README.md

metadata

license: cc-by-nc-4.0
pretty_name: VoxMem EPIC-KITCHENS share
tags:
  - egocentric
  - epic-kitchens
  - depth
  - camera-pose
  - video-qa

VoxMem EPIC-KITCHENS share

This share holds derived data of the VoxMem project, built from EPIC-KITCHENS-100. The videos are (c) the EPIC-KITCHENS team under CC BY-NC 4.0, and this share keeps that license: it is for non-commercial research. The original MP4s are not here.

Current version: 2026-10-18 (v3). It holds:

• the question set (frozen 2026-10-12), both tiers (accurate and rough), including the questions built with Gemini;
• one label row per question;
• the trainable lists;
• the token cache, 599 videos (the 572 of the previous version + 27 new ones, poses from EPIC-Fields only);
• the hand masks, the reference-frame tokens and their exact-frame 3D, and the fps table.

The voxels are not shared. The training machine builds them from the token cache with one script.

To train, use the code, not this page:

• the guide is docs/TRAIN_ELSEWHERE_EPIC_V3.md in the VoxMem repo, branch hdepic-gen-epic-questions;
• bash scripts/kit/download_epic_v3.sh downloads, checks and unpacks everything below into place.

Every path here is the path under the VoxMem data root (src/voxmem/paths.py), except the two reftok-4b / refgeo-4b tars, which unpack into <code>/results/cache/hdepic/.

v3 (2026-10-18)

path	what
epic-kitchens/tokens/qwen3vl-4b_512px/<video>.npz	599 files (3.43 TB), one per video. Each holds the Qwen3-VL-4B visual tokens (144 per 512x288 frame), similarity features, 3D points, camera pose, depth and the frame-rule table. Every frame is chosen by the frame rule: each wanted 1 fps frame exactly, else the nearest frame with a camera pose within +-0.5 s, else that second is dropped. sha256 in epic-kitchens/tokens/tokens_MANIFEST.tsv and SHA256SUMS_tokens.txt.
epic-kitchens/questions/questions_2026-10-18.tar	the question set: 2026-10-18/<type>/{accurate,rough}.jsonl for 25 question types, plus LIST.{csv,md} (also loose as LIST_2026-10-18.{csv,md}: per type, its counts, whether Gemini was involved, its sources) and reference_images.json
epic-kitchens/questions/labels_2026-10-18.jsonl	one row per question. The main columns: tier, gemini, split, coverage, n_frames, scene_id, voxel_tokens, est_tokens, trainable, reason (the first rule failed) and trainable_ref_train.
epic-kitchens/questions/train_2026-10-18.json + _split.json	the 68,465 trainable questions (64,439 train / 4,026 val, on 27 validation videos)
epic-kitchens/questions/train_2026-10-18_reftrain.json + _split.json	the 67,070 of them whose every reference image has 3D (for --ref_pos train)
epic-kitchens/hand_masks/hand_masks_2026-10-18.tar	box + palm hand masks, 599 videos each (unpacks to epic-kitchens/hand_masks/{box,palm}/). They supersede the loose 572-video hand_masks/{box,palm}/ files, whose rows are identical.
epic-kitchens/meta/fps_table_2026-10-18.tar	Videos/<P>/{frames,durations}.txt for the 599 videos: the voxel builder's fps table
epic-kitchens/reference_frames/reftok-4b_2026-10-18.tar	24,727 reference-image token files (144 x 2560, or 192 x 2560 for the two 4:3 videos)
epic-kitchens/reference_frames/refgeo-4b_2026-10-18.tar	23,015 exact-frame 3D files, one per reference image: per-token xyz and depth, camera pose, source. They come from the frame's EPIC-Fields pose, or a pose interpolated within 0.5 s, plus DA3 depth; otherwise from the nearest posed cached frame within 0.5 s.
epic-kitchens/reference_frames/ref512_jpg_2026-10-18.tar	the 512 px reference images the tokens were made from (needed only to re-bake them)
epic-kitchens/v3_MANIFEST.tsv	bytes, md5 and sha256 of every v3 file above except the tokens

Renamed on 2026-10-19 so every v3 file carries one date. The question files were named after the day the questions were frozen: questions_2026-10-12.tar, LIST_2026-10-12.{csv,md}, labels_2026-10-12.jsonl and train_1012[_reftrain][_split].json. Only the names changed. Every loose file is byte-identical, and the tar holds the same 65 files under the folder 2026-10-18/ instead of 2026-10-12/. The old files remain in this repo's history.

HD-EPIC reference geometry (added 2026-10-20)

path	what
hd-epic/reference_frames/hdepic_refgeo-4b_2026-10-18.tar	2,847 files (22 MB), one per reference image of the HD-EPIC VQA questions (2,650 questions). Each holds the camera pose (Aria SLAM world), the depth of each token and each token's MEAN 3D point (12 x 12 tokens of a 384 px, 95-degree rectified frame). md5 65381813477968f5deef7aebb04e88e1

• Who needs it: only someone TESTING on HD-EPIC a model trained with ref_pos=train. Unpack it into <code>/results/cache/hdepic/ (it creates refgeo-4b/).
• One recording is empty: the 51 references of P03-20240217-131219 hold NaN, because its released SLAM poses stop at 22.8 of 54.6 minutes. None of its questions is in a scored set.
• Where it comes from: HD-EPIC (Perrett et al., CVPR 2025). Cite it and follow its terms.
• The rest of the HD-EPIC test data is not public: no HD-EPIC images, videos or tokens are in this share.

Earlier versions (kept, not used by v3)

path	what
epic-kitchens/frames/<video>.tar	rectified 992 px frames of the 2026-10-06 version, <video>/<frame:06d>.jpg, 572 videos
epic-kitchens/questions/questions.tar	the 2026-10-06 question set (all.json, covered.json, split_videos.json)
epic-kitchens/train/train.tar	the v1 training lists (text_only/epic_v1_*, reference_frames/epic_v1_refzs_*)
epic-kitchens/reference_frames/reference_frames.tar	the v1 reference images
epic-kitchens/raw/*.tar	EPIC-Fields poses, Intermediate-Data (per-frame poses), VISOR and the EPIC annotations
<category>/outdated-2026-09-24/	the 2026-09-24 version (old layout: its members start with task_runtime/)

MD5SUMS.txt lists every tar and loose v3 file, and SHA256SUMS_tokens.txt lists every token file. bash download.sh <DATA> checks both and unpacks the tars that belong under the data root.