ToucheBench

Reference data, geometry annotations and semantic checklists for hand- and camera-conditioned egocentric video evaluation. This package contains 779 reference clips. Model predictions, answer sheets and leaderboards are excluded.

Subset	Clips	Purpose	Checklist
q1_q4	200	HOI4D hand PCK / object box and mask IoU	Objective metrics
q2	10	Moving-camera interactions	Final v1, 60 binary questions
q3	24	8 scenes × 3 alternative endings	Final v1, 144 binary questions
egolife300	300	Video-quality holdout	Holdout protocol
egostandard200	200	Video-quality holdout	Holdout protocol
hot3d45	45	Video-quality holdout, native 81@15 alignment	HOT3D protocol

Q1 and Q4 share the same 200 clips. Question counts are not clip counts. Holdouts are separate from Q1–Q4. The source declares EgoLife/EgoStandard holdouts disjoint from training; this packaging audit does not independently recheck the training corpus.

Layout

data/<subset>/<clip_id>/
  rgb.mp4                 # packed reference RGB, never a model prediction
  hand_control.mp4        # visualization of hand conditioning
  hand_motion.npz         # motion features
  camera.npz              # camera conditioning
  ...                     # subset-specific labels / first frame / original
annotations/
  <subset>.jsonl          # one reference clip per row, root-relative paths
  schema.json
  npz_schema.json         # actual array keys, shapes and dtypes
  protocols/*.json
  assets/umetrack_profile.json
checklists/
  q2.json                  # final English v1, 60 questions
  q3.json                  # final English v1, 144 questions
  q3_gt_cross_v1.json     # human GT cross-check, not a model score

All manifest paths resolve from the repository root. IDs use ASCII filenames. Original public-dataset sequence/window identifiers remain in metadata for reproducibility. Q2/Q3 recording filenames, dates and internal absolute paths are removed. The original source archive and private mapping are kept outside this release folder.

Data and annotation conventions

Packed RGB/control videos use 153 frames at 24 fps, 768×512. Frames are zero-indexed. Media and numerical annotations retain source values; object-dtype string labels in NPZ are converted to Unicode arrays so arrays can be read with allow_pickle=False. Motion features may be frame-aligned (153 steps) or latent-aligned (20 steps); preserve anchor_source_frame_ids when present. NPZ files are not a common geometry schema: consult annotations/npz_schema.json for each subset/file type. Estimated hand/camera controls are distinct from the official HOI4D reference labels.

Q1/Q4 include hand_j2d.npz, object.npz, object_mask.npz and full original source videos. Use packed-window frame mappings rather than assuming the full video starts at the benchmark window. Q1 uses PCK@0.05/0.10/0.20; Q4 uses object box/mask IoU. The source protocol does not fully specify every metric implementation detail (for example PCK normalization and invalid-label aggregation); consult the evaluation implementation before reproducing headline numbers.

Q2 retains source camera-quality flags, including rejected tracks. Q2 and Q3 checklists are final v1, confirmed by the dataset owner. Q2 questions and reference descriptions are translated into English; Q3 question meanings, roles and separability labels are preserved; two horizontal-line glyphs are expressed with their existing English descriptions. Some Q3 scenes have partial/low separability; see the GT cross-check annotations. Scene and ending labels are in English, and checklist clip IDs join the annotation manifests.

Semantic evaluation

checklists/q2.json and checklists/q3.json are standard JSON documents with metadata and a clips array. Each clip joins the annotation manifest through id. All checklist content is English.

Each Q2/Q3 clip has three easy and three hard binary questions. A pass means the question predicate holds, including negatively phrased predicates. Use the candidate RGB video, with frame 0 for context and frames 49, 80, 112, 152 for inspection; do not judge the control render. Report the applicable protocol and checklist version with pass rates. No model answers are provided in this package.

Q2 hard roles: S end state, C camera arrival, A interaction. Q3 hard roles: S target ending, X exclusion of alternatives, A interaction. Q3 uses 8 prefix latents (49 prefix pixels, about 2.04 s); the source teacher protocol instead starts a new video at original GT frame 49. These evaluation settings must be reported separately.