hyperscene_traj

Camera walkthroughs of annotated SketchUp indoor scenes: a surface-normal video, cameras, a caption and RGB per clip, 832x480, 24 fps, 124 frames. A clip's key is <scene>-<traj>; the two batches' trajectories differ, so never join them on key.

path	content
v0_1009/raw/	10,219 clips, webdataset: normal.mp4, instance.mkv, instance.json, traj.npz, meta.json (six-section caption)
v0_1009/select_final.json	the filtered clip list v0_1009/raw/ was built from
v0_1009/train/, v0_1009/val/	H3-Train Ref2VA first-frame split, 2,764 / 33 clips
v0_1009/lynnreal/<scene>/<scene>.tar	LynnReal-Omni RGB of 5,154 v0_1009/raw/ clips, <key>.mp4, no audio
v0_1009/physics_sim_test.tar	first physics-simulation test set
v0_1009/gpt-6/	archived normal-caption RGB experiment
raw_v2_1016/	20,044 clips, webdataset: normal.mp4, traj.npz, meta.json (photoreal caption), Cycles rgb.mp4 for some
flythrough/3dfront_wo_caption/	826 3D-FRONT flythroughs, 90 s each (2,164 frames), horizontal FOV 110°; no caption, no RGB. Webdataset as in raw_v2_1016/: normal.mp4, traj.npz (with its own intrinsics, fx = fy = 291.2863), meta.json (scores from the trajectory evaluation); key front_<3D-FRONT id>-fly_01
v1_1016/lynnreal/<scene>/<scene>.tar	LynnReal-Omni RGB of all 20,044 raw_v2_1016/ clips, <key>.mp4 with an AAC track
v1_1016/rel_rope/raw.safetensors	rel_cam_rope coverage of 20,042 raw_v2_1016/ clips, v1_1016/train's prope_path: the instances each token shows and their pixels, cameras, world-axis 3D boxes; Y up, metres
v1_1016/scene_preproc/	structure only of the 5,350 sk_hq.txt scenes, <scene>.mesh.meshopt.zst (meshopt + zstd, Y up, metres): every instance, backdrop and outlying piece removed; backdrops.json, outliers.json list the removed faces as indices into the full scene mesh
v1_1016/train/	H3-Train Ref2VA first-frame split of the 18,656 raw_v2_1016/ clips with a bbox video, two controls: bbox.safetensors (every box a translucent solid, composited per pixel in depth order), normal.safetensors
physics_sim_v2.tar	physics-simulation test set, 30 clips
sk_scene_shards/	the 14,299 source scenes, GLB and instance annotations, losslessly coded; decode with its sk_decode.py
sk_hq.txt	5,350 high-quality scene ids
floor_seg_hq.json	storeys of the sk_hq.txt scenes: height-histogram peaks grouped floor, layer_N, roof, each group noting whether the storey above it holds an instance; selects each group's largest peak, two in a middle group
floor_seg_hq_v2.json	the same peaks; selects the lowest floor peak and the highest of every other group
tools/	captioning tools

Normals are world-space in the axes of traj.npz's pose (Y up, OpenCV camera-to-world), stored as RGB = (n*0.5+0.5)*255, black where the ray misses; fx = fy = 554.3622, cx = 415.5, cy = 239.5 (flythrough/ clips: see their traj.npz intrinsics). instance.mkv holds instance id + 1 per pixel, 0 for none.