h2o-annotations-v1

Annotations only — no images, no video.

VITRA-style hand episodes for H2O (Two Hands Manipulating Objects), with per-hand instructions and paraphrases.

episodes	5,845
training samples (index_frame_pair rows)	200,332
annotation	MANO pose + world/camera joints + per-frame extrinsics
text	one instruction per episode + 1.86 paraphrases on average
images / video	not included — see Getting the frames below

What we did

Episodes are ours. The source release ships either raw video or differently-segmented clips, so we re-cut it with VITRA's method — speed minima of the 3D wrist in world space:

gaussian smooth (sigma=1.0) -> local speed minima in a fixed window (win=15, i.e. 0.5 s
at 30 Hz) -> merge runs shorter than min_seg=16 -> pad 2 frames on each end

sigma and win are quantities in time, converted per source frame rate. Left and right hands are cut independently, with the other hand's motion ignored.

Instructions are ours. Two rounds, both with Qwen3.5-122B-A10B-FP8: round 1 captions 8 frames per episode with the palm's future trajectory drawn on them; round 2 checks the sentence belongs to that hand, strips same-hand references ("Rinse the right hand." -> "Rinse the hand.", because training already prepends Left hand: ... Right hand: ...), and writes 1-3 paraphrases.

Episodes with no instruction are not included. Round 1 returns N/A when an episode shows no object interaction. Those episodes are excluded from both the archive and the index, so every episode here has a usable instruction.

Files

h2o.tar                 ->  Annotation/h2o/episodic_annotations/*.npy
episode_frame_index.npz        index_frame_pair (N,2) uint32 + index_to_episode_id (E,)

index_frame_pair row number is the sample id: row r = (episode ordinal, frame within that episode). len(index_frame_pair) is the size of the training set.

import numpy as np
# tar -xf h2o.tar
z = np.load("episode_frame_index.npz", allow_pickle=True)
ep_slot, frame_id = z["index_frame_pair"][sample_id]
eid = str(z["index_to_episode_id"][ep_slot])
d = np.load(f"Annotation/h2o/episodic_annotations/{eid}.npy", allow_pickle=True).item()
rgb_frame_id = int(d["video_decode_frame"][frame_id])

Each .npy is a dict with video_name, video_decode_frame, intrinsics, per-frame extrinsics (world->camera), anno_type (which hand this episode is for), text, text_rephrase, and a left/right dict holding beta, hand_pose, global_orient_worldspace, transl_worldspace, joints_worldspace, kept_frames. text[hand] = [(sentence, (0, T))] and text_rephrase[hand] = [([paraphrases...], (0, T))].

Getting the frames

video_decode_frame indexes the source video, which we do not redistribute. Get it from H2O (Two Hands Manipulating Objects) — https://taeinkwon.com/projects/h2o/, then decode by index (we use decord; a self-maintained sequential counter drifts silently if the decoder ever skips a frame).

Known limitations

• Paraphrase count averages 1.86, not a fixed number. Past 3 the model starts inventing; a sentence with no prepositional phrase honestly supports only one or two.
• Verified: the index lists exactly the episodes that have an instruction, every episode's stored frame count matches its index rows, and no index entry points at a missing episode.

---

Revision — 2026-09-08 (re-cut)

This release replaces the previous one. The previous episodes contained invalid frames and should not be used.

A defect in our episode-cutting step let frames with kept_frames == False (invalid hand pose — all-zero or NaN wrist coordinates) stay inside published episodes. The validity mask was only used to keep a cut point from landing on an invalid frame; it did not constrain what a segment contained. Worse, a run of invalid frames could suppress cutting altogether, so the gap was swallowed into one long segment instead of being excluded.

Every episode here is now built from a run of consecutive valid frames, so kept_frames is all-True by construction — verified over the whole collection: 568,369 episodes / 16.5 M frames, zero kept_frames == False. Both instruction rounds were regenerated for the new segmentation.

Segment counts and episode ids therefore changed, and the index and splits were rebuilt:

	previous	this release
episodes on disk	7,792	7,803
episodes published (with an instruction)	5,696	5,845
training samples	196,941	200,332

Splits are video_test (test owns whole videos disjoint from train/val; train and val share the remaining videos and are split at the episode level), balanced on frames at 90/5/5, seed 1.

dataset	episodes	training samples	our contribution	size	HF
EPIC-KITCHENS-100	149,570	4,019,534	episodes + text	8.70 GB	epic30-annotations-v1
EgoTouch	107,364	3,123,675	episodes + text + tactile	17.02 GB	egotouch-annotations-v1
GigaHands	70,486	2,266,087	episodes + text	2.89 GB	gigahands-annotations-v1
Ego-Exo4D	67,051	1,757,474	text only	4.09 GB	egoexo4d-annotations-v1
OakInk2	29,058	1,052,924	episodes + text	1.56 GB	oakink2-annotations-v1
TACO	23,757	736,136	episodes + text	1.34 GB	taco-annotations-v1
HOT3D	18,805	619,680	episodes + text	1.51 GB	hot3d-annotations-v1
ARCTIC	12,610	425,796	episodes + text	0.85 GB	arctic-annotations-v1
H2O	5,845	200,332	episodes + text	0.40 GB	h2o-annotations-v1
Tachin	2,669	79,584	episodes + text + tactile	3.38 GB	tachin-annotations-v1
total	487,215	14,281,222		41.7 GB

Something-Something V2 was dropped from the collection (12 fps against 30 fps everywhere else, so a 16-step action chunk spans 1.33 s instead of 0.53 s). The repository still exists but should not be used.