You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
These frames are decoded from YouTube videos of the Stereo4D dataset. By requesting access you confirm that you will use them only for non-commercial research, will not redistribute them, and accept the terms of the original Stereo4D release and of the video owners.
Log in or Sign Up to review the conditions and access this dataset content.
Stereo4D frames used by TrajQA-B
The exact video frames used to train and evaluate Sparse Trajectory Tokens for Metric Motion Queries in
Multimodal Language Models: for each of the 42,151 processed Stereo4D clips, the rectified left-eye frames at
512×512 (%05d.jpg, adaptive stride, 20–38 frames per clip), the camera-only gt_stereo4d.npz (camera poses and
timestamps) and stereo4d_meta.json. Scenes, trajectory tokens and questions are in the companion dataset
TrajQA-B; the code repository extracts these frames deterministically from the official Stereo4D release
(byte-identical), so this repo only saves that step.
Access is gated: the frames are derived from YouTube videos and are provided for non-commercial research use only, under the terms of the Stereo4D release; please do not redistribute them.
Layout
frames/frames-NN.tar (64 shards, ≈1.6 GB each) contain <clip>/…; frames/INDEX.tsv maps each clip to its shard.
The shard index of a clip is the same as in TrajQA-B/scenes/, so shard NN of both repos covers the same clips.
MANIFEST.tsv lists the SHA-256 of every tar.
mkdir -p data/stereo4d/frames && for t in frames/frames-*.tar; do tar xf "$t" -C data/stereo4d/frames; done
- Downloads last month
- 106