Request access to the E0 debug archives

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

These archives redistribute original clips from HOIGen-1M (gated on the Hub) and from Ego-Exo4D (released under its own licence agreement), beside SA-V and DROID clips (CC BY 4.0). Access is granted individually and only to people who use each source's clips under that source's terms and do not pass them on.

Log in or Sign Up to review the conditions and access this dataset content.

E0 action-progress debug archives

Debug snapshots of the E0 data pipeline of Whose Time Is It? Instance-Specific Action-Progress Control for Video Editing. Each archive is a self-contained zip with the same layout; every text field and progress curve inside is a model proposal (Qwen3-VL-32B-Instruct under a frozen prompt, SAM 2.1 masks), with no human review.

Earlier sample archives (removed 2026-09-24)

The eleven sample archives of the method versions v1 to v4 that this repo carried until 2026-09-24 (hoigen_debug_native_20260917.zip, the 100-clip v2 and v3 HOIGen samples, the 500-candidate v3 samples of SA-V, DROID and Ego-Exo4D in both actor rules, and the 50-clip v4 samples of all four sources) were removed when the per-source datasets below replaced them. Their content is superseded: the datasets below hold the same clip layout for far more clips under the v4 method. The archives still exist on the project's compute nodes.

Datasets (one folder per source, 2026-09-24)

The Todi campaign of 2026-09-24 is published as one dataset per source, all in the same layout, split into zip parts of at most 10 GB so that no single download is larger than that. Each folder <name>/ holds <name>-partNNN.zip (the clips: source video, frames and masks of a clip always in the same part), <name>-labels.zip (labels/, index.jsonl, README.md, the run records), and beside them <name>-index.jsonl, <name>-README.md and <name>-manifest.json (every file with size and SHA-256) to read before downloading anything. index.jsonl names the part of every clip. The exports ran as 48 to 192 parallel shards, so parts are small (1 to 2 GB) and many rather than few and 10 GB.

Folder Source Method and screen Pool decided Processed clips Parts Size
hoigen_v4gemma/ HOIGen-1M (whole clips, 4 to 10 s, 44 archives of the release) v4 detector-first; screen and judgements by Gemma 4 31B (google/gemma-4-31B-it, rev 842da379) 133,014 of 148,294 21,576 576 parts of at most 0.92 GB + labels 280.7 GB
droid_v4todi/ DROID (one 4 s window per episode) v4 detector-first; Qwen3.8-27B (Qwen/Qwen3.8-27B, rev 1d4bf0f2, no think block) 18,120 of 18,120 4,669 48 parts of at most 1.6 GB + labels 71.2 GB
egoexo_v4gemma/ Ego-Exo4D ego view (12 seeded 4 s windows per take) v4 detector-first; Gemma 4 31B 32,964 of 60,096 at the stop 2,835 64 parts of at most 1.6 GB + labels 74.6 GB

The two screen models differ a lot: Gemma 4 accepted about 15% of HOIGen clips and 10% of Ego-Exo4D windows where Qwen3.8-27B without its think block accepted 3.4% and 2%, with Gemma's accepted set almost containing Qwen's (HOIGen, 5,960 shared clips: 117 both, 761 Gemma only, 5 Qwen only). DROID was screened by Qwen3.8-27B before the switch and kept. Every text field is a model proposal; no human has reviewed any clip.

Layout of every archive

  • README.md with the conventions, index.jsonl with one row per processed clip (source, clip id, window start, duration, effective frame rate, actors with text fields and frame-0 box, 16 knot progress scores with validity, confidence and expected score, acceptance verdict, envelope fractions, relative paths), pipeline_log.jsonl, screen.jsonl, pipeline_summary.json.
  • labels/<clip>.json: the full label per clip (raw score words, per-knot distributions, extractor and segmenter configuration, dense-tracking record, discarded actors, timings).
  • clips/<clip>/source.mp4 for every processed clip: the original video, byte for byte at its native frame rate, verified against the SHA-256 recorded at extraction. Nothing is re-encoded or resampled.
  • clips/<clip>/frames/frame_00..48.png (512x320, aspect kept, grey padding), valid_image_mask.png, clip.json with the resize-and-padding transform and the selected source frame ids; clips/<clip>/masks/actor_<k>/affected_masks/ (the 49 working frames), .../dense_masks/ (every native frame of the window, with dense_masks.json timestamps and hashes), static_envelope.png, and the same for combined/.

Terms per source

  • HOIGen-1M clips: gated on the Hub; use under the HOIGen-1M terms, do not redistribute.
  • SA-V clips: CC BY 4.0.
  • DROID episodes: CC BY 4.0.
  • Ego-Exo4D clips: released under the Ego-Exo4D licence agreement; the two egoexo_* archives are available only to holders of their own Ego-Exo4D licence and must be used under its terms.

This repo is gated with manual approval for these reasons. Request access with the two boxes above; the labels and masks are the project's own work and carry no further terms.

Downloads last month
87