PerrinLin/handdiffuse-full

HandDiffuse — full-view face-mosaicked videos

This edition contains 115 paired episodes, 255,845 two-hand timesteps and 5,505 processed videos referenced by the existing accepted camera/time mappings. It also contains 86 separate historical motion-only groups (172 NPY files, 314,958 two-hand timesteps) under motion_only/. Left and right hands are counted together once per timestep. Historical group names do not establish independent capture identity.

Videos are face-mosaicked, re-encoded as H.264 and contain no audio stream. Dimensions and decoded frame count are preserved. Native-speed outputs retain source timestamps; new 4K60 half-speed outputs are checked against exactly doubled source timestamps, preserving every frame index. Selected motion arrays are unchanged. Original calibration values and namespace/ID bindings are preserved. Known excluded or unmapped cameras and nine unresolved paired candidates are absent. BEAT and seven unknown-origin groups are excluded. MANO model weights are not distributed.

episodes.json indexes the paired data. view_scope.json states the evidence carried for each retained camera: calibration/time mapping acceptance is not exhaustive visual-quality certification. The 115 previously selected single views have carried sampled review records; other views are retained as processed mapped views with the existing documented limitations. No new frame-map integrity audit is performed during this preparation.

Eight Ninja episodes retain explicitly approximate adjacent-frame alignment, including selected and alternative indices/PTS. 1207_random_5 and 1207_random_23 retain 51-column motion, original rest joints and same-version FBX companions. Other selected paired motion is 61-column. No new camera fit or invented beta is used.

Public availability does not assert an open dataset license. A formal license and citation have not been supplied. release_manifest.json contains relative paths, sizes and known hashes; checksums.sha256 is generated after upload hashing and remote verification and includes the processed videos.

Privacy processing and upload status

This repository is being populated incrementally. The planned total is 5,505 videos; consult upload_status.json for verified uploaded counts. Original unmasked videos are never uploaded here. A .privacy.json sidecar beside each processed video records source/output hashes, encoding and validation.

Face detection uses YuNet and CenterFace every five frames for native-speed videos, or every ten frames for new 4K60 half-speed videos. Detected face boxes are padded and their locations and sizes interpolated for masking every output frame. Visually reviewed adjustments may cover detector misses or remove hand/equipment false positives. Boundary positions can be held through occlusions. Sparse detection and sampled visual review are not a mathematical guarantee against all missed faces. Hand protection takes precedence: separate convex hulls of recorded MANO projections are expanded outward by 20 original-image pixels. Original source pixels are restored inside that region while face mosaic continues outside. Faces can remain visible where they overlap protected hands. Isolated unmapped boundary frames may remain original, as recorded in each sidecar. Unreviewed clips retain all detector tracks; automated checks are distinguished from sampled visual reviews in each sidecar. No identity recognition is performed.

Episode layout

episodes/<episode_id>/
  metadata.json
  calibration.json
  motion/
    poses_left.npy
    poses_right.npy
    joints_left.npy             # original rest joints for selected 51-column motion
    joints_right.npy
    output_left.fbx             # optional verified same-version source companion
    output_right.fbx
  cameras/
    primary_<original_id>/
      <original_video_id>.mp4
      frames.csv
    full60_<original_id>/       # optional supplementary original-ID namespace
      <original_video_id>.mp4
      frames.csv
  quality/
    boundary_frames.csv        # optional recorded boundary screen
    local_motion_events.json   # optional recorded local quality events

Actual optional files and camera directories are specified in metadata.json. Read those references rather than guessing filenames, pairing sorted directories, or assuming camera IDs are contiguous. A camera's video ID and calibration parameter ID are separate fields. A supplementary namespace may reuse a numeric ID without referring to the primary camera with that ID.

The package provides processed videos and preserves selected motion arrays. Excluded or unavailable rows are represented by metadata and masks; they are not silently removed, interpolated, or renumbered.

Read an episode

Use Python 3.10 or later. NumPy is required for numeric motion loading; the metadata and CSV reader uses the standard library. Place handdiffuse_reader.py on the Python import path.

from handdiffuse_reader import Episode

# Run from the package root containing episodes/.
episode = Episode(".", "tutting_19")
left = episode.load_motion("left")   # read-only memory map, [motion_rows, 51 or 61]
right = episode.load_motion("right")

camera_uid = next(iter(episode.cameras))
camera = episode.cameras[camera_uid]
original_parameters = episode.calibration(camera_uid)
video_path = episode.path(camera["video_file"])

for row in episode.iter_frame_map(camera_uid):
    if not row.available:
        # This motion row has no source frame in this particular camera.
        continue
    if not row.passes_recorded_screens:
        # Application choice: skip rows covered by recorded conservative screens.
        continue
    video_index = row.video_index
    source_presentation_time = row.video_pts * row.video_time_base
    # Check row.has_approximate_timing when exact motion/video correspondence matters.
    # Approximate rows expose row.alternative_video_index and its source PTS.
    motion_left = left[row.motion_index]
    motion_right = right[row.motion_index]
    # Decode the recorded source video index with a decoder of your choice.

passes_recorded_screens combines source availability with the recorded boundary and local motion screens. It does not certify unreviewed frames, anatomical accuracy, hidden fingers, physical contact, or freedom from collision. Flags are retained on each yielded row. Boundary and local exclusions are combined as an index union; overlapping reasons do not remove two rows.

Iterate each frame map to exhaustion to check its complete row count and agreement with the metadata's missing-source indices. Array shape/dtype checks and annotation bindings are checked on reading. The reference reader does not calculate file-content hashes, decode video content, reconstruct MANO meshes, or select a different motion version.

Motion parameters

All slices below use Python's zero-based, end-exclusive convention.

Columns	Meaning
0:3	Saved source root/wrist position
3:51	Sixteen axis-angle rotations, in radians: root orientation followed by fifteen local joint rotations in source MANO order
51:61	Ten stored shape coefficients, present only in 61-column arrays

The saved rotation vectors represent the full pose rather than PCA coefficients. For the documented 61-column reconstruction, use the saved pose and beta without adding another mean-pose rotation. Given raw posed model vertices and joints, bind the posed wrist to the stored position:

vertices_source = vertices_raw - joints_raw[0] + motion_row[0:3]
joints_source = joints_raw - joints_raw[0] + motion_row[0:3]

Retain each episode's documented reconstruction convention. Physical length units and absolute world-axis semantics have not been independently established for every capture family; retain source coordinates and calibration translation values together instead of applying an assumed scale or recentering operation.

A 51-column file contains no beta coefficients. Do not append zeros or infer a shape silently. Use its supplied original rest joints for forward kinematics, or use the optional original same-version FBX geometry and skinning. Matching FBX animation keys do not establish surface equality with an arbitrary MANO renderer or pose-corrective implementation.

MANO model weights are not included in the planned assets. Obtain any required model through its original provider under the applicable terms. This documentation does not grant a model license.

Camera parameters

calibration.json stores original values grouped by parameter namespace. metadata.json binds each listed camera to a namespace and original parameter ID. Calibration may contain additional original entries; an entry's presence does not by itself mean that its video or frame correspondence has been accepted.

For an original RT entry reshaped to 3 by 4, the convention is:

X_camera = RT[:, :3] @ X_source + RT[:, 3]
homogeneous_pixel = K @ X_camera
pixel = homogeneous_pixel[:2] / homogeneous_pixel[2]

Keep the original image size and distortion values, and follow the recorded image convention when applying distortion. K and RT are stored as flat numeric arrays in the current schema. Camera values and original IDs are retained; no newly fitted camera parameters are substituted by the reference reader.

Frame correspondence

Each camera has its own frames.csv with:

Field	Meaning
motion_frame_index	Row in both selected hand arrays, starting at zero
video_frame_index	Decoded source-video frame index, starting at zero; empty when unavailable
video_pts	Original decoded presentation timestamp; empty when unavailable
video_time_base	Rational units per timestamp tick
video_timestamp_seconds	Recorded timestamp in seconds
source_frame_available	Optional explicit 0/1 availability flag
alignment_certainty	Present on accepted approximate maps: APPROXIMATE_ADJACENT_FRAME_SELECTION
alternative_video_frame_index	Adjacent candidate source index on approximate maps
alternative_video_pts	Decoded PTS of the alternative, using the same time base
alternative_video_timestamp_seconds	Alternative decoded timestamp in seconds

Do not infer a universal motion FPS, a fixed downsampling factor, or capture timing from the FBX export clock. Empty source indices and timestamps remain explicit. A repeated or uneven source sampling interval is not corrected by inventing frames. The existence of a row mapping does not establish physical subframe exposure synchronization.

Eight Ninja episodes (8, 9, 13, 14, 17, 19, 22, 27) use owner-delegated approximate alignment. Each has one fixed default offset shared by its six cameras, plus an explicit adjacent alternative. The original extraction indices remain unrecovered; these candidates are not a proven hard timing-error bound. Camera metadata records alignment_certainty, both offsets, and exact_extraction_index_recovered: false. The reader exposes this on every Frame; passes_recorded_screens concerns motion-quality screens and does not remove or certify approximate timing. Applications requiring uniquely supported motion-to-video indices should filter row.has_approximate_timing. Original camera parameters and motion values remain unchanged.

Version and quality information

The selected pair is described by the episode metadata. Selection can favor an older motion version: filename recency and smoothness alone do not establish correspondence or hand quality. Quality annotations are bound to the selected motion hashes; do not reuse them with a different version.

Some episodes have conservative boundary or local correspondence screens. An interval can include unshown or uncertain transitions; it does not mean every included row is individually confirmed erroneous. Preserve event descriptions and their scope when reporting downstream filtering.

Remaining curation limitations include missing multi-view source groups, unresolved camera/time correspondences, unknown provenance for additional motion candidates, and sparse rather than exhaustive geometry review. The draft's candidate and camera coverage counts are not a complete-release or whole-episode quality claim.

Citation and licensing

A dataset license and formal citation have not been supplied. Public availability does not by itself grant an open license. No license is invented by this preparation. Keep source identities and machine-specific curation paths out of the public package.

Per-video playback clocks

From the half-speed processing update onward, remaining 4K videos recorded near 60fps retain every frame and play at half speed (60 to30fps;59.94 to29.97fps). Duration doubles. Already-published native-speed files retain their recorded policy; always inspect each video's .privacy.json. The original frames.csv, motion arrays, and calibration remain source-clock evidence. Frame indices are unchanged. For half-speed output, multiply source PTS/time by2, or read the accompanying <video>.playback_frames.csv, which explicitly binds source/output PTS and the unchanged MANO row indices. No motion data is resampled. New face detection samples every10 retained frames with interpolation; sideways supplemental detection remains every30 frames.