ise-ice-lab/PSI_VQA

PSI-VQA — AI City Challenge 2026, Track 3

PSI-VQA is a video question-answering benchmark built on the PSI 2.0 dataset, covering egocentric dashcam footage of pedestrian crossing scenarios. It is the OOD Test Set 2 for AI City Challenge 2026, Track 3: Driver Situation Awareness.

The dataset spans four complementary tasks, all unified under the tao-vl-reason-v1.0 schema (NVIDIA TAR Benchmark format). Each item pairs a short video clip with a question; the model must return a structured answer.

Task	Train	Test
Binary Crossing QA (BCQ)	245	55
Open QA	354	126
Multiple Choice QA (MCQ)	321	91
Temporal Localization	227	56
Total	1147	328

---

Tasks

1. Binary Crossing QA (BCQ)

Setting: A pedestrian is visible in the first few seconds of a dashcam clip. Annotators reached a clear majority consensus on whether this pedestrian intends to cross the road.

Input: Short video clip + natural-language question asking whether the pedestrian intends to cross in front of the recording car.

Output: "Yes" or "No" (exactly, no period).

File: train/bcq.json (245 items) | test_public/bcq_questions.json (55 items)

---

2. Open QA

Setting: A pedestrian's crossing intent was genuinely ambiguous — annotators disagreed. Three sub-questions are asked per video, one for each intent direction (cross, not cross, uncertain).

Input: Video clip + question asking for visual cues that support a specific intent direction.

Output: A bulleted list of visual cues (e.g., "- The pedestrian stepped off the curb.\n- The pedestrian made eye contact with the driver."), or "None" if no cues support that direction.

File: train/open_qa.json (354 items = 118 videos × 3 sub-questions) | test_public/open_qa_questions.json (126 items)

---

3. Multiple Choice QA (MCQ)

Setting: Same ambiguous-intent videos as Open QA. The question tests understanding of the pedestrian's specific behavioral or intentional state.

Input: Video clip + question with four labeled options (A–D) and the instruction "Answer with a single letter."

Output: A single letter: "A", "B", "C" or "D". Only the leading letter is scored; any option text after it is ignored.

File: train/mcq.json (321 items) | test_public/mcq_questions.json (91 items)

---

4. Temporal Localization

Setting: Egocentric dashcam clips from PSI 2.0. The task is to identify the time interval during which a road user or road factor most influences the driver's decision-making.

Input: Video clip (15 s single-cluster, or shorter sub-clip for multi-cluster scenarios) + natural-language question.

Output: JSON string with start and end timestamps: {"start": "MM:SS", "end": "MM:SS"} (sub-second precision allowed, e.g., "00:04.93").

The response must be the JSON object alone, with no commentary before or after it. A surrounding ```json code fence is tolerated; trailing text is not, and is scored as unparseable. MM:SS, MM:SS.mm and HH:MM:SS are all accepted for both fields.

Scoring is temporal IoU against the ground-truth interval. A prediction with start > end is not reordered, a prediction that does not overlap the ground truth scores 0, and any unparseable response scores 0. All of these are included in the mean.

File: train/temporal_localization.json (227 items) | test_public/temporal_localization_questions.json (56 items)

---

Submission Format

Submissions are a two-column CSV: item_index and prediction.

item_index,prediction
bfaa0b67a0385860,Yes
41dcf77f800ebcae,No
d1538f8c200b0bb4,"A) - The pedestrian is already in the roadway or has stepped into the road."
b22a0fcaac174951,"- The pedestrian stepped off the curb."
c3f9a12d44e7b081,"{""start"": ""00:04"", ""end"": ""00:07""}"

The item_index for each test question is provided in the test_public/*_questions.json files. All four tasks may be combined in a single submission CSV.

Per-task prediction format:

Task	Expected prediction
BCQ	"Yes" or "No"
Open QA	Bulleted cue list, or "None"
MCQ	Single letter A–D
Temporal Localization	{"start": "MM:SS", "end": "MM:SS"}

Evaluation is handled by the AI City Challenge evaluation server. The challenge submission portal is at https://www.aicitychallenge.org/2026-track3/.

---

Directory Structure

PSI_VQA/
├── README.md
├── train/
│   ├── bcq.json
│   ├── open_qa.json
│   ├── mcq.json
│   ├── temporal_localization.json
│   └── videos/
│       ├── clear/          # BCQ videos (clear-intent clips)
│       ├── ambiguous/      # Open QA and MCQ videos (ambiguous-intent clips)
│       └── temporal/       # Temporal localization clips (15 s + sub-clips)
└── test_public/
    ├── bcq_questions.json
    ├── open_qa_questions.json
    ├── mcq_questions.json
    ├── temporal_localization_questions.json
    └── videos/
        ├── clear/
        ├── ambiguous/
        └── temporal/

All annotation JSON files follow the tao-vl-reason-v1.0 schema:

{
  "format": "tao-vl-reason-v1.0",
  "metadata": { "type": "annotation", "task": "<task_name>", ... },
  "media_root": null,
  "items": [
    {
      "video_id": "PSI/<subdir>/<filename>.mp4",
      "question": "<natural-language prompt>",
      "answer": "<task-specific format>",
      "reasoning": "<annotation provenance or empty>"
    }
  ]
}

video_id paths are relative to the media_root. When loading locally, set media_root to the split root (e.g., train/videos/ for train items, test_public/videos/ for test items).

---

Access and License

This dataset is released under the TASI Benchmark Data Sharing Agreement (inherited from PSI 2.0). Use is restricted to academic and non-commercial research.

Full license text: TASI Benchmark Data Sharing Agreement

Source: PSI 2.0 / PSI 3.0 — NCSU Intelligent Cognitive Ergonomics Lab

---

Acknowledgments

Dataset curation is supported by the National Science Foundation under Award #2516587.

Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

---

Citation

If you use this dataset, please cite the PSI paper and acknowledge AI City Challenge 2026:

@inproceedings{jing2025psi,
  title={PSI: A Benchmark for Human Interpretation and Response in Traffic Interactions},
  author={Jing, Taotao and Chen, Tina and Tian, Renran and Chen, Yaobin and Domeyer, Joshua and Toyoda, Heishiro and Sherony, Rini and Ding, Zhengming},
  booktitle={Advances in Neural Information Processing Systems},
  volume={38},
  year={2025},
  url={https://proceedings.neurips.cc/paper_files/paper/2025/hash/436fb0fa57c75e0d2063b5bc19a21da1-Abstract-Datasets_and_Benchmarks_Track.html}
}

@misc{aicity2026track3,
  title={{AI City Challenge 2026, Track 3: Driver Situation Awareness}},
  year={2026},
  note={[citation to be added upon release]}
}