98617cfd.jpg", "reference_inputs/images/7b669d5d97127a61a1f6bbcf647c416f5fa2986e9c11de324e983feff34ea310.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000006", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093151_front_195; view front\nB: frame 1776093144_left_193; view left\nC: frame 1776093151_left_195; view left\nD: frame 1776093141_left_192; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9d8ea64782a1494ab5c19cd1a1daf4895b45198e3402c8c792e2c64da9498168.jpg", "reference_inputs/images/560936d10b80554cea4745e850f8157b9033abc772c58c80ac68e893e5aac8e5.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000008", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093034_right_160; view right\nB: frame 1776093037_right_161; view right\nC: frame 1776093043_right_163; view right\nD: frame 1776093043_front_163; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fe0542b6a0e90d0f44338ee4a0dcfc6b173bdfc5bc5384450d71ec6082dbac8d.jpg", "reference_inputs/images/7232495404d1d9d4ce0f9d2f62df76156c5a4f406cd7c771467e818aa96b313b.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000010", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092562_left_86; view left\nB: frame 1776092559_left_85; view left\nC: frame 1776092556_left_84; view left\nD: frame 1776092562_front_86; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6dbb6746fc0a64a22637ca5039e4c5787606f63653b809bcc08ddc1411048791.jpg", "reference_inputs/images/7a7bf043a154f4f22edde72453438df290ad4f02dfa8435c04a18065047c9f5c.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000011", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093107_front_181; view front\nB: frame 1776093107_right_181; view right\nC: frame 1776093098_right_178; view right\nD: frame 1776093094_right_177; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fd7b93c14326a1eb4b9351fed0b5ba18eb62a44d224f63ca26d3b4cce4f3c274.jpg", "reference_inputs/images/239451e4ce3e95c247c12b0c1b08cf6c2c98d79b6dd8624b69d1f02feff1c956.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000012", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000012", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094107_left_308; view left\nB: frame 1776094116_front_311; view front\nC: frame 1776094107_front_308; view front\nD: frame 1776094100_left_306; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ed9b688854ca82e7b099ca3fabbb18fbbbfbb81203b407655a68d5003906d68b.jpg", "reference_inputs/images/9d2b8b069a3c47c4225c591e251092d0eb5aeeb7d4adaf34b0428e463d3deda9.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000013", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092841_right_137; view right\nB: frame 1776092856_front_139; view front\nC: frame 1776092856_right_139; view right\nD: frame 1776092877_right_141; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a9108448fb86166e7f7ab10d12409941e046663809a2d1d945d933ebecb369a2.jpg", "reference_inputs/images/97a54b07c29bbbe54336b6d634527a491e4ec0b3dce35472ec79d730594c6534.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000014", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093047_right_164; view right\nB: frame 1776093065_right_170; view right\nC: frame 1776093056_right_167; view right\nD: frame 1776093056_front_167; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0117a32645a85c097df716bde720c02343013d47e982e7f3c9f8739ef2a31e70.jpg", "reference_inputs/images/a20403a51d6489fa63b5d7a2b5893eb60bb0eb6077200464791e2f7f45497039.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000015", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094069_left_296; view left\nB: frame 1776094062_front_294; view front\nC: frame 1776094062_left_294; view left\nD: frame 1776094053_left_291; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/380e66577c9ec880c8a14eaf33e4560890859dd68e66a3ec731ea516f70911fc.jpg", "reference_inputs/images/b9f6c2e977ff8de325b4a14aec7e67c3da3a36fa35f84b453520298ad92bc3a2.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000018", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093161_right_198; view right\nB: frame 1776093144_right_193; view right\nC: frame 1776093151_front_195; view front\nD: frame 1776093151_right_195; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b03c8badf6ca92200c08c0600bac28622d62565ab59f7929842d400e09e2d874.jpg", "reference_inputs/images/671a44855d6269ca116cc69e535a50ff1ca9098e6ed3b9651fe625c0e2ddc68d.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000019", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092666_front_114; view front\nB: frame 1776092666_right_114; view right\nC: frame 1776092660_front_112; view front\nD: frame 1776092673_front_116; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e4ba88ead16781e03227fe958569b51a9252d48de27a2371d473ea0253b67767.jpg", "reference_inputs/images/211a7e9593182c6972f2830f26e7be5845e4ea5f97d336a1108f74b606e721ae.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000020", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094072_left_297; view left\nB: frame 1776094078_right_299; view right\nC: frame 1776094088_right_302; view right\nD: frame 1776094078_front_299; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7d2619ecc8045dfec581bb5d67cc9912c9b38fe2160e7434afb7c84a19c12254.jpg", "reference_inputs/images/bacf6097a4a9f0e7624a03c9419e0290e71344ec0a3e43ce172092d8ec1abe36.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000022", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094238_front_350; view front\nB: frame 1776094244_front_352; view front\nC: frame 1776094253_left_355; view left\nD: frame 1776094244_left_352; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/32adb25d17272fdec60ac45551d66ee4edba04c131a6d3de4d55ed29aa603a7d.jpg", "reference_inputs/images/45a9df9b74c14520d94dd9bea39ec9e829d5b825dbbb8b6a78acac8abeffa560.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000024", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000024", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092929_front_146; view front\nB: frame 1776092929_left_146; view left\nC: frame 1776092909_front_144; view front\nD: frame 1776092950_left_148; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e0cc4417ef11dc892d9fb67d3a75c694809ab32111ab41409182028e4fd5febb.jpg", "reference_inputs/images/23ef7e0acd8294abe5f8eb9f7463d8caa02390abb9fd5a1bb66a7b63dc39f291.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000030", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094204_left_339; view left\nB: frame 1776094210_left_341; view left\nC: frame 1776094200_front_338; view front\nD: frame 1776094210_front_341; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7e254360b8c23238b41f358057878e89cbd854c829f88e4f358654575ca52521.jpg", "reference_inputs/images/7bc8d53469200f990dfd392413705908833ee84427a43da398602b695ed5f646.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000031", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000031", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092635_right_106; view right\nB: frame 1776092635_front_106; view front\nC: frame 1776092626_right_103; view right\nD: frame 1776092629_right_104; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f60a76258d28c6dc5609166a97a62cc763a7b288d4b1d33db8fd1b9c9eef34ec.jpg", "reference_inputs/images/5711a6193669579585ae5594a1f3777534f38e32227d916dbf86c922945f5aec.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000032", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092849_front_138; view front\nB: frame 1776092816_front_134; view front\nC: frame 1776092849_right_138; view right\nD: frame 1776092833_front_136; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e45c521edc4194f6f26c8a73c4c7b196bfcaf2765ac87b666c5448837a97ba0f.jpg", "reference_inputs/images/ab598132b51a2c2a3417995f0a58d7db528f157036c3816deb2857989f024f16.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000033", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000033", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092595_front_93; view front\nB: frame 1776092607_front_97; view front\nC: frame 1776092601_front_95; view front\nD: frame 1776092601_right_95; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/593125bcf6a7d2adb48925343233250923e5587bb62b7b247d744560213765a6.jpg", "reference_inputs/images/f510b323bcda937d468d491fd7310d27f1c4922e97ee6b988beace7ec7242a57.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000034", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000034", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094272_front_361; view front\nB: frame 1776094272_left_361; view left\nC: frame 1776094266_left_359; view left\nD: frame 1776094263_left_358; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2dc49b947b28aac8cb367fdfb816f8b0226aaaaf3838eb961fb94ee41f0e92b5.jpg", "reference_inputs/images/75c3bb0064893529667fddcebbc6468616f6aec8b1191fe75ff0e3dd45a966b9.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000037", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000037", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092632_front_105; view front\nB: frame 1776092632_left_105; view left\nC: frame 1776092623_front_102; view front\nD: frame 1776092641_front_108; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/de65a8157c8313f4147b1fe81431f3660c930dbf54b03f04d18c80d4a2bf2d7e.jpg", "reference_inputs/images/2d29a34b21ae7254406e4d0d55b7db0ba9c8d65fd7e7d6061d9b614137fa80b8.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000038", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000038", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094075_front_298; view front\nB: frame 1776094065_right_295; view right\nC: frame 1776094075_right_298; view right\nD: frame 1776094081_right_300; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/51d4c48d53ba1d7edcea0ba020abcd277e5e80a3f135700bf02ee4578ec54f67.jpg", "reference_inputs/images/90d0dae209660c37f2e126a145c01c1618dc4ca6797410750041cc843fad8efc.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000040", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000040", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092673_right_116; view right\nB: frame 1776092682_right_119; view right\nC: frame 1776092676_right_117; view right\nD: frame 1776092682_front_119; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/12185ced06d759a12445fa12448a76cb456b5e44b5949c37721d324023c6a094.jpg", "reference_inputs/images/4d29cb908b5ee2a372e4d3ceb781390828ba90cbe3ef02c3ac74f647c5aa8ea5.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000044", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000044", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092638_left_107; view left\nB: frame 1776092632_front_105; view front\nC: frame 1776092629_front_104; view front\nD: frame 1776092638_front_107; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e8f3bcd95128231007b882e593aae5efb0cc3d0909020ad23f6707109846e686.jpg", "reference_inputs/images/65f56679b9653fa53caf228d11d98aaa91e985c34e81817e5875d4453f0fbbd7.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000045", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000045", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092657_right_111; view right\nB: frame 1776092657_front_111; view front\nC: frame 1776092638_front_107; view front\nD: frame 1776092635_front_106; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/85faadba2ce8acb9465fa013512cd6eccfef27cc31b64489d7fb343fccc0933e.jpg", "reference_inputs/images/e2c65582ab271284f7f4812ac3759af5d65e72dd142102f59919b4eff74bae56.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000046", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000046", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094116_front_311; view front\nB: frame 1776094122_front_313; view front\nC: frame 1776094122_right_313; view right\nD: frame 1776094113_front_310; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/05539a8d79e56e84d0e2f125d7d2bc79e5400adee4aaedf5517bb980ac2a86d2.jpg", "reference_inputs/images/b6312673db382eec08876221c3a3229e21e11f63dfe970e1a89b06afb02e48ca.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000048", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000048", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093068_right_171; view right\nB: frame 1776093065_right_170; view right\nC: frame 1776093074_right_173; view right\nD: frame 1776093074_front_173; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8e9b0bfe54a146564dc1c9e5ec45b21c07e67a5e2512ba2bb62deea8c9777233.jpg", "reference_inputs/images/f722431c190f2bbd3cda947c6b2c32255a41c98e57f37a5e90e46b4d5fa75660.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000050", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000050", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093110_front_182; view front\nB: frame 1776093116_left_184; view left\nC: frame 1776093101_front_179; view front\nD: frame 1776093110_right_182; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/777128fa1d29408b224224787ef2076152aeb49f2b5bbf1c6324d50d5de35bf6.jpg", "reference_inputs/images/93498bf0d042ec5b1b4dbb64a04189223c71a0ec0ca1b69982609450a73ade2c.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000051", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000051", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094188_right_334; view right\nB: frame 1776094194_right_336; view right\nC: frame 1776094179_left_331; view left\nD: frame 1776094188_front_334; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3ef1a3c8036c0b2692c95f8913992a8522f3071b7aa5863fb86a370acbb8c8ee.jpg", "reference_inputs/images/5417d58a9efb178138726e0169ec6a84b2e531b38186a27b013163fb0e26cb7b.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000056", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000056", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093040_right_162; view right\nB: frame 1776093053_right_166; view right\nC: frame 1776093053_front_166; view front\nD: frame 1776093059_right_168; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8d7025b1eed96a6297b4abfa1c49e080679e129e320a398d36e3ce571486e694.jpg", "reference_inputs/images/85354f8662d0dac783cf9f507a56c7681745fc0ee2fbf4368e60361873486979.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000057", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000057", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093961_left_262; view left\nB: frame 1776093961_front_262; view front\nC: frame 1776093970_left_265; view left\nD: frame 1776093951_left_259; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/075e925e7c37434d3e4eef47302b990b24c4d67bfab899649be51b450d3fba66.jpg", "reference_inputs/images/c4ccf02b792f755910000bf9555b9cdc5f1e3f0c011515c637485d61dcdfb33b.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000059", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000059", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093116_right_184; view right\nB: frame 1776093123_front_186; view front\nC: frame 1776093113_right_183; view right\nD: frame 1776093123_right_186; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a051304b4ee41e56dcc73bd5aef0925c934326e489aee10f5824dd3acb77413b.jpg", "reference_inputs/images/1758956db06f24ebf4a26fcba9e04c18b5ba3575134a58e29286e40add9227fe.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000061", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000061", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093062_front_169; view front\nB: frame 1776093062_right_169; view right\nC: frame 1776093071_right_172; view right\nD: frame 1776093056_right_167; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/611f33dbb2f5cbbbd37444d020acf0e163f6a0f5b4b213482977f1abfae5642f.jpg", "reference_inputs/images/be1226c679ca642e7f3326f0db4264562ce6fca6469d4b384b78208dc907bafd.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000062", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000062", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093047_front_164; view front\nB: frame 1776093031_front_159; view front\nC: frame 1776093037_front_161; view front\nD: frame 1776093037_left_161; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9a2974de949e218f84b3bac1d8609d610d422b08b8bcddd8b7711a2852777155.jpg", "reference_inputs/images/d77214f70d778678a8083e5ecf33b09c1f0c9f1cb3deafe427c826019057a7f3.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000063", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000063", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092808_front_133; view front\nB: frame 1776092808_right_133; view right\nC: frame 1776092792_right_131; view right\nD: frame 1776092833_right_136; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5e5e934816b90b74a0974d1b4f382f697bbff3664ed5f40d3ef8263058059f8c.jpg", "reference_inputs/images/59d4b89a73b7666d70af014f8b86a78d64d140d6f71ec9a7b829521b1d6d0f46.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000066", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000066", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092898_right_143; view right\nB: frame 1776092877_front_141; view front\nC: frame 1776092898_front_143; view front\nD: frame 1776092919_front_145; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a7aa91d6525bb82518370c166ab08c97fb33f476fa73802bef0059b3e9440ace.jpg", "reference_inputs/images/40fb0c9322f475648876e0405feb4f2b0b2b9256858e559f9ca0a8518c333390.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000068", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000068", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094081_right_300; view right\nB: frame 1776094091_front_303; view front\nC: frame 1776094091_left_303; view left\nD: frame 1776094097_left_305; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a8e527963fbe8c6ea4f2b9fdd6d3cdf7e4f2f0961a18b24fd720cf002bc46f4d.jpg", "reference_inputs/images/2cd26065e33508750dcb9d5731a5a63329e148efdb156e13601bea7a19caf31f.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000069", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000069", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094179_left_331; view left\nB: frame 1776094188_left_334; view left\nC: frame 1776094179_front_331; view front\nD: frame 1776094172_front_329; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1beb73b0765db2333f778446e754a1e9359497cedcace35cdbe200a976331708.jpg", "reference_inputs/images/c59e9cbf7dfbfee8a9d78aaea695ba574aaa78ff5c1cf0986261dad7b1222a77.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000070", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000070", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094330_front_380; view front\nB: frame 1776094330_left_380; view left\nC: frame 1776094324_left_378; view left\nD: frame 1776094479_left_385; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6dbedbe85cbfa1451e9314aab787c2fc75e2d4ae5b3e93bab17480c495216615.jpg", "reference_inputs/images/c15e0683c03809e918d9222d52f58cc3cff878c5ef8a0412e8ce9dc320356770.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000072", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000072", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094266_left_359; view left\nB: frame 1776094272_front_361; view front\nC: frame 1776094272_right_361; view right\nD: frame 1776094281_left_364; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/fb99d6c087b8240dc599110c6cb9da24f495f6de87f0e22ffd61267ea8d7c8fd.jpg", "reference_inputs/images/3eb6338296f5a1168133f1e60bb09996dc9309a3a44deb5fa17f2953f7fd2742.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000074", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000074", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093154_left_196; view left\nB: frame 1776093161_left_198; view left\nC: frame 1776093179_left_201; view left\nD: frame 1776093161_front_198; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/942991b2ee434fd6e3518650176b9e0be995caae77093ca4f9901e9a45066880.jpg", "reference_inputs/images/6609e1851bec57f7aff20b35eb0478c68071ff4dabc34dffa186d62f07eafe3d.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_20m/000080", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_20m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_20m/all_questions.json", "question_id": "000080", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092645_front_109; view front\nB: frame 1776092632_front_105; view front\nC: frame 1776092645_right_109; view right\nD: frame 1776092635_front_106; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/baf32b10b2046e6c4e6733214ddab59ff0c7a6c2ce43112e277a6fd465d93d1c.jpg", "reference_inputs/images/fc6463de296c44831e0c14e65d4f7efea96e8be9aef50f55efd85d4bcf723d6b.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000002", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093973_front_266; view front\nB: frame 1776093964_left_263; view left\nC: frame 1776093973_left_266; view left\nD: frame 1776093967_left_264; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/dd863e54ef68fb734b1bc7be56bcb562acd767c62b76462df6140d5704bab169.jpg", "reference_inputs/images/80c27532f83cb488c272edddaceccfa5b9c91a20fb22c2b35539c25820679e6c.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000003", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093031_right_159; view right\nB: frame 1776093040_right_162; view right\nC: frame 1776093040_front_162; view front\nD: frame 1776093034_right_160; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b0896b0eca62db86188513855be5eb91383f58fd3fa87c502456219c73b530c0.jpg", "reference_inputs/images/4a3b3102864bd49efffc9abd63eb9d4230eea414331d194744b72088c793592a.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000004", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094234_front_349; view front\nB: frame 1776094241_front_351; view front\nC: frame 1776094234_left_349; view left\nD: frame 1776094225_front_346; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2fd8f8120dd1f2d2d17da5fa413b66d83e8e40d4c2c65c296befa5de96b8c680.jpg", "reference_inputs/images/e263b0bde970df7ea7dcd8ba1bad9213a77092456aa3f298ad3893ae7e1804cd.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000008", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093094_front_177; view front\nB: frame 1776093101_right_179; view right\nC: frame 1776093079_front_174; view front\nD: frame 1776093094_right_177; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8693cf8b202686a87725957fb75f06aef92933099e623afcf5bc07daa943536e.jpg", "reference_inputs/images/9349daccf27aad5d60e7016c76054490bb070a627d388775586930dbcefad807.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000010", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000010", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092604_right_96; view right\nB: frame 1776092595_front_93; view front\nC: frame 1776092604_front_96; view front\nD: frame 1776092598_front_94; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3a034ca21112b977813a3ad694df03e88aec707f8790f364c5664b4a6ca27374.jpg", "reference_inputs/images/553780a5caa1ba8dcb0520284d3ccce52899e8471f0f8d590acc82c53b6ec069.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000011", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094040_left_287; view left\nB: frame 1776094046_front_289; view front\nC: frame 1776094046_left_289; view left\nD: frame 1776094056_left_292; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d2c7a6f653ddab55ac9874fe86526ceb5c616d82832dc23883068e438a3c61bc.jpg", "reference_inputs/images/55eca9cf326375ec7aba2df52f5c40d040ddf7e7d5fe59c059cf552cb346fe0d.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000014", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093135_right_190; view right\nB: frame 1776093135_front_190; view front\nC: frame 1776093141_front_192; view front\nD: frame 1776093141_right_192; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/18d2c00b199d816dc243a2b67f1718c00d3158eac435f7078f2f6b81b5b91ce1.jpg", "reference_inputs/images/180b2c169b5aaeb771a97bd06a9906b8164d745090f4c557946ef1489f397867.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000015", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094297_front_369; view front\nB: frame 1776094290_right_367; view right\nC: frame 1776094297_left_369; view left\nD: frame 1776094303_right_371; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a9923432ee3b1cd004161cf0669e3f2592d344e9e9a6e846d75d1e670c349559.jpg", "reference_inputs/images/a8fd0f1a8712015fd2ad2c29c9aeb9ad9326286193142bbf1c856e3bdfdfe3b0.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000016", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093037_left_161; view left\nB: frame 1776093021_left_156; view left\nC: frame 1776093028_front_158; view front\nD: frame 1776093028_left_158; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a21e3a5cac08af83e927bcda93c8c407201fd28de9f8ac4b4f5333138466d66c.jpg", "reference_inputs/images/61507f63150827c2f382c79f28c6817f89e45056af08f2f4461d1cdc27ce21e1.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000017", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000017", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094116_front_311; view front\nB: frame 1776094116_left_311; view left\nC: frame 1776094107_front_308; view front\nD: frame 1776094122_front_313; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7cce4db4e7ddd8e35e4afb7d2a50645b3eec4274394bceda3f3dd8317e125171.jpg", "reference_inputs/images/ef813b37d18603e3d299e2f044db9a3cc49326d952e61e59349609ea325ecac7.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000019", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092929_front_146; view front\nB: frame 1776092939_front_147; view front\nC: frame 1776092909_front_144; view front\nD: frame 1776092929_right_146; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/75fff77463644e93370a3c2660d054ea6c6f53689bd42265eaef450843b5855e.jpg", "reference_inputs/images/6ee80009d57dd8cd7229bc3d95e47acce20e1721e15e18364d5e4e6ce56bb884.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000021", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092455_front_66; view front\nB: frame 1776092455_right_66; view right\nC: frame 1776092446_front_64; view front\nD: frame 1776092446_right_64; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6e20559c41d5109d21ba989a4f6e2da0eaa31537167aafe555c250fa21f53c57.jpg", "reference_inputs/images/0300b758bdaa5b086735c8ecc9ab4758b3aa03595c495ffb05ebae9012423c6c.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000024", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000024", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093126_front_187; view front\nB: frame 1776093132_right_189; view right\nC: frame 1776093132_front_189; view front\nD: frame 1776093123_front_186; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d2cebd29b90f602f54365f41e77dd915575721ab29f62ddf80eec9c2ee3b06e6.jpg", "reference_inputs/images/c8e5e08598f42aa581bcbe6fe9a04f7e8dca7c91b39e15f2d9a722596047e3aa.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000025", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092571_front_89; view front\nB: frame 1776092595_left_93; view left\nC: frame 1776092565_left_87; view left\nD: frame 1776092571_left_89; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ad3593b83d60049e24ab1d195fc3252a8f177f068c3fc673e7cabf86e798ff40.jpg", "reference_inputs/images/ab5075f35f34d1304e10ef5c779f2659715363ef6b7529195fd595981e09c2ab.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000027", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000027", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093059_right_168; view right\nB: frame 1776093059_front_168; view front\nC: frame 1776093065_right_170; view right\nD: frame 1776093053_right_166; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7a1aa661fbfdbeedb61f168bcf115d216fb721898946c056ce02f3a1c035b784.jpg", "reference_inputs/images/aec66578061d86a3a5c2ccd4ec3b6b3f2cd116b3820169d7f1c3e73363ca379f.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000029", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094210_left_341; view left\nB: frame 1776094210_front_341; view front\nC: frame 1776094204_front_339; view front\nD: frame 1776094200_left_338; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1ceeb32d1ec50d882c4e398e625bfcf7f35fb6186543d11f5d103b9e73f3168d.jpg", "reference_inputs/images/2ed6301f10915bf4f148d31021813a1b85b2ac9588969564abadfcb98c591111.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000030", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093065_right_170; view right\nB: frame 1776093074_right_173; view right\nC: frame 1776093065_front_170; view front\nD: frame 1776093059_right_168; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/91742b99d33812d89d368354e54e66962371fc8fb70ec9f627c55ec2c934cffd.jpg", "reference_inputs/images/3ce7ffed04b8baa4ba54d1bc3d6da6f5bbe965dbd5825af7386ad68f4205ced9.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000033", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000033", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094263_left_358; view left\nB: frame 1776094269_left_360; view left\nC: frame 1776094269_front_360; view front\nD: frame 1776094263_front_358; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b4c3b5c47f04392c3976477b0c0479f0386633fa30d09adf81197a7d7106ca90.jpg", "reference_inputs/images/e7ab1ded9a2430444be8b22d2d7d2d1f8e5e745846a36a9434f1b7eb0b361fff.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000035", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000035", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092682_front_119; view front\nB: frame 1776092682_right_119; view right\nC: frame 1776092673_front_116; view front\nD: frame 1776092698_front_121; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d92a50adc812e898612332715d38bf7d28ce0e75725c0067300db2bfab4c4b6d.jpg", "reference_inputs/images/efcdea482b5219366a07edf5977e535e2ef6be58256dba8f8ca6567a967f9fa7.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000036", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000036", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092833_front_136; view front\nB: frame 1776092816_front_134; view front\nC: frame 1776092833_right_136; view right\nD: frame 1776092856_front_139; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/10f5de2f922740a7b826c92132c3a54c8db8806c6900491b1c3542df6f4c12b8.jpg", "reference_inputs/images/35aa5865fedf695d2e457d488b3a14d3b7bcd711415ac9dcedb06884d6b0a78e.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000037", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000037", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093995_left_273; view left\nB: frame 1776094002_front_275; view front\nC: frame 1776093989_left_271; view left\nD: frame 1776093995_front_273; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8baef60a496e15a557123c85199fb685288234f8d9d094ee87defab043a0914d.jpg", "reference_inputs/images/e1b9a20d8821b8f0e933558d2da665e3b8be4d4a186b4694d2974e931c7ac7b7.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000038", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000038", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092347_right_49; view right\nB: frame 1776092338_right_46; view right\nC: frame 1776092326_front_44; view front\nD: frame 1776092338_front_46; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4522079847744186fa33481df219caaaba11da84f15b95385c6ef5e418ce4159.jpg", "reference_inputs/images/60ba1df1c32429aa51938191a8cecd79d3402a47e082be5cbeaa7d827f975a84.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000041", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000041", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094151_front_322; view front\nB: frame 1776094135_front_317; view front\nC: frame 1776094144_right_320; view right\nD: frame 1776094144_front_320; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/afa7d2df9354ca6fbb755d9f42213f38b05618145bb5bc2666d22f0762e6d4ad.jpg", "reference_inputs/images/8c21e4b33cb722057a9f9780de5907fb0fb89d9d8bd066ec89f77839172a919f.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000043", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000043", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093836_front_239; view front\nB: frame 1776093836_right_239; view right\nC: frame 1776093811_right_236; view right\nD: frame 1776093852_right_241; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/906a4cc04c4cdbbd34b057263682cf66b48da5cc9103b5d7391b0e2ac3b320c7.jpg", "reference_inputs/images/0ae90d86b968141ddde030d0e92a14c8a0d7cc00df574343467aa62d2012d2d7.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000044", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000044", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092877_front_141; view front\nB: frame 1776092898_front_143; view front\nC: frame 1776092898_right_143; view right\nD: frame 1776092919_front_145; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e48702e52b8b350991dcdf5e60cdb2b5f7e2d1f889d3ab76c68234550625867d.jpg", "reference_inputs/images/185d6e9da02e88488284536d5cb3b4faf0505105522816e0bdf8f7887485b718.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000046", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000046", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092350_right_50; view right\nB: frame 1776092334_right_45; view right\nC: frame 1776092341_right_47; view right\nD: frame 1776092341_front_47; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6373082ee2076524cdeec8abb3bf23d123c79e305872961765178aaa78db3fc1.jpg", "reference_inputs/images/f0a063c8c8ad8af23f534312a5bbef4c1e046850950d75ea71bbe418f4d0c8f9.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000048", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000048", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092824_front_135; view front\nB: frame 1776092849_front_138; view front\nC: frame 1776092808_front_133; view front\nD: frame 1776092824_right_135; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7cc4625019369a0659c784d818c9d05dba379ffc29a98d262e9f422107426c9b.jpg", "reference_inputs/images/bd01cc254eb72b1bf78c9ac020ea851f5171ef66eddbf120c275155dfe99c66e.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000049", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000049", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092970_front_150; view front\nB: frame 1776092939_left_147; view left\nC: frame 1776092950_left_148; view left\nD: frame 1776092970_left_150; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ce5540457992a0603798b67f9247db2e71c584e4767f62549e5d540da61b1fae.jpg", "reference_inputs/images/3173dc0852eb10e460f919a4028190795d71172c1129f58f3c9454007a2560b3.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000051", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000051", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094200_front_338; view front\nB: frame 1776094200_left_338; view left\nC: frame 1776094194_front_336; view front\nD: frame 1776094207_front_340; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/90289818097c17dea279c1aa342985c773ca32e347e00508d30a14df325a8b98.jpg", "reference_inputs/images/37c2111f1672992c7ddf289ebde20cf6ecd136e8bbbf4b7fc41cccd232ffed17.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000052", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000052", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093793_front_234; view front\nB: frame 1776093793_right_234; view right\nC: frame 1776093769_front_232; view front\nD: frame 1776093811_front_236; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/f059bdb9a2a3eca10c9b8a0bde3e28cc365568341c81c42124c5009dcb10d21c.jpg", "reference_inputs/images/75a06e20da2aa60458dc39fc31a7a7a1c7df51d937cad4e5cd2f2694514ab609.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000054", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000054", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092446_front_64; view front\nB: frame 1776092455_right_66; view right\nC: frame 1776092455_front_66; view front\nD: frame 1776092446_right_64; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/09d021d6b5ca25bc928360e4aa7d6fd36a19186434803fb7aac96deba17798f2.jpg", "reference_inputs/images/85cffe9fb4636a98468701863640d3f2d4535ec69c5699ec8299340308b096a9.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000055", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000055", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094129_front_315; view front\nB: frame 1776094129_right_315; view right\nC: frame 1776094135_right_317; view right\nD: frame 1776094119_right_312; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/09945833a52a5d5be2587784e706256d02d991b1ac16175ac113f0906d24c818.jpg", "reference_inputs/images/87a214b59dcd5ff80bc84d4dc600cfe34fb9f24b46bea03d38141f1f05bbbe5b.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000056", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000056", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093059_front_168; view front\nB: frame 1776093053_front_166; view front\nC: frame 1776093050_front_165; view front\nD: frame 1776093059_right_168; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/130240a8ddb61078b9c604211ab959f15747288ac8c4416550dc32779ac9a774.jpg", "reference_inputs/images/13838cb8889ab44bef4f8c2a23788e2eee8616511483f2c6bf8aacec894a8e8b.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000058", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000058", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092676_right_117; view right\nB: frame 1776092669_right_115; view right\nC: frame 1776092682_right_119; view right\nD: frame 1776092676_front_117; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6e081f7c629d9bc858e11446864077231604768e12f7747f84cc086e3ab25be4.jpg", "reference_inputs/images/c33f2ed9d32db654042a741a8b7f1b35905522b551b2341a6084422f58268a97.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000059", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000059", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093037_front_161; view front\nB: frame 1776093037_right_161; view right\nC: frame 1776093043_right_163; view right\nD: frame 1776093031_right_159; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/dc1ed63612a2b9791070ad56e2dec068c2a332ff419c21d7b6d40a7625d0af37.jpg", "reference_inputs/images/1699f619b82a241670d6cdc64a63cbd2ff816730f334e3c4b00e4a367122a688.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000060", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000060", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094287_front_366; view front\nB: frame 1776094290_front_367; view front\nC: frame 1776094297_left_369; view left\nD: frame 1776094297_front_369; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5f9073acf06a968f5f17ae2997ca61332641b0d6f7650763721c690e0670f299.jpg", "reference_inputs/images/6bad75864b4503fa8aa7bf36dedfed602ec9d367a528b9d8eeacbbf409c9d4a6.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000062", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000062", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093752_front_230; view front\nB: frame 1776093803_front_235; view front\nC: frame 1776093760_front_231; view front\nD: frame 1776093760_left_231; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7e4e6fa5252b4d81e37beeb4442b108fc89e78d874c0d79ee8344cc7e823c845.jpg", "reference_inputs/images/a16dc3f518a8ee8a8e0c34d405e092e2d53c9b43636274137bf17cb293e552cc.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000063", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000063", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092808_front_133; view front\nB: frame 1776092792_left_131; view left\nC: frame 1776092808_right_133; view right\nD: frame 1776092833_front_136; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c4d960fb8e3722c9837c1e8a02119cf8fbc983c821518c23f414ac3bebc75190.jpg", "reference_inputs/images/6e0e87da784e807c22461bc50e2e9bb664a81b14e87d508da5685df1ad1c3e63.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000067", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000067", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092801_front_132; view front\nB: frame 1776092816_front_134; view front\nC: frame 1776092801_right_132; view right\nD: frame 1776092754_front_127; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4ea72508c4819374ba5e2c90de667ff3e2eef83626881ae0ba245f9c83f0c914.jpg", "reference_inputs/images/34a89bd09941415fa1d6ede31ae1040cdd2d1fd2372e28a32f1bbe8830bef21e.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000069", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000069", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093116_left_184; view left\nB: frame 1776093126_left_187; view left\nC: frame 1776093126_front_187; view front\nD: frame 1776093116_front_184; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/08f81f8725f830e5863d723e3b575e5b8bb9ae67cfcbec57d385a06baed43f22.jpg", "reference_inputs/images/5c1298f09f4a3c3a342851e1508ed956cbc4839e4e4e026dbb2990d6f4b77271.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000075", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000075", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094275_right_362; view right\nB: frame 1776094284_front_365; view front\nC: frame 1776094290_right_367; view right\nD: frame 1776094284_right_365; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/98dd71a6482b534c499d97e6367ee83ca122a45bf63f78bb976fb8ea46e0c4b9.jpg", "reference_inputs/images/edc353e8d714df4ded5603169c09ec76563b48d69c4165129a0792ee0d494412.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000077", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000077", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093803_right_235; view right\nB: frame 1776093769_front_232; view front\nC: frame 1776093769_right_232; view right\nD: frame 1776093760_right_231; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/39da9f70447920af3a7c0056fffe2a9c21cf82e862e3a25ebb08fcba7f35037f.jpg", "reference_inputs/images/25f61e4f7e413219f67015d9d84221d57d3a8194a63ad567b7e2b8218a466d53.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000079", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000079", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093129_left_188; view left\nB: frame 1776093120_left_185; view left\nC: frame 1776093129_front_188; view front\nD: frame 1776093116_left_184; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0434e9c3d4a686968d1d4ab3edc0ac45ddba3417799dc8ab1e672ad5cdb3f261.jpg", "reference_inputs/images/5c9d45ee7bba8d7b94cbfcf9cd1f636ca3eabde41766be14ff8ded2f7da811c2.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000085", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000085", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092326_front_44; view front\nB: frame 1776092344_right_48; view right\nC: frame 1776092334_right_45; view right\nD: frame 1776092334_front_45; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/20e63a6bf38ecb12e507bb5d4e5b0929cf0e785976c7dc073f810940c8ff857d.jpg", "reference_inputs/images/db4fa2c91222ea3595d715b1c4bbf8f2b164bb647aeb09cc26fd388d3740d53c.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000090", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000090", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092341_right_47; view right\nB: frame 1776092344_right_48; view right\nC: frame 1776092350_front_50; view front\nD: frame 1776092350_right_50; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/937100458a7244095afdf2446e860f1e38ad17d94fd43d19dd379ea083f02e26.jpg", "reference_inputs/images/c0a3c345d621ab5cf7ca2640e751805f1fd79475ad3d6b926d42ad317e0271fa.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000096", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000096", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093071_left_172; view left\nB: frame 1776093079_front_174; view front\nC: frame 1776093079_left_174; view left\nD: frame 1776093068_left_171; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5b0766e2cf797f7f982e625939f7806415aa4de9de7a524647d655760948bca7.jpg", "reference_inputs/images/f1b8c635069ee2be95513a0002fb392c19d993a1dbf55a49333e3ab18925dda0.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000097", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000097", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094191_front_335; view front\nB: frame 1776094172_right_329; view right\nC: frame 1776094185_front_333; view front\nD: frame 1776094185_right_333; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a786c4a34a29ddb363321e9be7ddf0ad6fe460aad21ed8ee819a420d049c4e73.jpg", "reference_inputs/images/af197d5d7768d4738b9a9b51773a597b3a31d10cbec98b78c4106daea9dfbebc.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000100", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000100", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092981_right_151; view right\nB: frame 1776092997_front_153; view front\nC: frame 1776092997_right_153; view right\nD: frame 1776093021_right_156; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1d01d547e814dbb6128a17aa122b957f97cfa9a0d7b7192c81c85902b5bef371.jpg", "reference_inputs/images/162bc706289b0af66b349791c2aee17bd74287c7f340e80f8c0fc3dfa5be9550.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000102", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000102", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093869_right_243; view right\nB: frame 1776093852_right_241; view right\nC: frame 1776093852_front_241; view front\nD: frame 1776093828_right_238; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/61c19ce2a731a51865f83043af99683d4181b88ffb2a1c59d205df87dfb746cf.jpg", "reference_inputs/images/50e245e12d5ac09b68ccc87289af34fda64c5d3c8b990ca45f1ef7f4e4891db1.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000104", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000104", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094081_front_300; view front\nB: frame 1776094075_right_298; view right\nC: frame 1776094088_right_302; view right\nD: frame 1776094081_right_300; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e8cd352cc5ba99ba24d763736350c986727762fac16ddc10962fd3c228b52533.jpg", "reference_inputs/images/b7e78c21b59c55e0b08445f4d3c8acbba44845b4d2832c74e2233e64ed7cbcc9.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000107", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000107", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094110_left_309; view left\nB: frame 1776094103_front_307; view front\nC: frame 1776094116_left_311; view left\nD: frame 1776094110_front_309; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c61fd9baade91409b88694c8925cbca9e9b3cbfe31e8dee15e7df78b53e682f3.jpg", "reference_inputs/images/05b1bbe0fcc5450f900c8024652add254a98cbc7d8738fb28da14e593d156964.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000110", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000110", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093154_left_196; view left\nB: frame 1776093161_front_198; view front\nC: frame 1776093179_left_201; view left\nD: frame 1776093161_left_198; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/aa6a3ec7a6b97a88ea0dcd6df773c5261dd921027cdf86161165be4ddb1323c5.jpg", "reference_inputs/images/71ee3c7107c8faca9249cfb57e0651c0533891ad54b01730a77f78b36124c59c.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000112", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000112", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093047_front_164; view front\nB: frame 1776093047_right_164; view right\nC: frame 1776093037_front_161; view front\nD: frame 1776093040_front_162; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/999b91b83d48ef328ebe334141dccd20606bb0791a45da3cb792feab16ef23ac.jpg", "reference_inputs/images/eb406a7446fbfa77728eade8b7fe371e9b115c00a587dab5174492ae11f1453c.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000118", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000118", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094129_right_315; view right\nB: frame 1776094129_front_315; view front\nC: frame 1776094119_right_312; view right\nD: frame 1776094122_right_313; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2c03a5a6484134c856b754591961a180b602dae89e984c84010b2539eb68fbda.jpg", "reference_inputs/images/2edabd61fd0d34027021fd340230dfb9d180b5d5e1041893dccd8523c5eaea39.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000119", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000119", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092690_left_120; view left\nB: frame 1776092679_left_118; view left\nC: frame 1776092690_front_120; view front\nD: frame 1776092676_left_117; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3d04743e367ff479b83bb0e3c27c0b10fd794cb2f317c623c6c1741f17e05d49.jpg", "reference_inputs/images/256b69a0cb8e20ba61ad96eea040abb8e478882c18181a61b725043749661ac6.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000123", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000123", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094207_front_340; view front\nB: frame 1776094207_left_340; view left\nC: frame 1776094200_left_338; view left\nD: frame 1776094200_front_338; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/130364aa6f7af5552c5796e163348cb60da54b6a7a42db737392382b0ef99908.jpg", "reference_inputs/images/8ff947e81fec345e5fff1e142d264cefb0cd1f9f41d45b9ecfe7fcc6d74980bb.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000124", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000124", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094197_right_337; view right\nB: frame 1776094204_front_339; view front\nC: frame 1776094210_right_341; view right\nD: frame 1776094204_right_339; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b8d7c3b6da9c9cbdec613cdde26658542cb34359032c34d49fa3b91975d9cbad.jpg", "reference_inputs/images/86e9eb2b0bb96919985a4892a597e52451c0e9d982044a69491091922a8168fb.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000129", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000129", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093120_left_185; view left\nB: frame 1776093113_front_183; view front\nC: frame 1776093120_front_185; view front\nD: frame 1776093129_front_188; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e59297885724fdd71428432156fd21589fcd501d5f7e6df7ab25bb1538dbb21f.jpg", "reference_inputs/images/281e6f464e39549086a65db70cc978f0224f8138d93b3e77ed631c279423780d.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000130", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000130", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094107_left_308; view left\nB: frame 1776094116_left_311; view left\nC: frame 1776094107_front_308; view front\nD: frame 1776094100_front_306; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0ca54a2664d45055090f6f24a0a63cee057b2f1ffcd38a5fb48a26df05d8655c.jpg", "reference_inputs/images/ad5248f26bdf42565d0c7b9606493b61fd1a02c07ddf4a51d216f3bb8d271578.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000136", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000136", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092669_front_115; view front\nB: frame 1776092663_left_113; view left\nC: frame 1776092676_left_117; view left\nD: frame 1776092669_left_115; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/31dc20c66f06c71089587ae6edb85c69940e0a02014d3271590457ab55afa20f.jpg", "reference_inputs/images/0ca668eb79b72ef01b759ed5b4db75e0a50ffb4a71b5a20cf53801332b4c042c.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000137", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000137", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094160_front_325; view front\nB: frame 1776094160_left_325; view left\nC: frame 1776094151_front_322; view front\nD: frame 1776094166_front_327; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/eb0f7ac8348cfc5fb2cf695524c0af2bcf3ef96fc3976ec4a6a742d3adc7b444.jpg", "reference_inputs/images/e9e29076924a1b30809f97d3fbec1ddda6dfac4bf361161cb388738daf6b41fe.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000138", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000138", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092338_front_46; view front\nB: frame 1776092326_front_44; view front\nC: frame 1776092338_right_46; view right\nD: frame 1776092344_front_48; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2c1b9b7394a689958dc5047aaaf142b8c3d188a91c3a74b6b94151bb377c8f6a.jpg", "reference_inputs/images/14c366b9793addd069a52fd3b8e4c677e1bca98f540c196438fefad7ce2e5530.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_30m/000141", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_30m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_30m/all_questions.json", "question_id": "000141", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092673_left_116; view left\nB: frame 1776092666_front_114; view front\nC: frame 1776092682_front_119; view front\nD: frame 1776092673_front_116; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a01db7944534824b71018eb47b9215545d88b39cbd4bcf5f5fc1ed81000b7bfd.jpg", "reference_inputs/images/5f5bbdf7454ebdcefe97fc40d6b0cbaceaf76b2d3e73af92033555bd4f62880e.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000004", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000004", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093973_front_266; view front\nB: frame 1776093955_front_260; view front\nC: frame 1776093964_front_263; view front\nD: frame 1776093964_right_263; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2a99830de5c3aa734870639afd5c97252c3d302b886a349f75883065d508c6d4.jpg", "reference_inputs/images/a1ef7b50ae88109b73285be35e1008db758b2daa0cd9c58f0f0433c6647e0e2b.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000006", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093107_right_181; view right\nB: frame 1776093107_front_181; view front\nC: frame 1776093094_right_177; view right\nD: frame 1776093098_right_178; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bbfcb687256e5c62ee70e4dc31e013cadbe88f6cc2258bb7fcce16b7b34758be.jpg", "reference_inputs/images/df05cc1250584202ac6ed2f4d5eef242fb80831475aec5425a81bc2c0939adff.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000007", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000007", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094100_left_306; view left\nB: frame 1776094094_front_304; view front\nC: frame 1776094088_right_302; view right\nD: frame 1776094094_left_304; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/98be3cb7539a46948b4d63d5b6e142db9b1fc18a52914332581990de4b5193ec.jpg", "reference_inputs/images/18d2c51629beb4d2aa8ad4a3dff2eedc526380855e7d14b67a2c2b2a2e14a702.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000011", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000011", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092182_right_27; view right\nB: frame 1776092152_right_25; view right\nC: frame 1776092182_front_27; view front\nD: frame 1776092141_right_24; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7ebe738c8533772617dccf07e78074230358f675ee461e2aabba051d7c9340d7.jpg", "reference_inputs/images/4349da162c9cd1dda6adbe74b0778d4700f2c5f7561c94b809503826ada9c832.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000013", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000013", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092193_right_28; view right\nB: frame 1776092214_front_30; view front\nC: frame 1776092248_right_34; view right\nD: frame 1776092214_right_30; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/58a71ba1f3a38e77a90abecca0f4345897ed29c731975fb377f99930e761532a.jpg", "reference_inputs/images/dbc584bad672920f7bcedad8b7f2ac808214150d17cd852239d744d50349696c.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000014", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000014", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092553_front_83; view front\nB: frame 1776092559_left_85; view left\nC: frame 1776092553_left_83; view left\nD: frame 1776092544_left_80; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/85e9b5381d954016963a59b6fdf91012e32dc5860acd9bf042f23965f9508827.jpg", "reference_inputs/images/046377e98010b8de774389506a2b6044875a818bdc0ed773a371aab1cdd5515f.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000015", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092629_front_104; view front\nB: frame 1776092641_front_108; view front\nC: frame 1776092635_front_106; view front\nD: frame 1776092635_left_106; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/50603b11b120581c062565c23b4f1f58de8b2d9c9ab8b632817a9840aec6716d.jpg", "reference_inputs/images/998831eec6d5b57a3195edd2b8093cfdc816699d904b0164a267b0d02a5bccfd.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000016", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093053_front_166; view front\nB: frame 1776093059_front_168; view front\nC: frame 1776093050_front_165; view front\nD: frame 1776093059_right_168; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ef97a6b6013d0ce9a01b4269c2aee18aa0f4f71865962daecc5988a7cf5a0d34.jpg", "reference_inputs/images/bcf086b52eb42083679c872927c7bf0d37aa6390085a5ab3517db943eea7a50e.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000020", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093595_right_213; view right\nB: frame 1776093292_right_208; view right\nC: frame 1776093310_right_210; view right\nD: frame 1776093595_front_213; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4197c99f725f7b62c7f9418ee1c8cd82f1f62d10df6396bfe7a67b2c1f7931ee.jpg", "reference_inputs/images/72e6f4a700ce98a8459aea2e9312367fd86c7b5d6971ebf35aaf649519c299ec.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000021", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094194_right_336; view right\nB: frame 1776094188_right_334; view right\nC: frame 1776094194_front_336; view front\nD: frame 1776094200_right_338; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/eee0bfd79a08bbc775f7a1518c1b2c59056b0f394ceaece393998dfe250cbaa5.jpg", "reference_inputs/images/84f97040803f071fa51c49fbedae8f1b025a3be90fe3b62eb8af9eb3d9b869cb.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000022", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000022", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094179_left_331; view left\nB: frame 1776094188_left_334; view left\nC: frame 1776094194_left_336; view left\nD: frame 1776094188_front_334; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/54bc1594517eaa057a6e706aeff53424e3e1b846fa425f80b440ae6ce89ea75f.jpg", "reference_inputs/images/4b930dc040a12a535710f1a90ae848fdce40bff864e8a2e9ffddd8be772f47b4.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000023", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094324_left_378; view left\nB: frame 1776094333_left_381; view left\nC: frame 1776094500_left_387; view left\nD: frame 1776094333_front_381; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/845b28376c31689a1a3ba3805633cf413e648c7f9daabef4b8c492c797ba2c5b.jpg", "reference_inputs/images/ac3d01491316bb53b8ebffe2e2b4cb888df9c1ff250c9c9763d4dfd226ba3e82.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000024", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000024", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094126_front_314; view front\nB: frame 1776094116_front_311; view front\nC: frame 1776094126_left_314; view left\nD: frame 1776094119_right_312; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d91e812b0c92b33c667df7de4fb0968a0ebfd66ac724da60f3ab9f179d900a86.jpg", "reference_inputs/images/00a9b18d79ca315a49ad1da9b909194a99419d7e8a0bc271948840678b603223.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000025", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000025", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093828_right_238; view right\nB: frame 1776093803_right_235; view right\nC: frame 1776093828_front_238; view front\nD: frame 1776093811_right_236; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/650a6d1806b0c0e20bb26ce30484ac99efb303def45e5fc7d0a16f7f886c8e24.jpg", "reference_inputs/images/122db4a2a982fe2d8ccf3dde76d3f7b303187295fae45c86568f0a5b10dcd56b.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000030", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093961_front_262; view front\nB: frame 1776093970_right_265; view right\nC: frame 1776093951_right_259; view right\nD: frame 1776093961_right_262; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b0154040cbd58ec6ca8c5596a2faede58ec013a912836e53e7b8a1f646c56bcd.jpg", "reference_inputs/images/3635ace9cba72c0ed9aa5493882a1afc633821d032d0ba73971809768b2ecceb.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000032", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000032", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093769_left_232; view left\nB: frame 1776093793_left_234; view left\nC: frame 1776093811_left_236; view left\nD: frame 1776093793_front_234; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/baf5828b996035ff0a72e83fdc412e4c01c61c40de2b12338aa41ce9da462d36.jpg", "reference_inputs/images/cdd10e6492de45260425c943417cdc6789c0a708deb457feb5eafcaf77a51caf.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000041", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000041", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093869_front_243; view front\nB: frame 1776093852_left_241; view left\nC: frame 1776093852_front_241; view front\nD: frame 1776093828_front_238; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d5b041ef24bbe6168b2f5f66104d746d37aee76709acf888a5df1d6fcd84ea3b.jpg", "reference_inputs/images/0134878d97f19b62978c6baf730647df249ddeda1fa56a1fb71ced247563189c.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000044", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000044", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093101_front_179; view front\nB: frame 1776093101_right_179; view right\nC: frame 1776093089_right_176; view right\nD: frame 1776093094_right_177; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3351981f257135db236433e2a00e33d8feff9ced756c751aaf57fb6549743194.jpg", "reference_inputs/images/057f8a5dbf4e2ae1c4c81530da60a71885509ceb7312ce0faa6b02bbe13d689a.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000045", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000045", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094059_front_293; view front\nB: frame 1776094059_right_293; view right\nC: frame 1776094065_right_295; view right\nD: frame 1776094053_right_291; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/216f75f94fa4166e8084eb85f99ceabd6ce21d4cf81e0a79cdf7efae64846076.jpg", "reference_inputs/images/b66218c67df6e6b08188941e1f1b197c280c3d12198a5b11a477ceca9953c21e.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000047", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000047", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094315_front_375; view front\nB: frame 1776094324_left_378; view left\nC: frame 1776094309_front_373; view front\nD: frame 1776094315_left_375; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d2b667f5dadd22b6785856832e7628009f54142ffb15a4dbaa5cae6c5541de94.jpg", "reference_inputs/images/af3afb34b907e268edf580ab9bf6f29d195b18f9b317f78fe197975e32009fbc.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000055", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000055", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776092623_front_102; view front\nB: frame 1776092623_left_102; view left\nC: frame 1776092616_left_100; view left\nD: frame 1776092629_left_104; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/47641b67fbe2d073b63245388e7b7b50fc10e034220351ceeb108c37e5da8abe.jpg", "reference_inputs/images/d1c09572ba6889cba4ca8a85ed5d8f0405a76ae5008a4034f3f7e168c81f5617.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000063", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000063", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094188_left_334; view left\nB: frame 1776094194_front_336; view front\nC: frame 1776094200_front_338; view front\nD: frame 1776094194_right_336; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/42e37f9b8432a2ca1fe747f5bef84f3bbdeecbafccb2582d4fe33a10f30196e6.jpg", "reference_inputs/images/53d8f37d3a0f2982d48878144e187c89e7fcf58c61bb36aad2cc051aa019ce19.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000064", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000064", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093120_right_185; view right\nB: frame 1776093113_right_183; view right\nC: frame 1776093110_right_182; view right\nD: frame 1776093120_front_185; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e1b5e354d1d8fc82fc5884ca6e5c1dc4e58150fa6803b6de7ea97861072c8d0f.jpg", "reference_inputs/images/486507d95165986ab11002c792174a96f33611f83cc6f7b8dac32459170452d9.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000065", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000065", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094179_left_331; view left\nB: frame 1776094179_front_331; view front\nC: frame 1776094172_right_329; view right\nD: frame 1776094185_front_333; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7714b11ba9830e4abd06caf457aaf139ca317cffdb78fb0725b95de466e7ed8b.jpg", "reference_inputs/images/e2ff9fe947300b8b6d7e2352839f044686d961039c3962f59a384bb9607fe087.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000073", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000073", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776093936_left_254; view left\nB: frame 1776093936_front_254; view front\nC: frame 1776093933_left_253; view left\nD: frame 1776093945_left_257; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0d95713af95df5dd51d10fc9a0cf04b61cb00f31921326ab414f7386e761f921.jpg", "reference_inputs/images/18ed6a07ffb14685c6ec9e0bfae10c31649d52258b80bbe6f660aa4cfe979303.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/city06_ground_aerial_v2_50m/000074", "task_name": "view_transition_reasoning", "dataset_name": "city06_ground_aerial_v2_50m", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/city06_ground_aerial_v2_50m/all_questions.json", "question_id": "000074", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776094300_left_370; view left\nB: frame 1776094309_right_373; view right\nC: frame 1776094315_front_375; view front\nD: frame 1776094309_front_373; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0776293502197f7b148704a6b1090b81b26bfd43e30ce900dd9fba300ea997fe.jpg", "reference_inputs/images/0971963cac7d10ff47e6c2cd804049371eeeef90a1f93390f375c135be709e6a.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000001", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000001", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155296_front_113; view front\nB: frame 1776155289_front_111; view front\nC: frame 1776155289_right_111; view right\nD: frame 1776155278_front_108; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/a20d3d83054c04f71ed6b5c6312b939f05cc9d24b5e6affddb8ee38fddb1e4ee.jpg", "reference_inputs/images/0c2b1f8e768a74e6b58d1ad07b0a499d68fa37e8fddd9d079738c1efe8ba4e15.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000002", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000002", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155872_front_259; view front\nB: frame 1776155849_front_254; view front\nC: frame 1776155856_right_256; view right\nD: frame 1776155856_front_256; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2d51b2d0b65a721925f50de4f9152874e6d2736ecfdad71e9fd38a981c0f927d.jpg", "reference_inputs/images/06fdd514e786700ea7a23da184dd81266d49dc97e2163e66f7514c2f4eea2684.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000003", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000003", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155344_right_127; view right\nB: frame 1776155334_right_124; view right\nC: frame 1776155344_front_127; view front\nD: frame 1776155350_right_129; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/685858ecb859261a7e7bfad1b706afc38c9c7e1fcd340f4472e29512c2eaf1dc.jpg", "reference_inputs/images/fd2c5cf4aa2a59bb4d352409f347fe42269b09e6d4250b85cb1fe76176ada20e.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000005", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000005", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155442_right_157; view right\nB: frame 1776155442_front_157; view front\nC: frame 1776155436_right_155; view right\nD: frame 1776155448_right_159; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/140a9b4bbbbd508392dd2afde7d5a03faed822d428dd5e86048e9743fb4b9244.jpg", "reference_inputs/images/16fa9cf6208ab35fb0d826e48d6332ccd233ceca1dfcf6162507e182eb3bb565.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000006", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000006", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155826_front_247; view front\nB: frame 1776155833_front_249; view front\nC: frame 1776155839_front_251; view front\nD: frame 1776155833_right_249; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4e39da2d888a310aaafae69029530ff5e50dcdf438abd894baece4ea1a117d5d.jpg", "reference_inputs/images/7c0ab268b36f28c9505438130afdcfe4d225defd6b0e9dad757f6995d0ca7686.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000008", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000008", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155817_front_244; view front\nB: frame 1776155823_front_246; view front\nC: frame 1776155830_front_248; view front\nD: frame 1776155823_right_246; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ae13a9c5945de7b02e95c2240e094c09107ee5b85bc2e970146788b55fcef9bf.jpg", "reference_inputs/images/fbc41e957cdbae21b6379cdb6714f9471bb30dd1c1528a014ca95931ed98d96f.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000015", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000015", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155214_left_93; view left\nB: frame 1776155214_front_93; view front\nC: frame 1776155238_left_97; view left\nD: frame 1776155204_left_90; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/451adcef570413c32c7712e668f0ca2b5cf1fd94f59d92bb78b54d31e51dd6e5.jpg", "reference_inputs/images/b2dd64c7e9b1243b03ab736072cd05cee70050a33f8b0cb630cdda598a97ff74.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000016", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000016", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155011_left_45; view left\nB: frame 1776155004_front_43; view front\nC: frame 1776154995_left_40; view left\nD: frame 1776155004_left_43; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/354a9fb45c2fe63d6e95e0f029c8adf65148bc1ce2584a495b7b9053d2d4db5a.jpg", "reference_inputs/images/b7777dae8deeee8c87ad575e40d03c77275ed68ff04babe417c0c207d143b8de.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000018", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000018", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155014_right_46; view right\nB: frame 1776155024_right_49; view right\nC: frame 1776155014_front_46; view front\nD: frame 1776155007_front_44; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b129d9fa075374f861b6b449ded32804a231405fd649654b5d57abc69568319b.jpg", "reference_inputs/images/767039f4d005d1f842c7d5cbf721314ee24853b89c0d171394280de0020cd8e9.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000019", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000019", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155360_front_132; view front\nB: frame 1776155377_right_137; view right\nC: frame 1776155370_right_135; view right\nD: frame 1776155370_front_135; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/378f26489a575ecb49c7c1df8cb29a07bcca46229b38ad1c4a01bad5f2054c14.jpg", "reference_inputs/images/5383d33d7f0f593bcb25c66cbeade56d3e9c8e3a0603c354099e71ad04cf18d6.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000020", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000020", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155417_front_149; view front\nB: frame 1776155417_right_149; view right\nC: frame 1776155398_right_143; view right\nD: frame 1776155426_front_152; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6025ca88abbdcc2195d677f8e1a46f656a2528c3768bd9fb5d51f74b6322d5e9.jpg", "reference_inputs/images/a5e1cada3f1119ca40449a30b4e3ebc45eff5da93bdfa416f111d59c1ef34903.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000021", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000021", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776156281_right_327; view right\nB: frame 1776156274_right_325; view right\nC: frame 1776156274_front_325; view front\nD: frame 1776156267_right_323; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/635ab25ed245c8c152dfed54e61484546ea66ce647d882644b16bb1e3bf60a76.jpg", "reference_inputs/images/1f97d285d8679182d3172ceb6f831a4c20fe3a9abab6e23166fc1be25a84005b.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000023", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000023", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776154995_left_40; view left\nB: frame 1776154995_front_40; view front\nC: frame 1776154986_left_37; view left\nD: frame 1776155004_left_43; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d95145a5355ed1cde86bf2adffdd8b83edac842d4536b5817c57e1cbae3fa96c.jpg", "reference_inputs/images/516e287ec1804b32a2025a8e3f14d0608ab8695bb45fa091fa885cd756ac49fa.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000024", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000024", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155377_front_137; view front\nB: frame 1776155366_front_134; view front\nC: frame 1776155377_right_137; view right\nD: frame 1776155370_front_135; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5ec0e08da2373840cb4038db0dfd0868f4fc033a85850790c289d45ea33bea32.jpg", "reference_inputs/images/83688920d4eeeeeb236ced3091ff9e2e5272ba47fda98fe003c74eb5330d51f5.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000029", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000029", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155823_front_246; view front\nB: frame 1776155806_front_241; view front\nC: frame 1776155817_right_244; view right\nD: frame 1776155817_front_244; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/0316a4fbf1fac31c6a08903b8118c5d0f664e65eb8971e413852394456fb5cab.jpg", "reference_inputs/images/43d9bb45cd24e01d42e77efb8214de2611373827ec67827c65d390ca68e9ee75.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000030", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000030", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155787_front_235; view front\nB: frame 1776155780_front_233; view front\nC: frame 1776155793_front_237; view front\nD: frame 1776155787_right_235; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4888fd8dde17efd43dcac31675860b050a8babab2992066f0f85b05ba50ae50e.jpg", "reference_inputs/images/6eaa5bffe373a65a7df1e985ab43b3bae4ea2a9c8b0dd84e2f5cbbcd4ef071e2.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000033", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000033", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155407_right_146; view right\nB: frame 1776155398_right_143; view right\nC: frame 1776155407_front_146; view front\nD: frame 1776155401_right_144; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/def132f86ad6e70af3e3d8a7a3f67f549a6e0848242bcfada98c522b7aec394f.jpg", "reference_inputs/images/a23d9fa417833110e6a8409304fba6893217309507c6cc96927bbdb7cdd19de1.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000034", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000034", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155787_front_235; view front\nB: frame 1776155773_front_231; view front\nC: frame 1776155780_right_233; view right\nD: frame 1776155780_front_233; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/87322f999e28c03d8c72865649b7da193424c7316db612ec7ed50fa61289cba0.jpg", "reference_inputs/images/35f9a7f828063a684b360bd85b5e536fcda67eb84a43bf2073501753b58a9d29.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000039", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000039", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155360_right_132; view right\nB: frame 1776155360_front_132; view front\nC: frame 1776155347_right_128; view right\nD: frame 1776155353_front_130; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e80d7e3620fd48b979861c0d1b062d4c74ce53fa0b464286e266d04af38e2c05.jpg", "reference_inputs/images/b14fe2e66121de8e295ac39ef8f12c32327799074dd518a625d387ed0eb01c55.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000040", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000040", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155337_front_125; view front\nB: frame 1776155337_left_125; view left\nC: frame 1776155331_front_123; view front\nD: frame 1776155327_front_122; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9ebc2d7bf4d8af9a7c886efbcd8e8dc25a83c1c5d966a68e9235a160dc0c8e09.jpg", "reference_inputs/images/9216acdbcaf9be2b325f3780839a34bc0d4a625aa85841308e709efd973243ed.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000041", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000041", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155314_left_118; view left\nB: frame 1776155321_front_120; view front\nC: frame 1776155321_left_120; view left\nD: frame 1776155310_left_117; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bdf1174f597c13dfc08c4999062e97d6fb909ee43c428ec71bb2e78ac91d0000.jpg", "reference_inputs/images/76417b83f9d21b4fb8017ef57aa0307934c4e1ef887369102c864e0ab60ab9cd.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000042", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000042", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155420_front_150; view front\nB: frame 1776155426_right_152; view right\nC: frame 1776155426_front_152; view front\nD: frame 1776155436_right_155; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/454738419da85780eed3408425ad93b4af64b553f7b914c4685cd6688b605c0f.jpg", "reference_inputs/images/462826f7295466c5b902fe06f7dba03902cad18d17487cba994838144f029968.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000043", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000043", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155790_right_236; view right\nB: frame 1776155790_front_236; view front\nC: frame 1776155796_front_238; view front\nD: frame 1776155783_front_234; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/76096e018b0fde6dbfb53fb28502a84499a530c6dcfc1d4657a19b6594bc19a7.jpg", "reference_inputs/images/a0129d63db46e8eba377249c0179f4014572173374d022f06f9daf7ffd70f3cc.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000044", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000044", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776156128_right_285; view right\nB: frame 1776156118_right_282; view right\nC: frame 1776156135_right_287; view right\nD: frame 1776156128_front_285; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9f9335565c1c29f2dfbec3f237deb438e8a31570755f8f108fcefbc6992f55f3.jpg", "reference_inputs/images/cade71daaceb769b9d83068ff66c4e34105ae3e124f224c0b3edf75f34105d6f.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000048", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000048", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155363_front_133; view front\nB: frame 1776155347_front_128; view front\nC: frame 1776155353_front_130; view front\nD: frame 1776155353_left_130; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2dffc10a00c6943d2cb3527b0992b4a3587eb0eb74145e0eaf9f85a5e7a11d0c.jpg", "reference_inputs/images/d79278cfe4c413cac5890f5ca2665a8ea60c25b762271484ea33967e85ab5048.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000051", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000051", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776156150_right_291; view right\nB: frame 1776156142_right_289; view right\nC: frame 1776156150_front_291; view front\nD: frame 1776156139_right_288; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c1aa5919fb87c97d3fff3583631f433f2a849f95078a621f8157647587c3a998.jpg", "reference_inputs/images/6c64a12c4518966592f29ba0edbb8e96864721048a9c5140592e55d9e31b7de7.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000055", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000055", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776156271_front_324; view front\nB: frame 1776156281_front_327; view front\nC: frame 1776156271_right_324; view right\nD: frame 1776156264_front_322; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/4dadc32b67e2a92b77f94f28242c869f37dbd7b3af67d1aa3cc1aec58a04355d.jpg", "reference_inputs/images/87d4925f76d1157e466e9f12f42e71ebb31b2f16c3c806d1f0bf18ef8ef9cd65.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000056", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000056", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776156215_front_307; view front\nB: frame 1776156215_right_307; view right\nC: frame 1776156208_right_305; view right\nD: frame 1776156204_right_304; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/21ba9b614ccede8c53cf5f9976a5c858a878c89dfca0825245a3673f62edd3a5.jpg", "reference_inputs/images/6de269a8a2a3f4aafd6348f185617f2636a659210d87ea8f1a0fcb7ef935af5f.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000058", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000058", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155856_front_256; view front\nB: frame 1776155872_left_259; view left\nC: frame 1776155872_front_259; view front\nD: frame 1776156019_left_262; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/eecb49a01dcbef342a4549da6b5a41c423fb63e2e0879c463aa4fd0dda1dd437.jpg", "reference_inputs/images/6e256298b6cfa2c1bffc17b40645b8f29b7ee7ba08570408d6bc3e6e2edaac0e.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000061", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000061", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155748_right_223; view right\nB: frame 1776155761_right_227; view right\nC: frame 1776155754_right_225; view right\nD: frame 1776155754_front_225; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/d7d3acdd155a929153facccd31cfe2b573f0336468f894dea82b66c21204e111.jpg", "reference_inputs/images/3a331b4040ae1bf56b8cb17c8870402a7ec96d1b8b1edf2cbaef76b37ad41695.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000065", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000065", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155398_right_143; view right\nB: frame 1776155410_front_147; view front\nC: frame 1776155410_right_147; view right\nD: frame 1776155401_right_144; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/b1da0be79c88bc772c758eb4dffc54d8da159dbf41db273cfbb2383de3b936a7.jpg", "reference_inputs/images/00b17a5b2db78651b827f2b07e7b8fbf222d5f6a7d2276cb4581d538164deb7e.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000066", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000066", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155600_right_181; view right\nB: frame 1776155590_right_178; view right\nC: frame 1776155600_front_181; view front\nD: frame 1776155580_right_175; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8d82d2f9b27fa42bc2850c0eda89aaa3438b8115aa5b7b2075a2e07ed25328dc.jpg", "reference_inputs/images/9f71b5195e18b5d2d5ea8ca81ebfcc0995f82a0247bdf847224436c36812aba3.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000069", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000069", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155347_right_128; view right\nB: frame 1776155344_right_127; view right\nC: frame 1776155356_front_131; view front\nD: frame 1776155356_right_131; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/7f56b97527bec15d36b72f876ca9fa8f89a1243a9400108e700252b0a4057dd7.jpg", "reference_inputs/images/dc7e71f4e199170a0a3ff97af0bec11d2b71394ff12c4542bf859bb7a103d859.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000070", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000070", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155088_front_69; view front\nB: frame 1776155147_left_73; view left\nC: frame 1776155078_left_66; view left\nD: frame 1776155088_left_69; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/041f77b080f78da211c5b9e4354573707c5bea28acff5062720ebf0dc840b358.jpg", "reference_inputs/images/2344ebae2d969e28ff886f1b8f03c49ffdebb0de1e3ceb075be6a80d63a0263a.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000073", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000073", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155211_left_92; view left\nB: frame 1776155238_left_97; view left\nC: frame 1776155219_left_94; view left\nD: frame 1776155219_front_94; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/bedc32ddb3655bfacd46b38cb360f5fe494b1cd6b1eb604298d6e7e20a43f727.jpg", "reference_inputs/images/eec1bc9cd4b684bdacba2ab18847cffe6d06f59ababd292d938c014d0e377f7f.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000075", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000075", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155436_front_155; view front\nB: frame 1776155436_right_155; view right\nC: frame 1776155430_front_153; view front\nD: frame 1776155426_front_152; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/6fd6badd1842bc493e61a4e55d825ed70087b8fc58ac0adddec9c952dfdd45d4.jpg", "reference_inputs/images/fca24a2645fdfb1b92699de3302112c1c74253d1469771661703733f2a2db8a3.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000086", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000086", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155241_right_98; view right\nB: frame 1776155211_right_92; view right\nC: frame 1776155234_front_96; view front\nD: frame 1776155234_right_96; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/381663fff987c1b9f06c9a79b12c12dde19a44a7abb9d4e313ec6f5159245bcd.jpg", "reference_inputs/images/c00ec3ca9d0a78a008c3b98f7074d6be381a6142397e3281b8094384071e9f2f.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000087", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000087", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155204_right_90; view right\nB: frame 1776155204_front_90; view front\nC: frame 1776155194_front_87; view front\nD: frame 1776155231_front_95; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ff6aa7a1455b16dcaa3adba7ac86fb55547c22a8997d1fd944cb5fbc863cb9c2.jpg", "reference_inputs/images/2b24372e27e2bd704a32937cc8ba4e08eb66a8d17ca0271fe8197d1fda5afcda.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000088", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000088", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776156070_front_271; view front\nB: frame 1776156093_front_275; view front\nC: frame 1776156093_right_275; view right\nD: frame 1776156073_front_272; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/edd6c9163a9fe0a87ff8ad2e3cff5e5ad7fe7dd2d36526aefea97402963c2907.jpg", "reference_inputs/images/c176561f79e92d8e9b475bc240f8381fde5b83c610b3ed70c0cba05de2f33d2a.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000090", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000090", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155651_left_194; view left\nB: frame 1776155665_left_198; view left\nC: frame 1776155665_front_198; view front\nD: frame 1776155655_left_195; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/3385e9fbe817bcfe72568526bafc1c66ecaea3ff8d5ce9ed5c3d86043a248103.jpg", "reference_inputs/images/174e6d8ba0a5059b86b284aa7f2c6aedc1a72e43079ee1ed43fcd85775c0aea4.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000093", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000093", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155231_left_95; view left\nB: frame 1776155197_front_88; view front\nC: frame 1776155197_left_88; view left\nD: frame 1776155191_left_86; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/825e90476ee3fe7ce7072ff71173589fcdf60c69ef6fca5d2b1b91eaa419954d.jpg", "reference_inputs/images/b89d4ac06819b40a09e3360f4f816607c5724f14895f3b70356cff07e2134c3a.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000096", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000096", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155387_front_140; view front\nB: frame 1776155398_front_143; view front\nC: frame 1776155387_left_140; view left\nD: frame 1776155384_front_139; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/c5e9f263933f58eca163817551e81024ce9f0a4ed8608d81d4db4a518e3ac2b4.jpg", "reference_inputs/images/b825eb2826a323b6794d04b3b5eac2bc5d72cb41e0b4a5da6519579b7179cd75.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000101", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000101", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155423_front_151; view front\nB: frame 1776155417_front_149; view front\nC: frame 1776155423_right_151; view right\nD: frame 1776155436_right_155; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/1fcbcfe5e580a6b5ef261be74f18d9acf99e68ce26525e0a6a114e2221b99a4f.jpg", "reference_inputs/images/365a64c128897218503111653b8a462ee10fa0958504ae1985ecde819c92b1eb.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000106", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000106", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776154986_left_37; view left\nB: frame 1776154986_front_37; view front\nC: frame 1776154995_left_40; view left\nD: frame 1776154979_left_35; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/ccfc7c3198f4eed362b231083cccdf5d6f125736701f2a2a07cbba324b102e46.jpg", "reference_inputs/images/41effa7d61b270d1392d32c21879a54d96e6eea81cb9ffbbb30c3fef76e8a6aa.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000110", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000110", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155603_right_182; view right\nB: frame 1776155635_right_189; view right\nC: frame 1776155609_right_184; view right\nD: frame 1776155609_front_184; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/786b61c806fa53b8e0d41b8007ee248ef3ede0696a3c6066f0ef6ffa44cde876.jpg", "reference_inputs/images/ef8cc189c27b1cf4cf9f17d6d369457957b74fa94883569451a5c7b7c9ee48e0.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000113", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000113", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155317_right_119; view right\nB: frame 1776155327_right_122; view right\nC: frame 1776155337_right_125; view right\nD: frame 1776155327_front_122; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/34ac0cddf33737a0b9263eb1953f295037f00ec9a519d5b2ecb1172fd7922dec.jpg", "reference_inputs/images/2fe9020fdfa6810e50908f03dd2ca133915c3cab0fedb1a6ce82afdeb67a6b34.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000114", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000114", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155748_front_223; view front\nB: frame 1776155754_front_225; view front\nC: frame 1776155748_right_223; view right\nD: frame 1776155741_front_221; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/2b7d22548a2fdf96d2f6dc15295c250a0bf82705b7a7ec8f8251428823930c46.jpg", "reference_inputs/images/a85530851c47fe9a9523c552b9ab1cd101633d3e3bbd4b25ae9aea78f40d29c2.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000115", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000115", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155439_right_156; view right\nB: frame 1776155445_front_158; view front\nC: frame 1776155439_front_156; view front\nD: frame 1776155430_front_153; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5ed78fc184349bf6b8acbc8b5132818db96c331956e99eb10c86b705883a6adb.jpg", "reference_inputs/images/60d552cc870ac0c73344a3de2dd0482420886f4a0ed63fb796d9dc5fec76ecd1.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000116", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000116", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776156172_front_296; view front\nB: frame 1776156172_right_296; view right\nC: frame 1776156180_right_298; view right\nD: frame 1776156158_right_293; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/e2e5928d38f7d3013e79fd8a7e3c11db7806301ca5bcfe73bc17294ce5fc88b6.jpg", "reference_inputs/images/e0cdba13011b11ebc06eec5a1067f5b288fdc258865ab1ca9887bee4972405b3.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000121", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000121", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155658_front_196; view front\nB: frame 1776155665_front_198; view front\nC: frame 1776155672_front_200; view front\nD: frame 1776155665_right_198; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/5deedcd50de2d0b671aa7412451bf0aff32a298924ad98278ce7d442ebe620d4.jpg", "reference_inputs/images/7b95581dd7d7ac97f5dd5222610737850aecd451855f3e8890c03eae176e910e.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "D", "answer_index": 3}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000133", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000133", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776156301_left_333; view left\nB: frame 1776156294_front_331; view front\nC: frame 1776156301_front_333; view front\nD: frame 1776156308_left_335; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/8e99cd02ce1446268e9c21934c9ef5862b6d78c936f598705c3e91c60a499c03.jpg", "reference_inputs/images/ab5aa75da89f6c19751cbd17b146e7a2614bd9ac2f94cb50cb50e2bd6bc9b071.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "A", "answer_index": 0}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000137", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000137", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155241_front_98; view front\nB: frame 1776155241_right_98; view right\nC: frame 1776155231_right_95; view right\nD: frame 1776155234_right_96; view right\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/cc75a3343929f033c815bd4178493463929361d3d3cc594dd62a3aefd54a85f9.jpg", "reference_inputs/images/1cd1ad00a0c90f9e3d900f66299feb30b303c216b3e49e3988b6c35b650b4254.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "B", "answer_index": 1}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000138", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000138", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776154948_front_26; view front\nB: frame 1776154933_front_24; view front\nC: frame 1776154948_left_26; view left\nD: frame 1776154925_left_23; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/10bbb3dff09ccb7abe8a353391e4998ae0b4870b9a14e90a05d481e13c718460.jpg", "reference_inputs/images/a512e7d88a1270191297e6aa852bdafc2fdfc8fbc0dc9d52264369e551106909.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000142", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000142", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776155722_front_215; view front\nB: frame 1776155728_front_217; view front\nC: frame 1776155728_right_217; view right\nD: frame 1776155718_front_214; view front\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/908236d819204a1f12ca8765f2a0cbd61391ea6dbd8de348e83bdbd4a35ccf83.jpg", "reference_inputs/images/96d677f82c5b63f3379f4f798e7b12122b86affbe4ae58780a8044f98d924fda.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}
{"sample_id": "view_transition_reasoning/town1_ground_aerial_v2/000145", "task_name": "view_transition_reasoning", "dataset_name": "town1_ground_aerial_v2", "source_json": "benchmark/Benchmark_Questions_filtered/view_transition_reasoning/town1_ground_aerial_v2/all_questions.json", "question_id": "000145", "input_profile": "mcq_text_reference_v1", "prompt": "Task: View Transition Reasoning.\nYou will receive exactly two images from the same scene.\nImage 1 is the aerial query image with the marked ground teammate position and facing direction.\nImage 2 is the panel of candidate ground-view image options, and each tile is already labeled by option letter.\n\nYou are Aerial Agent, observing the scene from the aerial view. Image 1 is your aerial view. It marks your ground teammate's position with label G and shows the teammate's facing direction with an arrow. Image 2 shows candidate ground-view options captured from ground level and labeled A, B, C, D. Select the option that best matches what your ground teammate would most likely see from the marked position and facing direction. Use the aerial layout, local scene geometry, nearby structures, road/sidewalk context, and likely foreground occlusion around the marked viewpoint. Answer with exactly one option letter from A, B, C, D.\n\nImage usage note:\n- Use Image 1 to infer the teammate's local ground-level viewpoint from the marked position and heading.\n- Use Image 2 only as the answer-option panel; each labeled tile is one candidate ground view.\n- The hidden ground reference image is not provided during evaluation.\n\nAnswer options:\nA: frame 1776154904_front_17; view front\nB: frame 1776154894_left_14; view left\nC: frame 1776154904_left_17; view left\nD: frame 1776154891_left_13; view left\n\nTask: choose the one option whose image best matches the likely ground-level view at the marked position and heading.\nBenchmark mode instruction: respond with only one option letter from A, B, C, D. Do not output any explanation or extra text.", "images": ["reference_inputs/images/9bd0348a7160007920d805ab0fd3cdae1b1a682344fb32fcfc5ed34a2142bebc.jpg", "reference_inputs/images/582933f3c904fb0c9bdae594e22d6110f9286b298cdb1a28b67b97e443ee1e3a.jpg"], "image_roles": ["marked_aerial_query", "candidate_ground_panel"], "image_sizes": [[1024, 576], [1024, 698]], "option_labels": ["A", "B", "C", "D"], "answer": "C", "answer_index": 2}