Datasets

Motion dataset lab

Upload one or more videos, run a layer11_pose job, and download the dataset/ ZIP. Many videos at once (e.g. 30+) can use a lot of RAM — if the engine stops mid-job, run smaller batches (e.g. 10–15 clips) and restart run_server.py between runs. Set DATASET_ENGINE_URL=http://127.0.0.1:8787 in .env.local — the browser uses the same URL (via next.config.js) so jobs hit the local engine directly; the /api/dataset-engine proxy uses that variable too. AI suggestions use OPENROUTER_API_KEY on this Next.js server.

Stub engine: python run_server.py in Datasets/motion_dataset_engine — restores sample output when bulk_*_dataset/dataset exists.

Video file(s)

One or more videos share the same export path: merged dataset/global_stats.json (per-video rows + totals) and data.jsonl with source_video_index. A single file uses the same batch layout (one entry under videos[]).

Suggest action / pipeline / stride (first frame) — uses the first decoded frame from your video and OPENROUTER_API_KEY on the Next.js server. Review before starting the job.

High-quality export (optional)

Filters frames before writing data.jsonl. Rejected frames go to dataset/low_quality_frames.jsonl. See export_quality_report.json for formulas. Set lower-body min to 0 to disable that gate. Presets set thresholds and base weights; action_label nudges weights (e.g. walking vs upper-body).