Upload H2O shared-v1.0 (part 88)

H2O/shared-v1.0/script/check_dexycb_projection.py

@@ -0,0 +1,31 @@
+#!/usr/bin/env python3
+"""Check saved Dex-YCB source 3D joints against saved source 2D joints in sampled views."""
+import json
+from pathlib import Path
+import sys
+
+import numpy as np
+import torch
+
+root = Path(sys.argv[1])
+report = json.loads((root / 'doc/validation/STRUCTURAL_AUDIT.json').read_text())
+checks = []
+for name in report['sampled_views']:
+    view = root / name
+    load = lambda path: torch.load(view / path, weights_only=True, map_location='cpu').numpy()
+    k = load('camera/intrinsic.pt')
+    xyz = load('annotation/keypoints/xyz.pt')
+    uv = load('annotation/keypoints/uv.pt')
+    valid = load('annotation/keypoints/xyz_valid.pt') & load('annotation/keypoints/uv_valid.pt')
+    z = xyz[..., 2]
+    valid &= np.isfinite(xyz).all(axis=-1) & np.isfinite(uv).all(axis=-1) & (z > 0)
+    z = np.where(valid, z, 1)
+    pred = np.stack((k[:, 0, 0, None, None] * xyz[..., 0] / z + k[:, 0, 2, None, None],
+                     k[:, 1, 1, None, None] * xyz[..., 1] / z + k[:, 1, 2, None, None]), axis=-1)
+    errors = np.linalg.norm(pred[valid] - uv[valid], axis=-1)
+    checks.append(dict(view=name, points=int(errors.size), median_px=float(np.median(errors)) if errors.size else None,
+                       p95_px=float(np.percentile(errors, 95)) if errors.size else None))
+(root / 'doc/validation/KEYPOINT_PROJECTION_QC.json').write_text(json.dumps({
+    'scope': 'Saved source 3D joints projected with saved color intrinsics vs saved source 2D joints; independent of hand mesh rendering.',
+    'views': checks}, ensure_ascii=False, indent=2) + '\n')
+print(json.dumps(checks, ensure_ascii=False), flush=True)

H2O/shared-v1.0/script/config.json

@@ -0,0 +1,24 @@
+{
+  "root": "/home/xuzhefeng/Datasets/Hand-Datasets-Raw/H2O",
+  "out": "/home/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-process16-pipe",
+  "release_id": "shared-v1.0",
+  "limit": null,
+  "max_frames": null,
+  "mano_dir": "/home/xuzhefeng/Projects/WiLoR/mano_data",
+  "views": null,
+  "jpeg_full_check": false,
+  "workers": 4,
+  "view_backend": "processes",
+  "camera_pipeline": true,
+  "ffv1_slices": 16,
+  "view_workers": 16,
+  "ram_stage_dir": "/dev/shm/xuzhefeng-h2o-full-process16-20260930",
+  "ram_stage_gib": 32,
+  "stream_media": true,
+  "codec_threads": 10,
+  "decode_threads": 10,
+  "depth_threads": 4,
+  "subjects": null,
+  "include_casa": false,
+  "archive": null
+}

H2O/shared-v1.0/script/convert_dexycb_v1.0.py

@@ -0,0 +1,349 @@
+#!/usr/bin/env python3
+"""Dex-YCB source annotations → CONTRACT.md v1.0. No derived annotations."""
+import io
+import re
+import shutil
+import tarfile
+import tempfile
+from pathlib import Path
+
+import numpy as np
+from PIL import Image
+import yaml
+
+from warehouse_v1_0 import (View, build_record, find_file, locator, mano_models, mesh_asset,
+    model_profile, parser, pca_to_axis_angle, release, roots, safe_member, sequence_dir)
+
+IMG_W, IMG_H = 640, 480
+
+class _Loader(yaml.SafeLoader):
+    pass
+
+
+_Loader.add_constructor("tag:yaml.org,2002:python/tuple",
+                        lambda l, n: tuple(l.construct_sequence(n)))
+
+
+def yaml_load_bytes(data: bytes):
+    return yaml.load(io.BytesIO(data), Loader=_Loader)
+
+
+class Calibration:
+    def __init__(self, dexycb_root: Path):
+        self.intrinsics = {}   # serial -> {fx,fy,cx,cy,w,h}
+        self.extrinsics = {}   # group  -> {serial: [12] 行优先 3×4 cam→world}
+        self.betas = {}        # mano_calib_name -> [10]
+        calib_dir = dexycb_root / "calibration"
+        if calib_dir.is_dir():
+            self._load((p.as_posix(), p.read_bytes())
+                       for p in sorted(calib_dir.rglob("*")) if p.is_file())
+        else:
+            tar_path = dexycb_root / "calibration.tar.gz"
+            if not tar_path.exists():
+                raise FileNotFoundError(f"no calibration/ nor calibration.tar.gz in {dexycb_root}")
+            with tarfile.open(tar_path, "r:gz") as tf:
+                self._load((m.name, tf.extractfile(m).read()) for m in tf if m.isfile())
+
+    _RE_INTR = re.compile(r"(?:^|/)intrinsics/(\d+)_\d+x\d+\.yml$")
+    _RE_EXTRA = re.compile(r"(?:^|/)extrinsics_(\d{8}_\d{6})/extrinsics\.yml$")
+    _RE_MANO = re.compile(r"(?:^|/)mano_(\d{8}_\d{6}_subject-\d{2}_right)/mano\.yml$")
+
+    def _load(self, files):
+        for rel, data in files:
+            m = self._RE_INTR.search(rel)
+            if m:
+                c = yaml_load_bytes(data)["color"]
+                self.intrinsics[m.group(1)] = {"fx": float(c["fx"]), "fy": float(c["fy"]),
+                                               "cx": float(c["ppx"]), "cy": float(c["ppy"]),
+                                               "w": IMG_W, "h": IMG_H}
+                continue
+            m = self._RE_EXTRA.search(rel)
+            if m:
+                raw = yaml_load_bytes(data)["extrinsics"]
+                self.extrinsics[m.group(1)] = {
+                    str(s): [float(x) for x in np.asarray(v).ravel()] for s, v in raw.items()}
+                continue
+            m = self._RE_MANO.search(rel)
+            if m:
+                self.betas[m.group(1)] = [float(b) for b in yaml_load_bytes(data)["betas"]]
+
+
+
+def calibration_locator(path, member, **extra):
+    if path.is_dir():
+        return locator(path / member.removeprefix('calibration/'), **extra)
+    return locator(path, member, **extra)
+
+
+def prepare_models(root, destination):
+    archive = find_file(root, 'models.tar.gz')
+    if archive is None:
+        return {}, None
+    pattern = re.compile(r'^models/(\d{3}_[a-z0-9_]+)/(textured_simple\.obj(?:\.mtl)?|texture_map\.png)$')
+    names = set()
+    with tarfile.open(archive, 'r|gz') as tf:
+        for member in tf:
+            match = pattern.fullmatch(member.name)
+            if member.isfile() and match:
+                names.add(match[1])
+                path = safe_member(destination, '/'.join(match.groups()))
+                path.parent.mkdir(parents=True, exist_ok=True)
+                path.write_bytes(tf.extractfile(member).read())
+    if len(names) != 21:
+        raise ValueError('expected all 21 source YCB model directories')
+    return {i: destination / name for i, name in enumerate(sorted(names), 1)}, archive
+
+
+def convert_view(path, args, seq, serial, frames, meta, meta_loc, calib, calib_path, models, assets, asset_archive):
+    total = int(meta.get('num_frames', max(frames) + 1))
+    ids = list(range(min(total, args.max_frames) if args.max_frames else total))
+    for f in ids:
+        frames.setdefault(f, {})
+    if not ids:
+        raise ValueError('no RGB frames')
+    source_rows = []
+    for f in ids:
+        row = frames[f]
+        source_rows.append(dict(source_frame_id=f, sync_id=f'{seq}:{f}',
+            source_locator=row['color'][1] if 'color' in row else (meta_loc | {'source_frame_id': f, 'missing_rgb': True}), annotation_matches={
+                'labels': {'source_frame_id': f, 'source_timestamp_ns': None, 'clock_domain': None,
+                           'match_method': 'index_map', 'delta_ns': None, 'source_locator': row['labels'][1]}
+                for _ in [0] if 'labels' in row},
+            component_sources={k: v[1] for k, v in row.items()}))
+    view = View(path, 'dex-ycb', args.release_id, seq, serial, source_rows,
+                sampling='dense' if ids == list(range(ids[0], ids[-1] + 1)) else 'sparse')
+    view.valid_rules = {
+        'mano': 'Source pose_m present and not all zero; side model/calibration available',
+        'keypoints': 'Corresponding joint_2d/joint_3d present; whole-hand all -1 is missing, broadcast to its 21 points',
+        'pose': 'Source pose_y for this ycb_ids slot present and not an all-zero matrix',
+        'segment': 'Source seg array provided for this frame; region codes and entity references validated',
+        'relations': 'Source sequence ycb_grasp_ind resolves through ycb_ids; broadcast sequence grasp to selected frames',
+    }
+    view.full_jpeg = getattr(args, 'jpeg_full_check', False)
+    view.meta['view_type'] = 'exo'
+    view.meta['sampling']['selection'] = {'max_frames': args.max_frames, 'views': args.views}
+    subject = re.search(r'subject-(\d+)', seq)
+    if subject and 1 <= int(subject[1]) <= 10:
+        index = int(subject[1])
+        view.meta.update(split='val' if index == 7 else 'test' if index in (8, 9) else 'train',
+            split_scheme='Dex-YCB/s1', split_source_locator={
+                'url': 'https://github.com/NVlabs/dex-ycb-toolkit/blob/master/dex_ycb_toolkit/dex_ycb.py',
+                'rule': 's1 unseen-subject: 01-06/10 train, 07 val, 08-09 test'})
+        view.meta.pop('split_reason', None)
+    view.rgb([frames[f]['color'][0] if 'color' in frames[f] else None for f in ids], {'frames_ref': 'camera/frames.jsonl'})
+    n = len(ids)
+    intr = calib.intrinsics.get(serial)
+    K = None if intr is None else np.array([[intr['fx'], 0, intr['cx']], [0, intr['fy'], intr['cy']], [0, 0, 1]])
+    ext = calib.extrinsics.get(str(meta['extrinsics']), {}).get(serial)
+    E = None
+    if ext is not None:
+        E = np.eye(4)
+        E[:3] = np.asarray(ext).reshape(3, 4)
+        E = np.linalg.inv(E)
+    view.camera(K, calibration_locator(calib_path, f'calibration/intrinsics/{serial}_640x480.yml'), E,
+        calibration_locator(calib_path, f'calibration/extrinsics_{meta["extrinsics"]}/extrinsics.yml',
+                conversion='inverse of source T_world_from_camera'))
+    sides = meta.get('mano_sides')
+    if sides not in (['left'], ['right']):
+        raise ValueError(f'expected exactly one anatomical source hand: {sides}')
+    side = sides[0]
+    slot = ['left', 'right'].index(side)
+    cal_name = meta['mano_calib'][0] if isinstance(meta['mano_calib'], list) else meta['mano_calib']
+    labels = [dict(np.load(frames[f]['labels'][0])) if 'labels' in frames[f] else {} for f in ids]
+    label_loc = {'frames_ref': 'camera/frames.jsonl', 'source_field': 'component_sources.labels'}
+    motion = np.full((n, 2, 61), np.nan, np.float32)
+    mv = np.zeros((n, 2), bool)
+    if side in models and any('pose_m' in lab for lab in labels):
+        for i, lab in enumerate(labels):
+            if 'pose_m' not in lab:
+                continue
+            pose = np.asarray(lab['pose_m']).reshape(51)
+            if not np.any(pose):  # Official Dex-YCB missing-pose sentinel, independent of joints/seg.
+                continue
+            motion[i, slot] = np.r_[pose[:3], pca_to_axis_angle(pose[3:48], models[side]),
+                                     calib.betas[cal_name], pose[48:51]]
+            mv[i, slot] = True
+        view.tensor('mano', 'mano/axis_angle61/v1', 'annotation/mano/motion.pt', motion, mv,
+                    ['T', 'hand', 'parameter'], label_loc, model_profile=model_profile(models),
+                    calibration_source=calibration_locator(calib_path, f'calibration/mano_{cal_name}/mano.yml'),
+                    provenance={'kind': 'format_conversion', 'method': 'source PCA @ hands_components + hands_mean; explicit axis-angle consumed with flat_hand_mean=true'})
+    elif any('pose_m' in lab for lab in labels):
+        view.absent('mano', 'Source PCA exists; provide --mano-dir to expand it', 'dependency_missing')
+    else:
+        view.absent('mano', 'No source pose_m field')
+    for source_key, key, dim, unit in [('joint_2d', 'uv', 2, 'pixel'), ('joint_3d', 'xyz', 3, 'm')]:
+        if not any(source_key in x for x in labels):
+            view.absent('keypoints_' + key, 'No source ' + source_key)
+            continue
+        points = np.full((n, 2, 21, dim), np.nan, np.float32)
+        valid = np.zeros((n, 2, 21), bool)
+        for i, lab in enumerate(labels):
+            if source_key not in lab:
+                continue
+            p = np.asarray(lab[source_key]).reshape(21, dim)
+            points[i, slot] = p
+            valid[i, slot] = not np.all(p == -1)  # Official sentinel covers the whole source hand.
+            # An individual off-image point (-1, -1) is not evidence of missing annotation.
+        view.tensor('keypoints_' + key, f'keypoints/{key}/v1', f'annotation/keypoints/{key}.pt',
+                    points, valid, ['T', 'hand', 'keypoint', 'coordinate'], label_loc,
+                    keypoint_schema='mano21/v1', unit=unit)
+    ycb_ids = [int(x) for x in meta['ycb_ids']]
+    for j, oid in enumerate(ycb_ids):
+        entity = f'ycb_{oid:02d}'
+        view.object(entity, meta_loc, source_object_id=oid, category_id=oid)
+        poses = np.full((n, 4, 4), np.nan, np.float32)
+        valid = np.zeros(n, bool)
+        for i, lab in enumerate(labels):
+            if 'pose_y' in lab:
+                pose = np.asarray(lab['pose_y'])
+                if pose.shape != (len(ycb_ids), 3, 4):
+                    raise ValueError('pose_y shape differs from ycb_ids')
+                if np.any(pose[j]):  # Official all-zero missing object pose sentinel.
+                    poses[i] = np.eye(4)
+                    poses[i, :3] = pose[j]
+                    valid[i] = True
+        if any('pose_y' in lab for lab in labels):
+            view.tensor('pose_' + entity, 'object/rigid/v1', f'annotation/object/{entity}/pose.pt',
+                    poses, valid, ['T', 'row', 'column'], label_loc,
+                    convention='T_camera_from_object', unit='m')
+        else:
+            view.absent('pose_' + entity, 'No source pose_y field')
+        if oid in assets:
+            mesh_asset(view, entity, assets[oid], assets[oid] / 'textured_simple.obj',
+                locator(asset_archive, f'models/{assets[oid].name}/textured_simple.obj'))
+        else:
+            view.absent('mesh_' + entity, 'models.tar.gz unavailable', 'not_downloaded')
+    if any('seg' in lab for lab in labels):
+        mask_rows = []
+        width, height = view.meta['camera']['image_size']
+        for i, lab in enumerate(labels):
+            values = np.unique(lab.get('seg', np.zeros((height, width), np.uint8)))
+            mappings = []
+            for value in values:
+                value = int(value)
+                if value == 0:
+                    continue
+                if value != 255 and value not in ycb_ids:
+                    raise ValueError(f'unknown Dex-YCB region value {value}')
+                mappings.append(dict(mask_value=value, category_id='hand' if value == 255 else value,
+                    entity_id='hand_' + side if value == 255 else f'ycb_{value:02d}', part_id=None,
+                    source_instance_id=value))
+            mask_rows.append(dict(frame_index=i, labels=mappings))
+        view.raster('segment', (lab.get('seg', np.zeros((height, width), np.uint16)) for lab in labels),
+                    ['seg' in lab for lab in labels], label_loc, mask_rows)
+        view.meta['components']['segment'].update(coverage={'kind': 'selected_entities',
+            'entities': ['hand_' + side] + [f'ycb_{o:02d}' for o in ycb_ids]}, source_zero_meaning='background')
+    else:
+        view.absent('segment', 'No source seg field')
+    if any('aligned_depth_to_color' in frames[f] for f in ids):
+        def depths():
+            width, height = view.meta['camera']['image_size']
+            for f in ids:
+                item = frames[f].get('aligned_depth_to_color')
+                if item is None:
+                    yield np.full((height, width), np.nan, np.float32)
+                else:
+                    with Image.open(item[0]) as im:
+                        yield np.asarray(im).astype(np.float32) / 1000.0
+        view.raster('depth', depths(), None, {'frames_ref': 'camera/frames.jsonl',
+            'source_field': 'component_sources.aligned_depth_to_color', 'conversion': 'millimetres to metres'})
+    else:
+        view.absent('depth', 'No source aligned_depth_to_color frames')
+    grasp = meta.get('ycb_grasp_ind')
+    if grasp is not None:
+        if not 0 <= int(grasp) < len(ycb_ids):
+            raise ValueError('invalid ycb_grasp_ind')
+        relation = dict(relation_type='grasp', subject_id='hand_' + side,
+            object_id=f'ycb_{ycb_ids[int(grasp)]:02d}', status='positive', source_locator=meta_loc,
+            provenance={'kind': 'source', 'method': 'sequence ycb_grasp_ind resolved through ycb_ids; not per-frame contact'})
+        view.relations([{'frame_index': i, 'relations': [relation]} for i in range(n)], np.ones(n, bool), meta_loc)
+    else:
+        view.absent('relations', 'No source grasp/contact relation')
+    return view.finish()
+
+
+MEMBER = re.compile(r'^([^/]+)/(\d{8}_\d{6})/(meta\.yml|(\d{12})/(color|labels|aligned_depth_to_color)_(\d+)\.(jpg|npz|png))$')
+
+
+def convert_archive(archive, args, out, calib, calib_path, models, assets, asset_archive):
+    selected_views = set(args.views.split(',')) if args.views else None
+    done = 0
+    current, meta, meta_loc, rows = None, None, None, {}
+    with tempfile.TemporaryDirectory(prefix='dex-source-') as wd:
+        work = Path(wd)
+        def flush():
+            nonlocal done
+            if not rows:
+                return
+            seq = archive.name.split('.tar')[0] + '_' + current
+            reports = {}
+            with sequence_dir(out, seq) as target:
+                for serial, frames in sorted(rows.items()):
+                    reports[serial] = convert_view(target / serial, args, seq, serial, frames, meta,
+                                                   meta_loc, calib, calib_path, models, assets, asset_archive)
+            build_record(out, seq, reports)
+            for staged in list(work.iterdir()):
+                if staged.is_dir():
+                    shutil.rmtree(staged)
+                else:
+                    staged.unlink()
+            done += 1
+            print(f'Converted {seq}', flush=True)
+        with tarfile.open(archive, 'r|gz') as tf:
+            for member in tf:
+                match = MEMBER.fullmatch(member.name)
+                if not member.isfile() or not match:
+                    continue
+                seq = match[2]
+                if seq != current:
+                    flush()
+                    if args.limit and done >= args.limit:
+                        break
+                    current, meta, meta_loc, rows = seq, None, None, {}
+                data_kind = match[3]
+                if data_kind == 'meta.yml':
+                    data = tf.extractfile(member).read()
+                    meta = yaml_load_bytes(data)
+                    meta_loc = locator(archive, member.name, data)
+                    continue
+                serial, kind, f = match[4], match[5], int(match[6])
+                if selected_views and serial not in selected_views or args.max_frames and f >= args.max_frames:
+                    continue
+                data = tf.extractfile(member).read()
+                dest = safe_member(work, member.name)
+                dest.parent.mkdir(parents=True, exist_ok=True)
+                dest.write_bytes(data)
+                row = rows.setdefault(serial, {}).setdefault(f, {})
+                if kind in row:
+                    raise ValueError('duplicate archive frame')
+                row[kind] = (dest, locator(archive, member.name, data))
+            else:
+                flush()
+    return done
+
+
+def main(argv=None):
+    p = parser('dex-ycb', __doc__)
+    p.add_argument('--subjects')
+    args = p.parse_args(argv)
+    archives = sorted({a for r in roots(args.root) for a in r.glob('*-subject-*.tar.gz')})
+    if args.subjects:
+        subjects = {s.strip().zfill(2) for s in args.subjects.split(',')}
+        archives = [a for a in archives if re.search(r'subject-(\d+)', a.name)[1] in subjects]
+    if not archives:
+        p.error('no Dex-YCB subject archives found')
+    cp = find_file(args.root, 'calibration.tar.gz') or find_file(args.root, 'calibration')
+    if cp is None:
+        p.error('calibration source missing')
+    calib = Calibration(cp.parent)
+    models = mano_models(args.mano_dir)
+    out = release(args, 'dex-ycb')
+    with tempfile.TemporaryDirectory(prefix='dex-models-') as wd:
+        assets, asset_archive = prepare_models(args.root, Path(wd))
+        for archive in archives:
+            convert_archive(archive, args, out, calib, cp, models, assets, asset_archive)
+    return 0
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())

H2O/shared-v1.0/script/convert_h2o_v1.0.py

@@ -0,0 +1,678 @@
+#!/usr/bin/env python3
+"""H2O native RGB-D, MANO, joints and objects → CONTRACT.md v1.0 (no rendered masks)."""
+from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
+import io
+import json
+import multiprocessing
+import os
+import re
+import shutil
+import tarfile
+import tempfile
+import threading
+from pathlib import Path
+import zipfile
+
+# Bound native libraries before loading NumPy/Torch; each worker still uses independent FFmpeg processes.
+for key in ('OMP_NUM_THREADS', 'OPENBLAS_NUM_THREADS', 'MKL_NUM_THREADS', 'NUMEXPR_NUM_THREADS'):
+    os.environ[key] = '1'
+
+import numpy as np
+from PIL import Image
+import torch
+
+from warehouse_v1_0 import (View, build_record, check_mesh, find_file, locator, mano_models, mesh_asset,
+    model_profile, parser, release, roots, safe_member, sequence_dir, sha256)
+
+OBJECTS = {1: 'book', 2: 'espresso', 3: 'lotion', 4: 'spray',
+           5: 'milk', 6: 'cocoa', 7: 'chips', 8: 'cappuccino'}
+MEMBER = re.compile(r'^(?:h2o_CASA/)?(subject\d+)/([^/]+)/(\d+)/(cam\d)/(.+)$')
+FRAME = re.compile(r'(rgb|depth|cam_pose|hand_pose|hand_pose_mano|obj_pose|obj_pose_rt)/(\d+)\.(png|jpg|txt)$')
+
+
+def load_splits(root):
+    archive = find_file(root, 'label_split.zip')
+    result = {}
+    if archive is not None:
+        with zipfile.ZipFile(archive) as z:
+            for split in ('train', 'val', 'test'):
+                member = f'label_split/action_{split}.txt'
+                for line_no, line in enumerate(z.read(member).decode().splitlines(), 1):
+                    fields = line.split()
+                    if len(fields) < 2 or fields[1] == 'path':
+                        continue
+                    key = fields[1]
+                    if key in result and result[key][0] != split:
+                        raise ValueError('conflicting official splits')
+                    result[key] = (split, locator(archive, member, record=line_no))
+    return result
+
+
+def load_assets(root, work):
+    archive = find_file(root, 'object.zip')
+    if archive is None:
+        return {}, None
+    with zipfile.ZipFile(archive) as z:
+        for member in z.infolist():
+            if member.is_dir():
+                continue
+            path = safe_member(work, member.filename)
+            path.parent.mkdir(parents=True, exist_ok=True)
+            path.write_bytes(z.read(member))
+    result = {}
+    for oid, name in OBJECTS.items():
+        directory = work / 'object' / name
+        aliases = []
+        # Source cocoa.obj -> ovomaltine.mtl -> cocoa.png, but the ZIP names the
+        # supplied texture ovomaltine.png. Keep all originals and add a byte-identical alias.
+        if name == 'cocoa' and not (directory / 'cocoa.png').exists():
+            mtl = directory / 'ovomaltine.mtl'
+            if mtl.is_file() and 'map_Kd cocoa.png' in mtl.read_text().splitlines():
+                texture = directory / 'ovomaltine.png'
+                shutil.copyfile(texture, directory / 'cocoa.png')
+                aliases.append({'file': 'cocoa.png', 'sha256': sha256(texture),
+                                'source': locator(archive, 'object/cocoa/ovomaltine.png')})
+        candidates = list(directory.glob('*.obj'))
+        if len(candidates) != 1:
+            raise ValueError(f'expected one source OBJ for {name}')
+        check_mesh(candidates[0], directory)  # Fail before decoding any large subject archive.
+        result[oid] = (directory, candidates[0], aliases)
+    return result, archive
+
+
+def convert_view(path, args, seq, cam, rows, intr, splits, models, assets, asset_archive):
+    codec_threads = getattr(args, 'codec_threads', 1)
+    ids = list(range(min(rows), max(rows) + 1))
+    for f in ids:
+        rows.setdefault(f, {})
+    n = len(ids)
+    if not n:
+        raise ValueError('no RGB frames')
+    split, split_loc = splits.get(seq, ('unknown', None))
+    frame_rows = [dict(source_frame_id=f, sync_id=f'{seq}:{f}', source_locator=rows[f]['rgb'][1] if 'rgb' in rows[f] else {'sequence': seq, 'camera': cam, 'source_frame_id': f, 'missing_rgb': True},
+                       component_sources={k: item[1] for k, item in rows[f].items()},
+                       annotation_matches={k: dict(source_frame_id=f, source_timestamp_ns=None,
+                           clock_domain=None, match_method='index_map', delta_ns=None)
+                           for k in rows[f] if k != 'rgb'}) for f in ids]
+    view = View(path, 'H2O', args.release_id, seq.replace('/', '_'), cam, frame_rows, split=split)
+    view.valid_rules = {
+        'mano': 'Per-hand hand_pose_mano flag equals 1; corresponding side model available',
+        'keypoints': "Per-hand hand_pose flag equals 1, broadcast to that hand's 21 source points independently of MANO",
+        'pose': 'Source obj_pose_rt class equals this nonzero object class',
+        'points': 'Source obj_pose class equals this nonzero object class, broadcast to its 21 points',
+    }
+    view.full_jpeg = getattr(args, 'jpeg_full_check', False)
+    view.meta['view_type'] = 'ego' if cam == 'cam4' else 'exo' if cam in ('cam0', 'cam1', 'cam2', 'cam3') else 'unknown'
+    view.meta['sampling']['selection'] = {'max_frames': args.max_frames, 'views': args.views}
+    if split_loc:
+        view.meta.update(split_scheme='official_action_split', split_source_locator=split_loc)
+    # In the measured 275 H2O views, exact FFV1 was always smaller than exact H.264.
+    view.rgb([rows[f]['rgb'][0] if 'rgb' in rows[f] else None for f in ids],
+             {'frames_ref': 'camera/frames.jsonl'}, ffv1_first=True, codec_threads=codec_threads,
+             stream_source=getattr(args, 'stream_media', False))
+    loc = {'frames_ref': 'camera/frames.jsonl', 'source_field': 'component_sources'}
+    if intr is not None:
+        fx, fy, cx, cy, width, height = np.loadtxt(io.BytesIO(intr[0])).ravel()
+        if [int(width), int(height)] != view.meta['camera']['image_size']:
+            raise ValueError('source intrinsics resolution mismatch')
+        K = np.array([[fx, 0, cx], [0, fy, cy], [0, 0, 1]])
+        view.camera(K, intr[1])
+    camps = np.full((n, 4, 4), np.nan, np.float32)
+    cv = np.zeros(n, bool)
+    data = []
+    for i, f in enumerate(ids):
+        lab = {k: np.loadtxt(io.BytesIO(item[0])).ravel() for k, item in rows[f].items() if k not in ('rgb', 'depth')}
+        data.append(lab)
+        if 'cam_pose' in lab:
+            camps[i] = np.linalg.inv(lab['cam_pose'].reshape(4, 4))
+            cv[i] = True
+    if cv.any():
+        view.tensor('extrinsic', 'transform/se3/v1', 'camera/extrinsic.pt', camps, cv,
+                    ['T', 'row', 'column'], loc, convention='T_camera_from_world',
+                    source_convention='T_world_from_camera')
+    motion, mv = np.full((n, 2, 61), np.nan, np.float32), np.zeros((n, 2), bool)
+    points, pv = np.full((n, 2, 21, 3), np.nan, np.float32), np.zeros((n, 2, 21), bool)
+    object_ids = set()
+    for i, lab in enumerate(data):
+        if 'hand_pose_mano' in lab:
+            hands = lab['hand_pose_mano'].reshape(2, 62)
+            for side, row in enumerate(hands):
+                if row[0] not in (0, 1):
+                    raise ValueError('unexpected H2O MANO presence flag')
+                if row[0] == 1:
+                    motion[i, side] = np.r_[row[4:52], row[52:62], row[1:4]]
+                    mv[i, side] = True
+        if 'hand_pose' in lab:
+            hands = lab['hand_pose'].reshape(2, 64)
+            for side, row in enumerate(hands):
+                if row[0] not in (0, 1):
+                    raise ValueError('unexpected H2O keypoint presence flag')
+                if row[0] == 1:
+                    points[i, side] = row[1:].reshape(21, 3)
+                    pv[i, side] = True
+        for key in ('obj_pose', 'obj_pose_rt'):
+            if key in lab and lab[key][0] != 0:
+                object_ids.add(int(lab[key][0]))
+    if any('hand_pose_mano' in lab for lab in data):
+        if len(models) == 2:
+            view.tensor('mano', 'mano/axis_angle61/v1', 'annotation/mano/motion.pt', motion, mv,
+                        ['T', 'hand', 'parameter'], loc, model_profile=model_profile(models))
+        else:
+            view.absent('mano', 'Source parameters exist; --mano-dir must identify both model profiles', 'dependency_missing')
+    if any('hand_pose' in lab for lab in data):
+        view.tensor('keypoints_xyz', 'keypoints/xyz/v1', 'annotation/keypoints/xyz.pt', points, pv,
+                    ['T', 'hand', 'keypoint', 'coordinate'], loc, keypoint_schema='mano21/v1', unit='m')
+    for oid in sorted(object_ids):
+        if oid not in OBJECTS:
+            raise ValueError(f'unknown H2O object class {oid}')
+        entity = OBJECTS[oid]
+        view.object(entity, loc, category_id=oid, source_object_id=oid)
+        poses, ov = np.full((n, 4, 4), np.nan, np.float32), np.zeros(n, bool)
+        pts, kv = np.full((n, 21, 3), np.nan, np.float32), np.zeros((n, 21), bool)
+        for i, lab in enumerate(data):
+            if 'obj_pose_rt' in lab and lab['obj_pose_rt'][0] == oid:
+                poses[i] = lab['obj_pose_rt'][1:].reshape(4, 4)
+                ov[i] = True
+            if 'obj_pose' in lab and lab['obj_pose'][0] == oid:
+                pts[i] = lab['obj_pose'][1:].reshape(21, 3)
+                kv[i] = True
+        if ov.any():
+            view.tensor('pose_' + entity, 'object/rigid/v1', f'annotation/object/{entity}/pose.pt',
+                        poses, ov, ['T', 'row', 'column'], loc, convention='T_camera_from_object', unit='m')
+        if kv.any():
+            view.tensor('points_' + entity, 'object_keypoints/camera/v1',
+                        f'annotation/object/{entity}/keypoints/xyz.pt', pts, kv,
+                        ['T', 'keypoint', 'coordinate'], loc, keypoint_schema='H2O_official_object21',
+                        keypoint_names=['center'] + [f'corner_{i}' for i in range(8)] + [f'edge_midpoint_{i}' for i in range(12)],
+                        source_order='official obj_point_order.jpg: center, 8 corners, 12 edge midpoints', unit='m')
+        if oid in assets:
+            directory, obj, aliases = assets[oid]
+            asset_loc = locator(asset_archive, f'object/{entity}/{obj.name}')
+            if aliases:
+                asset_loc['texture_aliases'] = aliases
+            mesh_asset(view, entity, directory, obj, asset_loc)
+        else:
+            view.absent('mesh_' + entity, 'Source object.zip unavailable', 'not_downloaded')
+    if any('depth' in rows[f] for f in ids):
+        width, height = view.meta['camera']['image_size']
+        depth_threads = getattr(args, 'depth_threads', 1)
+        def read_depth(f):
+            item = rows[f].get('depth')
+            if item:
+                with Image.open(item[0]) as im:
+                    return np.asarray(im).astype(np.float32) / 1000.0
+            return np.full((height, width), np.nan, np.float32)
+        def depths():
+            if depth_threads == 1:
+                yield from map(read_depth, ids)
+            else:
+                with ThreadPoolExecutor(max_workers=depth_threads) as pool:
+                    # Ordered batches bound decoded arrays to depth_threads frames per view.
+                    for start in range(0, len(ids), depth_threads):
+                        yield from pool.map(read_depth, ids[start:start + depth_threads])
+        view.raster('depth', depths(), None, loc | {'source_field': 'component_sources.depth',
+            'definition': 'https://github.com/taeinkwon/h2oplayer/blob/main/utils.py',
+            'conversion': 'registered millimetre camera-Z to metres; no viewer range clipping'},
+            codec_threads=codec_threads, stream=getattr(args, 'stream_media', False))
+    else:
+        view.absent('depth', 'No source depth frames')
+    view.absent('segment', 'H2O provides no native segmentation; rendering deferred to v2.0')
+    view.absent('relations', 'No source contact/grasp annotation')
+    return view.finish()
+
+
+def convert_archive(archive, args, out, splits, models, assets, asset_archive, view_pool=None, ram_budget=None):
+    selected_views = set(args.views.split(',')) if args.views else None
+    subjects = {'subject' + s.strip() for s in args.subjects.split(',')} if args.subjects else None
+    current, rows, intrinsics, done = None, {}, {}, 0
+    published_views = set()
+    camera_jobs, closed_cameras = {}, set()
+    last_camera = None
+    pipeline_active = False
+    ram_reserved = 0
+    try:
+        with tempfile.TemporaryDirectory(prefix='h2o-source-') as wd, \
+                tempfile.TemporaryDirectory(prefix='h2o-ram-', dir=args.ram_stage_dir) as rd:
+            work, ram_work = Path(wd), Path(rd)
+            def queue_camera(cam):
+                if cam in closed_cameras:
+                    return
+                closed_cameras.add(cam)
+                if cam not in published_views and rows.get(cam):
+                    camera_jobs[cam] = view_pool.submit(convert_view, work / '.converted' / cam,
+                        args, '/'.join(current), cam, rows[cam], intrinsics.get(cam),
+                        splits, models, assets, asset_archive)
+            def flush():
+                nonlocal done, ram_reserved
+                if not rows:
+                    return
+                seq = '/'.join(current)
+                reports, pending = {}, {}
+                for cam, frames in sorted(rows.items()):
+                    existing = out / 'data' / seq.replace('/', '_') / cam
+                    if (existing / 'metainfo.json').is_file():
+                        old_frames = [json.loads(line) for line in (existing / 'camera/frames.jsonl').read_text().splitlines()]
+                        if [f['source_frame_id'] for f in old_frames] != list(range(min(frames), max(frames) + 1)):
+                            raise ValueError('existing view has different sampling; choose a new release-id')
+                        for frame in old_frames:
+                            incoming = frames.get(frame['source_frame_id'], {})
+                            previous = frame.get('component_sources', {})
+                            if set(previous) != set(incoming) or any(previous[k].get('sha256') != v[1].get('sha256') for k, v in incoming.items()):
+                                raise ValueError('duplicate H2O view differs from the published source; choose a new release-id')
+                    else:
+                        pending[cam] = frames
+                if pending:
+                    if pipeline_active:
+                        for cam in pending:
+                            queue_camera(cam)
+                        failure = None
+                        for cam, job in camera_jobs.items():
+                            try:
+                                reports[cam] = job.result()
+                            except Exception as exc:
+                                failure = failure or exc
+                        if failure is not None:
+                            raise failure
+                    with sequence_dir(out, seq.replace('/', '_')) as target:
+                        if pipeline_active:
+                            for cam in pending:
+                                shutil.move(work / '.converted' / cam, target / cam)
+                        elif view_pool is None:
+                            for cam, frames in pending.items():
+                                reports[cam] = convert_view(target / cam, args, seq, cam, frames,
+                                    intrinsics.get(cam), splits, models, assets, asset_archive)
+                        else:
+                            jobs = {cam: view_pool.submit(convert_view, target / cam, args, seq, cam,
+                                frames, intrinsics.get(cam), splits, models, assets, asset_archive)
+                                for cam, frames in pending.items()}
+                            failure = None
+                            # Every sibling must finish before sequence_dir can remove its inputs.
+                            for cam, job in jobs.items():
+                                try:
+                                    reports[cam] = job.result()
+                                except Exception as exc:
+                                    failure = failure or exc
+                            if failure is not None:
+                                raise failure
+                    build_record(out, seq.replace('/', '_'), reports)
+                for directory in (work, ram_work):
+                    for staged in list(directory.iterdir()):
+                        if staged.is_dir():
+                            shutil.rmtree(staged)
+                        else:
+                            staged.unlink()
+                if ram_budget is not None:
+                    lock, used, limit = ram_budget
+                    with lock:
+                        used[0] -= ram_reserved
+                        ram_reserved = 0
+                done += 1
+                print(f'{"Converted" if pending else "Verified existing"} {seq}', flush=True)
+            # Uncompressed CASA supports seeking over large frames; gzip is single-pass.
+            mode = 'r:' if archive.suffix == '.tar' else 'r|gz'
+            try:
+                with tarfile.open(archive, mode) as tf:
+                    for member in tf:
+                        match = MEMBER.fullmatch(member.name)
+                        if not member.isfile() or not match:
+                            continue
+                        subj, scene, seq, cam, tail = match.groups()
+                        if subjects and subj not in subjects:
+                            continue
+                        key = (subj, scene, seq)
+                        if key != current:
+                            flush()
+                            if args.limit and done >= args.limit:
+                                break
+                            current, rows, intrinsics = key, {}, {}
+                            camera_jobs.clear()
+                            closed_cameras.clear()
+                            last_camera = None
+                            pipeline_active = args.camera_pipeline
+                            published_views = {view.name for view in
+                                (out / 'data' / '_'.join(key)).glob('cam*')
+                                if (view / 'metainfo.json').is_file()}
+                        if selected_views and cam not in selected_views:
+                            continue
+                        if pipeline_active and cam != last_camera:
+                            if last_camera is not None:
+                                queue_camera(last_camera)
+                            if cam in closed_cameras:
+                                # Complete queued readers before accepting late fields for those cameras.
+                                for job in camera_jobs.values():
+                                    try:
+                                        job.result()
+                                    except Exception:
+                                        pass
+                                shutil.rmtree(work / '.converted', ignore_errors=True)
+                                camera_jobs.clear()
+                                closed_cameras.clear()
+                                pipeline_active = False
+                                print(f'Camera-order fallback: {"/".join(current)}', flush=True)
+                            last_camera = cam
+                        frame = FRAME.fullmatch(tail)
+                        if tail != 'cam_intrinsics.txt' and not frame:
+                            continue
+                        if frame and args.max_frames and int(frame[2]) >= args.max_frames:
+                            continue
+                        data = tf.extractfile(member).read()
+                        if tail.endswith('.txt'):
+                            payload = data  # Small source annotations stay in RAM, not thousands of HDD files.
+                        else:
+                            staging = work
+                            if ram_budget is not None and cam not in published_views:
+                                lock, used, limit = ram_budget
+                                needed = (len(data) + 4095) // 4096 * 4096
+                                with lock:
+                                    if used[0] + needed <= limit:
+                                        used[0] += needed
+                                        ram_reserved += needed
+                                        staging = ram_work
+                            payload = safe_member(staging, member.name)
+                            if cam not in published_views:
+                                payload.parent.mkdir(parents=True, exist_ok=True)
+                                payload.write_bytes(data)
+                        # Published views still verify source bytes, but never read their staging paths.
+                        item = (payload, locator(archive, member.name, data))
+                        if tail == 'cam_intrinsics.txt':
+                            intrinsics[cam] = item
+                        else:
+                            kind, f = frame[1], int(frame[2])
+                            row = rows.setdefault(cam, {}).setdefault(f, {})
+                            if kind in row:
+                                raise ValueError('duplicate H2O frame')
+                            row[kind] = item
+                    else:
+                        flush()
+            finally:
+                # Never delete staging while a closed-camera job can still read from it.
+                for job in camera_jobs.values():
+                    try:
+                        job.result()
+                    except Exception:
+                        pass
+    finally:
+        if ram_budget is not None:
+            lock, used, limit = ram_budget
+            with lock:
+                used[0] -= ram_reserved
+    return done
+
+
+def main(argv=None):
+    p = parser('H2O', __doc__)
+    p.set_defaults(release_id='shared-v1.0')
+    p.add_argument('--workers', type=int, choices=range(1, 5), default=1,
+                   help='Independent subject archives in parallel, 1..4; benchmark before using >2')
+    p.add_argument('--view-backend', choices=('threads', 'processes'), default='threads',
+                   help='Use spawn processes to isolate Python camera processing')
+    p.add_argument('--camera-pipeline', action='store_true',
+                   help='Overlap contiguous-camera extraction and conversion; reject interleaved input')
+    p.add_argument('--ffv1-slices', type=int, choices=(16,),
+                   help='Optional FFV1 v3 slice parallelism; decoded frames must still match exactly')
+    p.add_argument('--view-workers', type=int, choices=range(1, 17), default=1,
+                   help='Shared camera-view conversion pool, independent of archive readers')
+    p.add_argument('--ram-stage-dir', type=Path, help='Optional tmpfs staging; excess PNGs fall back to disk')
+    p.add_argument('--ram-stage-gib', type=int, choices=range(1, 65), default=32,
+                   help='Shared PNG tmpfs quota across all archive readers, not per worker')
+    p.add_argument('--stream-media', action='store_true',
+                   help='Pipe RGB/depth frames to FFmpeg instead of spooling huge RAW files')
+    p.add_argument('--codec-threads', type=int, choices=(1, 2, 4, 8, 10), default=4,
+                   help='FFV1 slice threads per encode; default 4, use 10 for the current full run')
+    p.add_argument('--decode-threads', type=int, choices=(1, 2, 4, 8, 10), default=1,
+                   help='FFmpeg threads for source RGB decoding and decoded-media checks')
+    p.add_argument('--depth-threads', type=int, choices=(1, 2, 4), default=1,
+                   help='Bounded ordered source depth PNG loading per worker')
+    p.add_argument('--subjects')
+    p.add_argument('--include-casa', action='store_true')
+    p.add_argument('--archive', type=Path, help='Convert one explicit source archive')
+    args = p.parse_args(argv)
+    if args.camera_pipeline and args.view_workers < 2:
+        p.error('--camera-pipeline requires at least two --view-workers')
+    os.environ['HAND_WAREHOUSE_FFMPEG_THREADS'] = str(args.decode_threads)
+    if args.ffv1_slices is None:
+        os.environ.pop('HAND_WAREHOUSE_FFV1_SLICES', None)
+    else:
+        os.environ['HAND_WAREHOUSE_FFV1_SLICES'] = str(args.ffv1_slices)
+    if args.include_casa and args.workers != 1:
+        p.error('CASA can share sequence IDs with subject archives; convert it sequentially in a separate release')
+    archives = [args.archive.expanduser()] if args.archive else sorted(
+        {a for root in roots(args.root) for a in root.glob('subject*_v1_1.tar.gz')})
+    if args.subjects and not args.archive:
+        requested = {'subject' + s.strip() for s in args.subjects.split(',')}
+        archives = [arc for arc in archives if arc.name.split('_', 1)[0] in requested]
+    if args.include_casa:
+        casa = find_file(args.root, 'h2o_CASA.tar')
+        if casa:
+            archives.append(casa)
+    if not archives:
+        p.error('no H2O archives found')
+    out = release(args, 'H2O')
+    source_mapping = next(p for p in (Path(__file__).resolve().parents[1] / 'docs/h2o-RAW_TO_SHARED.md',
+                                     Path(__file__).resolve().parents[1] / 'doc/RAW_TO_SHARED.md',
+                                     out / 'doc/RAW_TO_SHARED.md') if p.is_file())
+    mapping = out / 'doc/RAW_TO_SHARED.md'
+    if mapping.exists() and mapping.read_bytes() != source_mapping.read_bytes():
+        raise ValueError('RAW_TO_SHARED.md changed; choose a new release-id')
+    if not mapping.exists():
+        shutil.copyfile(source_mapping, mapping)
+    os.environ.pop('HAND_WAREHOUSE_MEDIA_TMPDIR', None)
+    ram_budget = None
+    if args.ram_stage_dir is not None:
+        args.ram_stage_dir = args.ram_stage_dir.expanduser().resolve()
+        if not args.ram_stage_dir.is_relative_to(Path('/dev/shm').resolve()):
+            p.error('--ram-stage-dir must be on /dev/shm; it is temporary input, not an output dataset')
+        args.ram_stage_dir.mkdir(parents=True, exist_ok=True)
+        limit = args.ram_stage_gib * 1024**3
+        if shutil.disk_usage(args.ram_stage_dir).free < limit + 8 * 1024**3:
+            p.error('not enough tmpfs headroom for the requested shared staging quota')
+        ram_budget = (threading.Lock(), [0], limit)
+        if args.stream_media:
+            os.environ['HAND_WAREHOUSE_MEDIA_TMPDIR'] = str(args.ram_stage_dir)
+    torch.set_num_threads(1)
+    models, splits = mano_models(args.mano_dir), load_splits(args.root)
+    executor = ProcessPoolExecutor if args.view_backend == 'processes' else ThreadPoolExecutor
+    options = {'mp_context': multiprocessing.get_context('spawn')} if args.view_backend == 'processes' else {'thread_name_prefix': 'h2o-view'}
+    with tempfile.TemporaryDirectory(prefix='h2o-assets-') as wd, \
+            executor(max_workers=args.view_workers, **options) as views:
+        assets, asset_archive = load_assets(args.root, Path(wd))
+        def convert_one(archive):
+            return convert_archive(archive, args, out, splits, models, assets, asset_archive,
+                                   views if args.view_workers > 1 else None, ram_budget)
+        if args.workers == 1:
+            for archive in archives:
+                convert_one(archive)
+        else:
+            with ThreadPoolExecutor(max_workers=args.workers) as pool:
+                for archive, count in zip(archives, pool.map(convert_one, archives), strict=True):
+                    print(f'Archive {archive.name}: {count} sequences', flush=True)
+    return 0
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())
+
+# TODO(v2.0): 以下旧 MANO 网格和渲染分割代码仅保留备查；v1.0 不生成标注。
+# def load_mano_model(path: Path):
+#     """MPI MANO pkl → numpy 字典。shapedirs 在 pkl 内为 chumpy Select（对 (778,3,20) 基
+#     底 flat 选取前 10 个 shape 分量；idxs 已核实等价于 base[:,:,:10]）。"""
+#     with open(path, "rb") as f:
+#         d = _ManoUnpickler(f, encoding="latin1").load()
+#     sd = d["shapedirs"]
+#     base = np.asarray(sd.a.x, np.float64)                       # (778, 3, 20)
+#     shapedirs = base.reshape(-1)[np.asarray(sd.idxs)].reshape(base.shape[0], 3, -1)
+#     jreg = d["J_regressor"]
+#     jreg = np.asarray(jreg.todense()) if hasattr(jreg, "todense") else np.asarray(jreg)
+#     return {"v_template": np.asarray(d["v_template"], np.float64),
+#             "shapedirs": shapedirs,
+#             "posedirs": np.asarray(d["posedirs"], np.float64),
+#             "weights": np.asarray(d["weights"], np.float64),
+#             "J_regressor": np.asarray(jreg, np.float64),
+#             "kintree": np.asarray(d["kintree_table"], np.int64),
+#             "faces": np.asarray(d["f"], np.int64)}
+#
+#
+# def rodrigues(r):
+#     """axis-angle (N,3) → 旋转矩阵 (N,3,3)。公式 I + sinθ·K + (1−cosθ)·K²（已对 scipy 核验）。"""
+#     r = np.asarray(r, np.float64).reshape(-1, 3)
+#     th = np.linalg.norm(r, axis=1)
+#     R = np.zeros((len(r), 3, 3))
+#     nz = th > 1e-12
+#     R[~nz] = np.eye(3)
+#     k = r[nz] / th[nz, None]
+#     K = np.zeros((len(k), 3, 3))
+#     K[:, 0, 1], K[:, 0, 2], K[:, 1, 2] = -k[:, 2], k[:, 1], -k[:, 0]
+#     K[:, 1, 0], K[:, 2, 0], K[:, 2, 1] = k[:, 2], -k[:, 1], k[:, 0]
+#     s, c = np.sin(th[nz])[:, None, None], np.cos(th[nz])[:, None, None]
+#     R[nz] = np.eye(3) + s * K + (1 - c) * (K @ K)
+#     return R
+#
+#
+# def _mano_chain(model, betas, pose):
+#     """返回 (v_shaped, J, Rs, G)：G 为 16 个关节的全局变换（平移列 = 关节位置）。"""
+#     v = model["v_template"] + np.einsum("nkc,c->nk", model["shapedirs"], betas)
+#     J = model["J_regressor"] @ v
+#     Rs = rodrigues(np.asarray(pose, np.float64).reshape(16, 3))
+#     kt = model["kintree"]
+#     G = np.zeros((16, 4, 4))
+#     for j in range(16):
+#         p = kt[0, j] if j else -1
+#         loc = np.eye(4)
+#         loc[:3, :3] = Rs[j]
+#         loc[:3, 3] = J[j] - (J[p] if j else 0.0)
+#         G[j] = loc if not j else G[p] @ loc
+#     return v, J, Rs, G
+#
+#
+# def mano_forward(model, betas, pose, transl):
+#     """MANO LBS 前向（经典蒙皮 M_j = G_j·G0_j⁻¹；对 hand_pose 21 关节实证 0.0 mm）。
+#     pose: 48 axis-angle（[0:3] 全局）；返回相机系米顶点 (778,3)。"""
+#     v, J, Rs, G = _mano_chain(model, betas, pose)
+#     lrot = (Rs[1:] - np.eye(3)).ravel()
+#     v_posed = v + np.einsum("nkc,c->nk", model["posedirs"], lrot)
+#     kt = model["kintree"]
+#     G0 = np.zeros((16, 4, 4))
+#     for j in range(16):
+#         p = kt[0, j] if j else -1
+#         loc = np.eye(4)
+#         loc[:3, 3] = J[j] - (J[p] if j else 0.0)
+#         G0[j] = loc if not j else G0[p] @ loc
+#     M = np.einsum("nj,jab->nab", model["weights"], G @ np.linalg.inv(G0))
+#     out = np.einsum("nab,nb->na", M[:, :3, :3], v_posed) + M[:, :3, 3]
+#     return out + np.asarray(transl, np.float64)
+#
+#
+# def mano_chain_joints(model, betas, pose, transl):
+#     """16 个关节位置（相机系米）= 链平移列；自检用。"""
+#     _, _, _, G = _mano_chain(model, betas, pose)
+#     return G[:, :3, 3] + np.asarray(transl, np.float64)
+#
+#
+# def rasterize(verts_cam, faces, labels, W, H, fx, fy, cx, cy):
+#     z = verts_cam[:, 2]
+#     P = faces
+#     ok = (z[P] > 0.02).all(1)                 # 任一顶点在相机后 → 丢弃该三角形
+#     P, labels = P[ok], labels[ok]
+#     if len(P) == 0:
+#         return np.zeros((H, W), np.uint16), np.zeros(0, np.int64)
+#     x = verts_cam[:, 0] / z * fx + cx
+#     y = verts_cam[:, 1] / z * fy + cy
+#     xa = np.maximum(0, np.minimum(np.floor(x[P[:, 0]]), np.minimum(
+#         np.floor(x[P[:, 1]]), np.floor(x[P[:, 2]]))).astype(np.int64))
+#     xb = np.minimum(W - 1, np.maximum(np.ceil(x[P[:, 0]]), np.maximum(
+#         np.ceil(x[P[:, 1]]), np.ceil(x[P[:, 2]]))).astype(np.int64))
+#     ya = np.maximum(0, np.minimum(np.floor(y[P[:, 0]]), np.minimum(
+#         np.floor(y[P[:, 1]]), np.floor(y[P[:, 2]]))).astype(np.int64))
+#     yb = np.minimum(H - 1, np.maximum(np.ceil(y[P[:, 0]]), np.maximum(
+#         np.ceil(y[P[:, 1]]), np.ceil(y[P[:, 2]]))).astype(np.int64))
+#     xa, ya = np.minimum(xa, xb), np.minimum(ya, yb)   # 完全出画的三角形 → 空 bbox
+#     wdt, hgt = xb - xa + 1, yb - ya + 1
+#     cnt = wdt * hgt
+#     tot = int(cnt.sum())
+#     tri = np.repeat(np.arange(len(P)), cnt)
+#     k = np.arange(tot) - (np.cumsum(cnt) - cnt)[tri]
+#     px = xa[tri] + k % wdt[tri] + 0.5
+#     py = ya[tri] + k // wdt[tri] + 0.5
+#     ax, ay = x[P[tri, 0]], y[P[tri, 0]]
+#     bx, by = x[P[tri, 1]], y[P[tri, 1]]
+#     cx_, cy_ = x[P[tri, 2]], y[P[tri, 2]]
+#     area = ((x[P[:, 1]] - x[P[:, 0]]) * (y[P[:, 2]] - y[P[:, 0]])
+#             - (y[P[:, 1]] - y[P[:, 0]]) * (x[P[:, 2]] - x[P[:, 0]]))
+#     areat = area[tri]
+#     e0 = (bx - ax) * (py - ay) - (by - ay) * (px - ax)
+#     e1 = (cx_ - bx) * (py - by) - (cy_ - by) * (px - bx)
+#     e2 = (ax - cx_) * (py - cy_) - (ay - cy_) * (px - cx_)
+#     s = np.sign(areat)
+#     inside = (np.abs(areat) > 1e-9) & ((e0 * s >= 0) & (e1 * s >= 0) & (e2 * s >= 0))
+#     px, py, tri = px[inside], py[inside], tri[inside]
+#     at = areat[inside]
+#     w0, w1, w2 = e1[inside] / at, e2[inside] / at, e0[inside] / at
+#     zz = w0 * z[P[tri, 0]] + w1 * z[P[tri, 1]] + w2 * z[P[tri, 2]]
+#     pix = py.astype(np.int64) * W + px.astype(np.int64)
+#     order = np.lexsort((zz, pix))             # 同像素按深度升序 → 首条即最近
+#     pix, tri = pix[order], tri[order]
+#     first = np.unique(pix, return_index=True)[1]
+#     mask = np.zeros(H * W, np.uint16)
+#     win = tri[first]
+#     mask[pix[first]] = labels[win] + 1        # 临时标签 +1（0=背景）
+#     return mask.reshape(H, W), np.bincount(labels[win] + 1)
+#
+#
+# # ============================================================================
+# # ffmpeg 写出与核验（与 convert_dexycb.py 同构）
+# # ============================================================================
+#             verts, faces_all, labels = [], [], []
+#             for ti, (tid, kind) in enumerate(tracks):
+#                 if kind == "hand_left" and b["mano"][0] > 0.5:
+#                     verts.append(mano_forward(mano_models["left"], b["mano"][52:62],
+#                                               b["mano"][4:52], b["mano"][1:4]))
+#                     faces_all.append(mano_models["left"]["faces"])
+#                 elif kind == "hand_right" and b["mano"][62] > 0.5:
+#                     verts.append(mano_forward(mano_models["right"], b["mano"][114:124],
+#                                               b["mano"][66:114], b["mano"][63:66]))
+#                     faces_all.append(mano_models["right"]["faces"])
+#                 elif isinstance(kind, tuple) and int(b["objp"][0]) == kind[1]:
+#                     M = np.asarray(b["objrt"][1:17], np.float64).reshape(4, 4)
+#                     V0, F0 = meshes[kind[1]]
+#                     verts.append(V0 @ M[:3, :3].T + M[:3, 3])
+#                     faces_all.append(F0)
+#                 else:
+#                     continue
+#                 labels.append(np.full(len(faces_all[-1]), ti))
+#             if verts:
+#                 V = np.concatenate(verts)
+#                 offs = np.cumsum([0] + [len(v) for v in verts[:-1]])
+#                 F = np.concatenate([fa + o for fa, o in zip(faces_all, offs)])
+#                 mask, counts = rasterize(V, F, np.concatenate(labels),
+#                                          IMG_W, IMG_H, fx, fy, cx, cy)
+#             else:
+#                 mask, counts = np.zeros((IMG_H, IMG_W), np.uint16), np.zeros(0, np.int64)
+#             lut = np.zeros(n_tracks + 1, np.uint16)
+#             for ti in range(n_tracks):
+#                 present = ti + 1 < len(counts) and counts[ti + 1] > 0
+#                 if present:
+#                     presence[tracks[ti][0]][f] = True
+#                     if active_id[ti] is None:
+#                         pid = next(i for i in range(1, 65535) if i not in used)
+#                         used.add(pid)
+#                         active_id[ti], seg_start[ti] = pid, f
+#                         frame_maps[ti].append({"id": pid})
+#                     lut[ti + 1] = active_id[ti]
+#                 elif active_id[ti] is not None:
+#                     frame_maps[ti][-1]["frames"] = [seg_start[ti], f - 1]
+#                     segments[ti].append([seg_start[ti], f - 1])
+#                     used.discard(active_id[ti])
+#                     active_id[ti] = None
+#             mw.write(lut[mask])
+
+
+# TODO(v2.0): 旧物体尺寸派生代码；v1.0 仅保留源物体关键点。
+#                 pts = np.asarray(b["objp"][1:64], np.float64).reshape(21, 3)
+#                 half = np.abs((pts[1:9] - pts[0]) @ M[:3, :3]).max(0)
+#                 obj_recs.append({
+#                     "track_id": f"o{classes.index(c)}", "frame": f,
+#                     "mesh_ref": f"h2o/{OBJ_CLASSES[c]}.obj",
+#                     "pose": {"quat": rotmat_to_quat_wxyz(M[:3, :3])[0].tolist(),
+#                              "trans": M[:3, 3].tolist()},
+#                     "bbox_center3d": pts[0].tolist(),
+#                     "bbox_size": (2 * half).tolist()})

H2O/shared-v1.0/script/convert_hoi4d_v1.0.py

@@ -0,0 +1,595 @@
+#!/usr/bin/env python3
+"""HOI4D native observations and annotations → CONTRACT.md v1.0."""
+from collections import Counter
+from concurrent.futures import ThreadPoolExecutor
+from contextlib import ExitStack
+from functools import lru_cache
+import gzip
+import io
+import json
+import os
+from pathlib import Path
+import pickle
+import re
+import shutil
+import shlex
+import posixpath
+import subprocess
+import tempfile
+import tarfile
+import threading
+import zipfile
+
+# Each worker owns one Python thread; prevent native libraries spawning CPU pools.
+for key in ('OMP_NUM_THREADS', 'OPENBLAS_NUM_THREADS', 'MKL_NUM_THREADS', 'NUMEXPR_NUM_THREADS'):
+    os.environ[key] = '1'
+
+import numpy as np
+from PIL import Image
+from scipy.spatial.transform import Rotation
+
+import torch
+
+from warehouse_v1_0 import (View, build_record, check_mesh, decode_args, find_file, json_write, locator,
+    mano_models, model_profile, parser, probe, release, run, safe_member, sequence_dir)
+
+CATEGORIES = {
+    1:  dict(cad="rigid/ToyCar", groups=[("main", [1], "Toy Car", 1)],
+             hands={2: "right"}, objmap={"Toycar": "main"}),
+    2:  dict(cad="rigid/Mug", groups=[("a", [1], "Mug", 2), ("b", [3], "Mug", 2)],
+             hands={2: "right"}, objmap={"Watercup": "a"}),
+    3:  dict(cad="articulated/Laptop", groups=[("main", [1, 3], "Laptop", 3)],
+             hands={2: "right"},
+             objmap={"Laptopdisplay": "main", "Laptopkeyboard": "main"}),
+    4:  dict(cad="articulated/StorageFurniture",
+             groups=[("main", [1, 3, 4], "Storage Furniture", 4),
+                     ("bottle", [5], "Bottle", 5)],
+             hands={2: "right"},
+             objmap={"Lockerbody": "main", "Lockersldingdoor": "main",
+                     "Lockerdrawer": "main", "Safebox": "main"}),
+    5:  dict(cad="rigid/Bottle", groups=[("main", [1], "Bottle", 5),
+                                         ("mug", [3], "Mug", 2)],
+             hands={2: "right"}, objmap={"bottleddrinks": "main"}),
+    6:  dict(cad="articulated/Safe", groups=[("main", [1, 3], "Safe", 6),
+                                              ("bottle", [4], "Bottle", 5)],
+             hands={2: "right"}, objmap={"Safebox": "main", "Safedoor": "main"}),
+    7:  dict(cad="rigid/Bowl", groups=[("main", [1], "Bowl", 7),
+                                       ("ball", [3], "Ball", None)],
+             hands={2: "right"}, objmap={"bowl": "main"}),
+    8:  dict(cad="articulated/Bucket",
+             groups=[("a", [1, 3], "Bucket", 8), ("b", [4, 5], "Bucket", 8)],
+             hands={2: "right", 6: "left"},
+             objmap={"bucket": "a", "Buckethandle": "a"}),
+    9:  dict(cad="articulated/Scissors", groups=[("main", [1, 3], "Scissors", 9)],
+             hands={2: "right"},
+             objmap={"scissors": "main", "scissors2": None}),   # scissors2：场景道具，无掩码
+    11: dict(cad="articulated/Pliers", groups=[("main", [1, 3], "Pliers", 11)],
+             hands={2: "right"}, objmap={"Pliersleft": "main", "Pliersright": "main"}),
+    12: dict(cad="rigid/Kettle", groups=[("main", [1], "Kettle", 12),
+                                         ("mug", [3], "Mug", 2)],
+             hands={2: "right"}, objmap={"kettle": "main"}),
+    13: dict(cad="rigid/Knife", groups=[("main", [1], "Knife", 13),
+                                        ("apple", [3], "Apple", None)],
+             hands={2: "right"}, objmap={"knife": "main"}),
+    14: dict(cad="articulated/TrashCan",
+             groups=[("main", [1, 3], "Trash Can", 14), ("bottle", [4], "Bottle", 5)],
+             hands={2: "right"}, objmap={"Dustbinbase": "main", "Dustbincover": "main"}),
+    17: dict(cad="articulated/Lamp", groups=[("main", [1, 3, 4, 5], "Lamp", 17)],
+             hands={2: "right", 6: "left"},
+             objmap={"Lampbase": "main", "Tablelampfirstjoint": "main",
+                     "Tablelampsecondjoint": "main", "Thirdjointoftablelamp": "main"}),
+    18: dict(cad="articulated/Stapler",
+             groups=[("main", [1, 3, 4], "Stapler", 18), ("paper", [5], "Paper", None)],
+             hands={2: "right", 6: "left"},       # 6=左手：官方 label.csv 缺列，实测修正
+             objmap={"Staplerbase": "main", "Staplercover": "main"}),
+    20: dict(cad="rigid/Chair", groups=[("main", [1], "Chair", 20)],
+             hands={2: "right", 3: "left"}, objmap={"chair": "main"}),
+}
+
+
+# Official HOI4D definitions/motion segmentation/label.csv; no spatial side inference.
+REGION_NAMES = {1: {1: 'Toy Car', 2: 'Right Hand'}, 2: {1: 'Mug', 2: 'Right Hand', 3: 'Mug'}, 3: {1: 'Screen', 2: 'Right Hand', 3: 'Keyboard'}, 4: {1: 'Door', 2: 'Right Hand', 3: 'Body of the Cabinet', 4: 'Drawer', 5: 'Bottle'}, 5: {1: 'Bottle', 2: 'Right Hand', 3: 'Mug'}, 6: {1: 'Body of the Safe', 2: 'Right Hand', 3: 'Door', 4: 'Bottle'}, 7: {1: 'Bowl', 2: 'Right Hand', 3: 'Ball'}, 8: {1: 'Body of the Bucket', 2: 'Right Hand', 3: 'Handle', 4: 'Body of the Bucket', 5: 'Handle', 6: 'Left Hand'}, 9: {1: 'Left Part of the Scissors', 2: 'Right Hand', 3: 'Right Part of the Scissors'}, 11: {1: 'Left Part of the Pliers', 2: 'Right Hand', 3: 'Right Part of the Pliers'}, 12: {1: 'Kettle', 2: 'Right Hand', 3: 'Mug'}, 13: {1: 'Knife', 2: 'Right Hand', 3: 'Apple'}, 14: {1: 'Body of the Trash Can', 2: 'Right Hand', 3: 'Lid', 4: 'Bottle'}, 17: {1: 'Base', 2: 'Right Hand', 3: 'Tube', 4: 'Top Link', 5: 'Bottom Link', 6: 'Left Hand'}, 18: {1: 'Base', 2: 'Right Hand', 3: 'Slot', 4: 'Lid', 5: 'Paper'}, 20: {1: 'Chair', 2: 'Right Hand', 3: 'Left Hand'}}
+
+
+class SplitFiles(io.RawIOBase):
+    """Read concatenated source gzip parts without concatenating 119 GB on disk."""
+    def __init__(self, paths):
+        self.paths = iter(paths)
+        self.current = None
+
+    def readable(self):
+        return True
+
+    def readinto(self, buffer):
+        while True:
+            if self.current is None:
+                path = next(self.paths, None)
+                if path is None:
+                    return 0
+                self.current = path.open('rb')
+            n = self.current.readinto(buffer)
+            if n:
+                return n
+            self.current.close()
+            self.current = None
+
+    def close(self):
+        if self.current is not None:
+            self.current.close()
+        super().close()
+
+
+def extract_depths(root, sequences, work):
+    directory = find_file(root, 'DEPTH')
+    if directory is None:
+        return {}, []
+    parts = sorted(directory.glob('HOI4D_depth_video.tar.gz*'), key=lambda p: int(p.name.split('.gz')[1] or 0))
+    if not parts:
+        return {}, []
+    wanted = {f'HOI4D_depth_video/{seq}/align_depth/depth_video.avi': seq for seq in sequences}
+    found = {}
+    with SplitFiles(parts) as joined, io.BufferedReader(joined) as inp, gzip.GzipFile(fileobj=inp) as gz:
+        with tarfile.open(fileobj=gz, mode='r|') as tf:
+            for member in tf:
+                if member.isfile() and member.name in wanted:
+                    seq = wanted[member.name]
+                    dest = work / (seq.replace('/', '_') + '.avi')
+                    with tf.extractfile(member) as src, dest.open('wb') as dst:
+                        shutil.copyfileobj(src, dst)
+                    found[seq] = (dest, {'files': [str(p.resolve()) for p in parts], 'member': member.name})
+                    if len(found) == len(wanted):
+                        break
+    return found, parts
+
+
+def palette_mask(data):
+    with Image.open(io.BytesIO(data)) as im:
+        arr = np.asarray(im)
+    if arr.ndim == 2:
+        return arr.astype(np.uint16)
+    if arr.shape[-1] != 3:
+        raise ValueError('unsupported HOI4D mask representation')
+    # VOC's 256-color palette interleaves eight label bits into the high RGB bits.
+    # Invert it directly instead of sorting every pixel with np.unique.
+    if np.any(arr & np.array([31, 31, 63], dtype=np.uint8)):
+        raise ValueError('source mask has colors outside official VOC palette')
+    mask = np.zeros(arr.shape[:2], dtype=np.uint16)
+    for bit in range(8):
+        mask |= ((arr[..., bit % 3] >> (7 - bit // 3)) & 1).astype(np.uint16) << bit
+    return mask
+
+
+# Four workers each reuse annotation and hand indices instead of rebuilding the ZIP directory.
+@lru_cache(maxsize=8)
+def index_frames(z):
+    index = {}
+    for name in z.namelist():
+        parent, _, filename = name.rpartition('/')
+        stem, dot, suffix = filename.rpartition('.')
+        if not dot or not stem.isdigit() or suffix not in ('png', 'json', 'pickle'):
+            continue
+        entries = index.setdefault((parent + '/', '.' + suffix), {})
+        f = int(stem)
+        if f in entries:
+            raise ValueError(f'duplicate frame number under {parent}')
+        entries[f] = name
+    return index
+
+
+def frame_members(z, prefix, suffix):
+    return index_frames(z).get((prefix, suffix), {})
+
+
+def convert_sequence(seq, args, target, rgbz, annz, hpz, calz, cadz, paths, models, depth):
+    # ponytail: cache only one sequence's encoded annotation bytes; no decoded masks or CAD cache.
+    @lru_cache(maxsize=None)
+    def read_annotation(archive, name):
+        return archive.read(name)
+
+    category = int(re.search(r'/C(\d+)/', seq)[1])
+    definition = CATEGORIES.get(category)
+    if definition is None:
+        raise ValueError(f'category mapping unavailable: {category}')
+    prefix = f'HOI4D_annotations/{seq}'
+    mask_members = frame_members(annz, prefix + '/2Dseg/mask/', '.png')
+    shifted = frame_members(annz, prefix + '/2Dseg/shift_mask/', '.png')
+    if not mask_members:
+        mask_members = shifted
+    pose_members = frame_members(annz, prefix + '/objpose/', '.json')
+    hand_members = {side: frame_members(hpz, f'Hand_pose/handpose_{side}_hand/{seq}/', '.pickle')
+                    for side in ('left', 'right')} if hpz else {'left': {}, 'right': {}}
+    with tempfile.TemporaryDirectory(prefix='hoi4d-source-') as wd:
+        work = Path(wd)
+        member = f'HOI4D_release/{seq}/align_rgb/image.mp4'
+        video = work / 'rgb.mp4'
+        with rgbz.open(member) as src, video.open('wb') as dst:
+            shutil.copyfileobj(src, dst)
+        info = probe(video)
+        if 'nb_frames' not in info:
+            info['nb_frames'] = json.loads(run(['ffprobe', '-v', 'error', '-count_frames', '-select_streams', 'v:0',
+                '-show_entries', 'stream=nb_read_frames', '-of', 'json', str(video)], stdout=subprocess.PIPE).stdout)['streams'][0]['nb_read_frames']
+        n = int(info['nb_frames'])
+        if args.max_frames:
+            n = min(n, args.max_frames)
+        if shifted and mask_members != shifted and any(f not in mask_members and f in shifted for f in range(n)):
+            raise ValueError('shift_mask has frames absent from mask; correspondence unverified')
+        frames = []
+        for f in range(n):
+            sources = {}
+            if f in mask_members:
+                sources['segment'] = locator(paths['annotations'], mask_members[f], read_annotation(annz, mask_members[f]))
+            if f in pose_members:
+                sources['object'] = locator(paths['annotations'], pose_members[f], read_annotation(annz, pose_members[f]))
+            for side in ('left', 'right'):
+                if f in hand_members[side]:
+                    sources['hand_' + side] = locator(paths['hand'], hand_members[side][f], read_annotation(hpz, hand_members[side][f]))
+            frames.append(dict(source_frame_id=f, source_locator=locator(paths['release'], member, frame=f),
+                component_sources=sources, annotation_matches={k: dict(source_frame_id=f, source_timestamp_ns=None,
+                    clock_domain=None, match_method='index_map', delta_ns=None) for k in sources}))
+        view = View(target, 'HOI4D', args.release_id, seq.replace('/', '_'), 'ego', frames)
+        view.valid_rules = {
+            'mano': 'Matched side pickle contains poseCoeff, beta, trans and corresponding model available; zero angles are not missing',
+            'keypoints': 'Matched side pickle contains kps2D; independently provided source points, no visibility gate',
+            'pose': 'Matched source object record contains rotation and center; record identity retained',
+            'dimensions': 'Matched source object record contains dimensions; size/mesh relation remains source_unspecified',
+            'segment': 'Source mask or shift_mask present with verified palette codes and entity/part mapping',
+        }
+        view.full_jpeg = getattr(args, 'jpeg_full_check', False)
+        view.meta['view_type'] = 'ego'
+        view.meta['sampling']['selection'] = {'max_frames': args.max_frames, 'source_sequence': seq}
+        view.rgb([], locator(paths['release'], member), video=video, count=n)
+        intr_name = f'camera_params/{seq.split("/")[0]}/intrin.npy'
+        if calz and intr_name in calz.namelist():
+            view.camera(np.load(io.BytesIO(calz.read(intr_name))), locator(paths['camera'], intr_name))
+        extr_name = prefix + '/3Dseg/output.log'
+        if extr_name in annz.NameToInfo:
+            lines = [line for line in annz.read(extr_name).decode().splitlines() if line.strip()]
+            if len(lines) % 5:
+                raise ValueError('invalid Open3D camera trajectory')
+            extr, valid = np.full((n, 4, 4), np.nan, np.float32), np.zeros(n, bool)
+            for i in range(0, len(lines), 5):
+                idx = int(lines[i].split()[0])
+                if idx < n:
+                    if valid[idx]:
+                        raise ValueError('duplicate trajectory index')
+                    extr[idx] = np.linalg.inv(np.loadtxt(io.StringIO('\n'.join(lines[i + 1:i + 5]))))
+                    valid[idx] = True
+            view.tensor('extrinsic', 'transform/se3/v1', 'camera/extrinsic.pt', extr, valid,
+                ['T', 'row', 'column'], locator(paths['annotations'], extr_name),
+                convention='T_camera_from_world', source_convention='Open3D LOG camera-to-world; inverted')
+        loc = {'frames_ref': 'camera/frames.jsonl', 'source_field': 'component_sources'}
+        motion, mv = np.full((n, 2, 61), np.nan, np.float32), np.zeros((n, 2), bool)
+        points, pv = np.full((n, 2, 21, 2), np.nan, np.float32), np.zeros((n, 2, 21), bool)
+        for slot, side in enumerate(('left', 'right')):
+            for f, name in hand_members[side].items():
+                if f >= n:
+                    continue
+                data = pickle.loads(read_annotation(hpz, name), encoding='latin1')
+                if all(k in data for k in ('poseCoeff', 'beta', 'trans')):
+                    motion[f, slot] = np.r_[np.asarray(data['poseCoeff']).reshape(48),
+                        np.asarray(data['beta']).reshape(10), np.asarray(data['trans']).reshape(3)]
+                    mv[f, slot] = True  # Zero angles are a valid pose, not a missing sentinel.
+                if 'kps2D' in data:
+                    points[f, slot] = np.asarray(data['kps2D']).reshape(21, 2)
+                    pv[f, slot] = True
+        if mv.any():
+            if all(side in models for slot, side in enumerate(('left', 'right')) if mv[:, slot].any()):
+                view.tensor('mano', 'mano/axis_angle61/v1', 'annotation/mano/motion.pt', motion, mv,
+                    ['T', 'hand', 'parameter'], loc, model_profile=model_profile(models))
+            else:
+                view.absent('mano', 'Source MANO exists; provide --mano-dir to fix model identity', 'dependency_missing')
+        if pv.any():
+            view.tensor('keypoints_uv', 'keypoints/uv/v1', 'annotation/keypoints/uv.pt', points, pv,
+                ['T', 'hand', 'keypoint', 'coordinate'], loc, keypoint_schema='mano21/v1', unit='pixel',
+                source_method='Upstream fitted MANO projection already present in kps2D')
+        objects = {}
+        groups = {key: (vals, name, cls) for key, vals, name, cls in definition['groups']}
+        def ensure_object(key, source_loc, category_id=None, frame=None, source_id=None):
+            if key not in objects:
+                view.object(key, source_loc, scope='frame' if frame is not None else 'sequence',
+                    scope_id=f'{seq}:{frame}' if frame is not None else None,
+                    category_id=category_id, source_object_id=source_id)
+                objects[key] = {}
+        mappings = []
+        mask_valid = [f in mask_members for f in range(n)]
+        if any(mask_valid):
+            width, height = view.meta['camera']['image_size']
+            def masks():
+                for f in range(n):
+                    labels = []
+                    if f in mask_members:
+                        mask = palette_mask(read_annotation(annz, mask_members[f]))
+                        for value in np.unique(mask):
+                            value = int(value)
+                            if value in (0, 255):
+                                if value == 255:
+                                    raise ValueError('HOI4D 255 region semantics unverified')
+                                continue
+                            if REGION_NAMES[category].get(value) in ('Left Hand', 'Right Hand'):
+                                entity, part, cls = 'hand_' + REGION_NAMES[category][value].split()[0].lower(), None, 'hand'
+                            else:
+                                group = next((k for k, (vals, _, _) in groups.items() if value in vals), None)
+                                entity = group or f'unknown_region_{value}'
+                                cls = groups[group][2] if group else None
+                                part = f'region_{value}' if group and len(groups[group][0]) > 1 else None
+                                ensure_object(entity, frames[f]['component_sources']['segment'], cls)
+                                if part:
+                                    objects[entity].setdefault('parts', {})[part] = {'part_id': part, 'parent_id': None,
+                                        'has_model': False, 'mesh_ref': None, 'source_region_code': value, 'name': REGION_NAMES[category].get(value)}
+                            labels.append(dict(mask_value=value, category_id=cls, entity_id=entity,
+                                               part_id=part, source_instance_id=value,
+                                               source_attributes={'region_name': REGION_NAMES[category].get(value)}))
+                    else:
+                        mask = np.zeros((height, width), np.uint16)
+                    mappings.append(dict(frame_index=f, labels=labels))
+                    yield mask
+            view.raster('segment', masks(), mask_valid, loc, mappings)
+            view.meta['components']['segment']['coverage'] = {'kind': 'unknown'}
+            if any(label['part_id'] for row in mappings for label in row['labels']):
+                view.meta['components']['segment']['segmentation_type'] = 'part'
+        else:
+            view.absent('segment', 'No native segmentation for this source sequence')
+        records = {}
+        for f, name in pose_members.items():
+            if f >= n:
+                continue
+            data = json.loads(read_annotation(annz, name))
+            # isEffective is not used to erase existing numerical annotations.
+            source_records = data.get('dataList', data.get('objects'))
+            if source_records is None:
+                raise ValueError('unrecognized source object annotation structure')
+            if category == 2 and source_records and (len(source_records) != 1 or
+                    source_records[0].get('id') != 1 or source_records[0].get('label') != 'Watercup'):
+                raise ValueError('C2 native pose differs from the verified one-object target; cannot map to mask region 1')
+            label_counts = Counter(record['label'] for record in source_records)
+            per_frame = set()
+            for record_index, record in enumerate(source_records):
+                label = record['label']
+                group = definition['objmap'].get(label)
+                # Pinned official ObjTracking reads dataList[0] with segmentation label 1.
+                source_primary = category == 2 and record_index == 0 and record.get('id') == 1 and group == 'a'
+                ambiguous = group is not None and sum(g[2] == groups[group][2] for g in groups.values()) > 1 and not source_primary
+                ambiguous = ambiguous or (label_counts[label] > 1 and not source_primary) or group is None
+                entity = f'pose_frame_{f}_{record_index}' if ambiguous else group
+                ensure_object(entity, locator(paths['annotations'], name, record=record_index),
+                    groups[group][2] if group else None, frame=f if ambiguous else None, source_id=record.get('id'))
+                if ambiguous:
+                    view.build['unresolved'].append({'component': entity, 'status': 'source_semantics_unverified',
+                        'reason': 'Source pose retained with frame identity; no guessed association to same-class mask instances'})
+                label_key = (entity, label)
+                if label_key in per_frame:
+                    raise ValueError(f'ambiguous repeated object label {label}; source identity must be resolved')
+                per_frame.add(label_key)
+                records.setdefault(label_key, {})[f] = (record, locator(paths['annotations'], name, record=record.get('id')))
+        if category == 2 and any(entity == 'a' for entity, _ in records):
+            view.build['validation'].append({'component': 'C2_source_pose_mask_identity',
+                'rule': 'Official ObjTracking process_data.py reads dataList[0] with semantic mask label 1',
+                'source_revision': 'afa7bf37b02b7e33cf245acd08b98c207ccc175c',
+                'source_pose_id': 1, 'mask_value': 1, 'matched_frames': sum(len(by_frame) for (entity, _), by_frame in records.items() if entity == 'a')})
+        for (entity, label), by_frame in records.items():
+            part_labels = [k[1] for k in records if k[0] == entity]
+            articulated = definition['cad'].startswith('articulated/') and entity == definition['groups'][0][0]
+            part = 'pose_' + re.sub(r'[^a-zA-Z0-9_]', '_', label) if articulated or len(part_labels) > 1 else None
+            base = f'annotation/object/{entity}'
+            if part:
+                base += '/parts/' + part
+                objects[entity].setdefault('parts', {})[part] = dict(part_id=part, parent_id=None,
+                    has_model=False, mesh_ref=None, source_label=label)
+            poses, valid = np.full((n, 4, 4), np.nan, np.float32), np.zeros(n, bool)
+            dimensions = np.full((n, 3), np.nan, np.float32)
+            dv = np.zeros(n, bool)
+            for f, (record, _) in by_frame.items():
+                if all(k in record for k in ('rotation', 'center')):
+                    poses[f] = np.eye(4)
+                    poses[f, :3, :3] = Rotation.from_euler('XYZ', [record['rotation'][k] for k in ('x', 'y', 'z')]).as_matrix()
+                    poses[f, :3, 3] = [record['center'][k] for k in ('x', 'y', 'z')]
+                    valid[f] = True
+                if 'dimensions' in record:
+                    dimensions[f] = [record['dimensions'][k] for k in ('length', 'width', 'height')]
+                    dv[f] = True
+            key = entity + ('_' + part if part else '')
+            view.tensor('pose_' + key, 'transform/se3/v1' if part else 'object/rigid/v1', base + '/pose.pt', poses, valid,
+                ['T', 'row', 'column'], loc, convention='T_camera_from_part' if part else 'T_camera_from_object',
+                unit='m', source_label=label, source_euler_order='XYZ',
+                definition='https://github.com/hoi4d/ObjTracking/blob/main/process_data.py')
+            if dv.any():
+                view.tensor('dimensions_' + key, 'object/dimensions/v1', base + '/dimensions.pt', dimensions, dv,
+                    ['T', 'dimension'], loc, dimensions_semantics='source_unspecified', unit='m', scale_applied_to_mesh=False)
+        if cadz:
+            primary = definition['groups'][0][0]
+            if category != 2 and sum(g[2] == category for g in groups.values()) > 1:
+                primary = 'source_cad_instance'
+            instance = int(re.search(r'/N(\d+)/', seq)[1])
+            mesh_prefix = f'HOI4D_CAD_Model_for_release/{definition["cad"]}/{instance:03d}'
+            candidates = [name for name in cadz.namelist() if not name.endswith('/') and
+                          (name == mesh_prefix + '.obj' or name.startswith(mesh_prefix + '/'))]
+            if candidates:
+                # Retain OBJ material/texture dependencies beside their original names.
+                queue = list(candidates)
+                for name in queue:
+                    if Path(name).suffix.lower() not in ('.obj', '.mtl'):
+                        continue
+                    for line in cadz.read(name).decode().splitlines():
+                        fields = shlex.split(line, comments=True)
+                        if not fields:
+                            continue
+                        refs = fields[1:] if fields[0] == 'mtllib' else ([fields[-1]] if
+                            fields[0].startswith('map_') or fields[0] in ('bump', 'norm', 'disp', 'decal') else [])
+                        for ref in refs:
+                            dependency = posixpath.normpath(posixpath.join(posixpath.dirname(name), ref))
+                            if dependency not in cadz.NameToInfo:
+                                raise ValueError(f'missing CAD dependency: {dependency}')
+                            if dependency not in candidates:
+                                candidates.append(dependency)
+                                queue.append(dependency)
+                ensure_object(primary, locator(paths['cad'], mesh_prefix), category)
+                asset_dir = view.path / f'annotation/object/{primary}/assets'
+                asset_refs = []
+                for name in candidates:
+                    relative = posixpath.relpath(name, mesh_prefix if definition['cad'].startswith('articulated/') else posixpath.dirname(mesh_prefix))
+                    dest = safe_member(asset_dir, relative)
+                    dest.parent.mkdir(parents=True, exist_ok=True)
+                    dest.write_bytes(cadz.read(name))
+                    ref = dest.relative_to(view.path).as_posix()
+                    if dest.suffix.lower() == '.obj':
+                        asset_key = re.sub(r'[^a-zA-Z0-9_]', '_', relative)
+                        view.component('mesh_' + primary + '_' + asset_key, 'mesh/obj/v1', ref, ['vertex'],
+                            locator(paths['cad'], name), unit='m', valid=True,
+                            scale_applied_to_mesh=False, dimensions_semantics='source_unspecified')
+                        asset_refs.append(ref)
+                        if definition['cad'].startswith('articulated/'):
+                            part_id = 'asset_' + asset_key
+                            objects[primary].setdefault('parts', {})[part_id] = dict(part_id=part_id,
+                                parent_id=None, has_model=True, mesh_ref=ref, source_asset=relative,
+                                note='Source asset; correspondence to motion part labels remains unverified')
+                for asset_ref in asset_refs:
+                    mesh_check = check_mesh(view.path / asset_ref, asset_dir)
+                    view.build['validation'].append({'component': asset_ref, **mesh_check,
+                        'unit_basis': 'official ObjTracking loads source canonical CAD without scaling in metre camera coordinates',
+                        'geometric_acceptance': 'physical dimensions-to-mesh relation remains source_unspecified'})
+                metadata_path = view.path / f'annotation/object/{primary}/metainfo.json'
+                metadata = json.loads(metadata_path.read_text())
+                metadata.update(has_model=bool(asset_refs), asset_refs=asset_refs)
+                if definition['cad'].startswith('rigid/') and len(asset_refs) == 1:
+                    metadata['mesh_ref'] = asset_refs[0]
+                json_write(metadata_path, metadata)
+                if definition['cad'].startswith('articulated/'):
+                    view.build['unresolved'].append({'component': 'object_part_asset_association',
+                        'status': 'source_semantics_unverified',
+                        'reason': 'All source assets and motion labels retained, but no guessed correspondence or joint_state'})
+        for entity, info in objects.items():
+            if 'parts' in info:
+                ref = f'annotation/object/{entity}/articulation.json'
+                json_write(view.path / ref, {'parts': list(info['parts'].values()), 'joints': [],
+                    'hierarchy_semantics': 'semantic_containment',
+                    'note': 'Unassociated source region, motion and asset nodes; original CAD trees retained in assets'})
+                view.component('articulation_' + entity, 'object/articulated/v1', ref, [], loc)
+                meta_path = view.path / f'annotation/object/{entity}/metainfo.json'
+                metadata = json.loads(meta_path.read_text())
+                metadata['articulation_ref'] = ref
+                json_write(meta_path, metadata)
+        if depth is not None:
+            dpath, dloc = depth
+            dmeta = probe(dpath)
+            if [dmeta['width'], dmeta['height']] != view.meta['camera']['image_size']:
+                raise ValueError('HOI4D depth grid differs from RGB')
+            if dmeta['pix_fmt'] != 'gray16le':
+                raise ValueError('unexpected HOI4D depth pixel format')
+            width, height = view.meta['camera']['image_size']
+            raw = work / 'depth.raw'
+            limits = ['-frames:v', str(n)] if args.max_frames else []
+            run(decode_args(dpath, 'gray16le') + limits + ['-y', str(raw)])
+            if raw.stat().st_size != n * width * height * 2:
+                raise ValueError('depth/RGB frame count or grid mismatch')
+            def depths():
+                with raw.open('rb') as inp:
+                    for _ in range(n):
+                        yield np.frombuffer(inp.read(width * height * 2), '<u2').reshape(height, width).astype(np.float32) / 1000
+            view.raster('depth', depths(), None, dloc | {'conversion': 'registered millimetre camera-Z to metres'})
+        else:
+            if find_file(args.root, 'DEPTH'):
+                raise ValueError('DEPTH source exists, but no matching depth video was found')
+            view.absent('depth', 'Depth source not downloaded', 'not_downloaded')
+        return view.finish()
+
+
+def main(argv=None):
+    p = parser('HOI4D', __doc__)
+    p.set_defaults(release_id='shared-v1.0')
+    # Four workers reached 21 observed OS threads including timing; stay below the 24-thread budget.
+    p.add_argument('--workers', type=int, choices=range(1, 5), default=1,
+                   help='Independent views in parallel, 1..4; codec/BLAS threads fixed at 1')
+    p.add_argument('--tasks', help='Comma-separated substrings of source sequence paths')
+    p.add_argument('--sessions', help='Comma-separated sessions')
+    p.add_argument('--categories', help='Comma-separated category numbers')
+    args = p.parse_args(argv)
+    if args.views and args.views != 'ego':
+        p.error('HOI4D source camera ID is ego')
+    paths = {key: find_file(args.root, name) for key, name in {
+        'release': 'HOI4D_release.zip', 'annotations': 'HOI4D_annotations.zip', 'hand': 'HOI4D_Hand_pose.zip',
+        'camera': 'camera_params.zip', 'cad': 'HOI4D_CAD_Model_for_release.zip'}.items()}
+    if not paths['release'] or not paths['annotations']:
+        p.error('HOI4D release and annotation ZIPs are required')
+    out = release(args, 'HOI4D')
+    mapping = Path(__file__).resolve().parent.parent / 'docs/hoi4d-RAW_TO_SHARED.md'
+    published_mapping = out / 'doc/RAW_TO_SHARED.md'
+    if published_mapping.exists() and published_mapping.read_bytes() != mapping.read_bytes():
+        raise ValueError('RAW_TO_SHARED.md changed; choose a new release-id')
+    if not published_mapping.exists():
+        shutil.copyfile(mapping, published_mapping)
+    torch.set_num_threads(1)
+    models = mano_models(args.mano_dir)
+    with ExitStack() as stack, tempfile.TemporaryDirectory(prefix='hoi4d-depth-') as wd:
+        archives = {key: stack.enter_context(zipfile.ZipFile(path)) if path else None for key, path in paths.items()}
+        sequences = sorted(name.removeprefix('HOI4D_release/').removesuffix('/align_rgb/image.mp4')
+                           for name in archives['release'].namelist() if name.endswith('/align_rgb/image.mp4'))
+        if args.tasks:
+            sequences = [s for s in sequences if any(token in s for token in args.tasks.split(','))]
+        if args.sessions:
+            sequences = [s for s in sequences if s.split('/')[0] in args.sessions.split(',')]
+        if args.categories:
+            sequences = [s for s in sequences if re.search(r'/C(\d+)/', s)[1] in args.categories.split(',')]
+        if args.limit:
+            sequences = sequences[:args.limit]
+        if not sequences:
+            p.error('no matching source sequences')
+        depths, _ = extract_depths(args.root, sequences, Path(wd))
+        worker = threading.local()
+        worker_stacks = []
+        def init_worker():
+            owned = ExitStack()
+            try:
+                # ZipFile.read mutates a shared seek position: never share a handle across workers.
+                worker.archives = {key: owned.enter_context(zipfile.ZipFile(path)) if path else None
+                                   for key, path in paths.items()}
+                index_frames(worker.archives['annotations'])
+                if worker.archives['hand']:
+                    index_frames(worker.archives['hand'])
+            except BaseException:
+                owned.close()
+                raise
+            worker_stacks.append(owned)
+        def convert_one(seq):
+            source = archives if args.workers == 1 else worker.archives
+            name = seq.replace('/', '_')
+            with sequence_dir(out, name) as target:
+                report = convert_sequence(seq, args, target / 'ego', source['release'], source['annotations'],
+                    source['hand'], source['camera'], source['cad'], paths, models, depths.get(seq))
+            build_record(out, name, {'ego': report})
+            return seq
+        if args.workers == 1:
+            for seq in map(convert_one, sequences):
+                print(f'Converted {seq}', flush=True)
+        else:
+            try:
+                with ThreadPoolExecutor(max_workers=args.workers, initializer=init_worker) as pool:
+                    for seq in pool.map(convert_one, sequences):
+                        print(f'Converted {seq}', flush=True)
+            finally:
+                for owned in worker_stacks:
+                    owned.close()
+    return 0
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())
+
+# TODO(v2.0): 旧 bbox IoU 实例重分配属于推断，注释保留，不在 v1.0 执行。
+#                         # 同类多实例（C8 双桶 / C2 双杯）：2dBox 可用时按 IoU 复核归属
+#                         cls = next(x[3] for x in cat["groups"] if x[0] == gk)
+#                         sibs = [g for g in cat["groups"]
+#                                 if g[0] != gk and cls is not None and g[3] == cls]
+#                         box = (d.get("2dBox") or {}).get("camera1")
+#                         box = (box[0][0], box[0][1], box[1][0], box[1][1]) \
+#                             if box and len(box) == 2 else None
+#                         if box is not None and sibs:
+#                             gvals = next(x[1] for x in cat["groups"] if x[0] == gk)
+#                             best, best_iou = gk, box_iou(box, bbox_of(mask, gvals))
+#                             for s in sibs:
+#                                 if ("obj", s[0]) not in tid_of_group:
+#                                     continue
+#                                 i = box_iou(box, bbox_of(mask, s[1]))
+#                                 if i > best_iou + 0.1 and i > 0.15:
+#                                     best, best_iou = s[0], i
+#                             if best != gk:
+#                                 gk, iou_fixups = best, iou_fixups + 1

H2O/shared-v1.0/script/convert_visor_v1.0.py

@@ -0,0 +1,271 @@
+#!/usr/bin/env python3
+"""VISOR source polygons and contact → CONTRACT.md v1.0, without rasterization or side inference."""
+from collections import Counter
+import hashlib
+import json
+import re
+import shutil
+import tempfile
+from pathlib import Path
+import zipfile
+
+import numpy as np
+import torch
+
+from warehouse_v1_0 import (View, build_record, find_file, jsonl_write, locator, parser,
+                            release, roots, sequence_dir, source)
+
+FRAME = re.compile(r'frame_(\d+)')
+HAND = {300: 'left', 301: 'right'}
+GLOVE = {303: 'left', 304: 'right'}
+STATES = {'hand-not-in-contact': 'negative', 'glove-not-in-contact': 'negative',
+          'none-of-the-above': 'target_unannotated', 'inconclusive': 'unknown'}
+
+
+def hand_side(ann):
+    if ann['class_id'] in HAND:
+        return HAND[ann['class_id']]
+    if ann['class_id'] not in (60, 303, 304):
+        return None
+    raw = ann.get('on_which_hand')
+    raw = raw if isinstance(raw, list) else [raw]
+    sides = set()
+    for item in raw:
+        if item == 'left hand':
+            sides.add('left')
+        if item in ('right hand', 'rigth hand'):
+            sides.add('right')
+    if len(sides) == 1:
+        return next(iter(sides))
+    if not sides and ann['class_id'] in GLOVE:
+        return GLOVE[ann['class_id']]
+    return None
+
+
+def convert_video(video, split, annotation, ann_loc, rgb_zip, rgb_loc, mapping, args, target):
+    indexed = sorted(enumerate(annotation['video_annotations']), key=lambda row: int(FRAME.search(row[1]['image']['name'])[1]))
+    if args.max_frames:
+        indexed = indexed[:args.max_frames]
+    if len({row['image']['name'] for _, row in indexed}) != len(indexed):
+        raise ValueError('duplicate VISOR frame')
+    keys = lambda row, ann: (row['image']['subsequence'], ann['class_id'], ann['name'])
+    collisions = set()
+    for _, row in indexed:
+        collisions.update(key for key, number in Counter(keys(row, a) for a in row['annotations']).items() if number > 1)
+    frames, paths = [], []
+    with tempfile.TemporaryDirectory(prefix='visor-rgb-') as wd, zipfile.ZipFile(rgb_zip) as z:
+        work = Path(wd)
+        for original_index, row in indexed:
+            name = row['image']['name']
+            data = z.read(name)
+            path = work / Path(name).name
+            path.write_bytes(data)
+            paths.append(path)
+            loc = rgb_loc | {'nested_member': name, 'sha256': hashlib.sha256(data).hexdigest()}
+            frames.append(dict(source_frame_id=int(FRAME.search(name)[1]), source_locator=loc,
+                source_image=row['image'], original_frame_name=mapping.get(name),
+                annotation_matches={'segment': dict(source_frame_id=name, source_timestamp_ns=None,
+                    clock_domain=None, match_method='index_map', delta_ns=None,
+                    source_locator=ann_loc | {'record': f'video_annotations[{original_index}]'})}))
+        view = View(target, 'EPIC-KITCHENS-VISOR', args.release_id, video, 'ego', frames, sampling='sparse', split=split)
+        view.valid_rules = {
+            'segment': 'Selected sparse annotation frame present; all source polygons finite with at least three vertices and entity mapping valid',
+            'relations': 'At least one source in_contact_object value present; positive targets resolve in the same frame; unknown is an explicit usable status',
+        }
+        view.full_jpeg = getattr(args, 'jpeg_full_check', False)
+        view.meta.update(view_type='ego', split_scheme='VISOR_GroundTruth_SparseAnnotations', split_source_locator=ann_loc)
+        view.meta['sampling']['selection'] = {'representation': 'official sparse annotations', 'max_frames': args.max_frames}
+        view.rgb(paths, rgb_loc)
+    entities = {}
+    polygons, mappings, relations, relation_valid = [], [], [], []
+    for f, (original_index, row) in enumerate(indexed):
+        frame_loc = ann_loc | {'record': f'video_annotations[{original_index}]'}
+        id_map, labels, polys = {}, [], []
+        for code, ann in enumerate(row['annotations'], 1):
+            if code >= 65535:
+                raise ValueError('too many source instances')
+            sid = ann['id']
+            if sid in id_map:
+                raise ValueError('duplicate source instance ID within a frame')
+            side = hand_side(ann)
+            if side is not None:
+                entity = 'hand_' + side
+            else:
+                key = keys(row, ann)
+                identity = key + (sid, row['image']['name']) if key in collisions else key
+                entity = 'object_' + hashlib.sha256(json.dumps(identity).encode()).hexdigest()[:20]
+                if entity not in entities:
+                    scope = 'frame' if key in collisions else 'subsequence'
+                    scope_id = f'{video}:{row["image"]["name"]}' if scope == 'frame' else f'{video}:{key[0]}'
+                    view.object(entity, frame_loc | {'source_instance_id': sid}, scope, scope_id,
+                        category_id=ann['class_id'], name=ann['name'],
+                        association_basis='source_frame_instance' if scope == 'frame' else 'official_subsequence_entity_name')
+                    entities[entity] = True
+            id_map[sid] = entity
+            for polygon in ann.get('segments', []):
+                p = np.asarray(polygon)
+                if p.ndim != 2 or p.shape[1] != 2 or len(p) < 3 or not np.isfinite(p).all():
+                    raise ValueError(f'invalid source polygon {sid}')
+            polys.append(dict(mask_value=code, polygons=ann.get('segments', [])))
+            labels.append(dict(mask_value=code, category_id=ann['class_id'], entity_id=entity,
+                               part_id=None, source_instance_id=sid,
+                               source_attributes={k: v for k, v in ann.items() if k != 'segments'}))
+        rr = []
+        provided = False
+        for ann in row['annotations']:
+            raw = ann.get('in_contact_object')
+            values = raw if isinstance(raw, list) else [raw]
+            for value in values:
+                if value is None:
+                    continue
+                provided = True
+                status = STATES.get(value)
+                obj = None
+                if status is None:
+                    if value not in id_map:
+                        raise ValueError(f'unresolved source contact target {value} in {video} frame {f}')
+                    status, obj = 'positive', id_map[value]
+                rr.append(dict(relation_type='contact', subject_id=id_map[ann['id']], object_id=obj,
+                    status=status, source_value=value, source_locator=frame_loc | {'source_instance_id': ann['id']},
+                    provenance=source('Resolve source contact ID against instances in this frame')))
+        polygons.append(dict(frame_index=f, instances=polys))
+        mappings.append(dict(frame_index=f, labels=labels))
+        relations.append(dict(frame_index=f, relations=rr))
+        relation_valid.append(provided)
+    jsonl_write(view.path / 'annotation/segment/polygons.jsonl', polygons)
+    jsonl_write(view.path / 'annotation/segment/instances.jsonl', mappings)
+    torch.save(torch.ones(len(frames), dtype=torch.bool, device='cpu'), view.path / 'annotation/segment/mask_valid.pt')
+    view.component('segment', 'segment/instance_polygons/v1', 'annotation/segment/polygons.jsonl', ['T'], ann_loc,
+        valid_ref='annotation/segment/mask_valid.pt', valid_axes=['T'], segmentation_type='instance',
+        mapping_ref='annotation/segment/instances.jsonl', coverage={'kind': 'unknown', 'exhaustive_ref': 'annotation/segment/instances.jsonl'})
+    view.relations(relations, relation_valid, ann_loc)
+    for component in ('depth', 'mano', 'keypoints_uv', 'keypoints_xyz', 'intrinsic', 'extrinsic'):
+        view.absent(component, 'Not provided by VISOR sparse source annotations')
+    return view.finish()
+
+
+def main(argv=None):
+    p = parser('EPIC-KITCHENS-VISOR', __doc__)
+    p.add_argument('--videos', help='Comma-separated video names')
+    p.add_argument('--splits', default='train,val')
+    args = p.parse_args(argv)
+    wanted = set(args.videos.split(',')) if args.videos else None
+    splits = set(args.splits.split(','))
+    if not splits <= {'train', 'val'}:
+        p.error('only annotated train/val splits are available')
+    out = release(args, 'EPIC-KITCHENS-VISOR')
+    archive = find_file(args.root, 'EPIC-KITCHENS VISOR.zip')
+    with tempfile.TemporaryDirectory(prefix='visor-source-') as wd:
+        work = Path(wd)
+        if archive:
+            with zipfile.ZipFile(archive) as z:
+                suffix = 'GroundTruth-SparseAnnotations/annotations/'
+                members = [n for n in z.namelist() if suffix in n and n.endswith('.json')]
+                prefix = members[0].split(suffix)[0]
+                fm = json.loads(z.read(prefix + 'frame_mapping.json'))
+                done = 0
+                for member in sorted(members):
+                    split, filename = member.split(suffix)[1].split('/')
+                    video = Path(filename).stem
+                    if split not in splits or wanted and video not in wanted:
+                        continue
+                    data = z.read(member)
+                    rgb_member = prefix + f'GroundTruth-SparseAnnotations/rgb_frames/{split}/{video.split("_")[0]}/{video}.zip'
+                    rgb_zip = work / (video + '.zip')
+                    with z.open(rgb_member) as inp, rgb_zip.open('wb') as dest:
+                        shutil.copyfileobj(inp, dest)
+                    with sequence_dir(out, video) as dest:
+                        report = convert_video(video, split, json.loads(data), locator(archive, member, data),
+                            rgb_zip, locator(archive, rgb_member), fm.get(video, {}), args, dest / 'ego')
+                    build_record(out, video, {'ego': report})
+                    rgb_zip.unlink()
+                    done += 1
+                    print(f'Converted {video}', flush=True)
+                    if args.limit and done >= args.limit:
+                        break
+        else:
+            root = next((r for r in roots(args.root) if (r / 'GroundTruth-SparseAnnotations').is_dir()), None)
+            if root is None:
+                p.error('VISOR outer ZIP or extracted source directory not found')
+            fm = json.loads((root / 'frame_mapping.json').read_text()) if (root / 'frame_mapping.json').exists() else {}
+            done = 0
+            for split in sorted(splits):
+                for ann in sorted((root / 'GroundTruth-SparseAnnotations/annotations' / split).glob('*.json')):
+                    video = ann.stem
+                    if wanted and video not in wanted:
+                        continue
+                    rgb = root / f'GroundTruth-SparseAnnotations/rgb_frames/{split}/{video.split("_")[0]}/{video}.zip'
+                    with sequence_dir(out, video) as dest:
+                        report = convert_video(video, split, json.loads(ann.read_text()), locator(ann), rgb,
+                            locator(rgb), fm.get(video, {}), args, dest / 'ego')
+                    build_record(out, video, {'ego': report})
+                    done += 1
+                    if args.limit and done >= args.limit:
+                        return 0
+    if not done:
+        p.error('no matching annotated videos found')
+    return 0
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())
+
+# TODO(v2.0): 旧距离猜侧和单层掩码生成仅注释保留；v1.0 不执行。
+#                     if side is None and len(frame_hands) == 1:
+#                         # 空间门槛：脱下放在远处的手套不并入（实测双峰 0..25px / ≥44px）
+#                         hb = frame_hands[0][1]
+#                         gb = _bbox_of(polys)
+#                         if hb and gb and _bbox_dist(gb, hb) <= GLOVE60_MERGE_BBOX_DIST:
+#                             side = frame_hands[0][0]
+#                             stats["glove60_merged_into_sole_hand"] += 1
+#                     if side is None:
+#                         stats["dropped_glove60_unresolved"] += 1
+#                         id_disp[mid] = ("dropped", "glove60")
+#                         continue
+# def rasterize_and_encode(vdir, frames, frame_draw, zip_path, n, stats):
+#     """流式读 zip 内 jpg + 光栅化多边形 → rgb.mkv / mask.mkv；返回 (w, h)。"""
+#     from PIL import Image, ImageDraw
+#
+#     with zipfile.ZipFile(zip_path) as zf:
+#         member_of = {}
+#         for nm in zf.namelist():
+#             member_of[os.path.basename(nm)] = nm
+#         names = [im["name"] for im in frames]
+#         missing = [nm for nm in names if nm not in member_of]
+#         if missing:
+#             raise RuntimeError(f"{len(missing)} annotated frames missing in zip, "
+#                                f"e.g. {missing[:3]}")
+#         stats["zip_extra_members"] = len(member_of) - len(names)
+#
+#         # 首帧定分辨率（官方 1920×1080，实测每帧校验）
+#         img0 = Image.open(io.BytesIO(zf.read(member_of[names[0]]))).convert("RGB")
+#         w, h = img0.size
+#
+#         bin_img = Image.new("L", (w, h), 0)
+#         bin_draw = ImageDraw.Draw(bin_img)
+#         canvas = np.zeros((h, w), np.uint16)
+#         with VideoPipe(vdir / "mask.mkv", w, h, FPS, "mask") as mw, \
+#                 VideoPipe(vdir / "rgb.mkv", w, h, FPS, "rgb") as rw:
+#             for f in range(n):
+#                 canvas[:] = 0
+#                 for pid, polys in frame_draw[f]:
+#                     bin_draw.rectangle((0, 0, w, h), fill=0)
+#                     for poly in polys:
+#                         bin_draw.polygon([(float(x), float(y)) for x, y in poly], fill=255)
+#                     arr = np.asarray(bin_img)
+#                     canvas[arr > 0] = pid          # 后画覆盖先画：物体→手（modal）
+#                 mw.write(canvas)
+#                 img = img0 if f == 0 else Image.open(
+#                     io.BytesIO(zf.read(member_of[names[f]]))).convert("RGB")
+#                 if img.size != (w, h):
+#                     raise RuntimeError(f"frame {names[f]}: size {img.size} != {(w, h)}")
+#                 rw.write(np.asarray(img))
+#             n_mask, n_rgb = mw.n, rw.n
+#     if (n_rgb, n_mask) != (n, n):
+#         raise RuntimeError(f"wrote rgb={n_rgb} mask={n_mask} != {n}")
+#     got = (count_frames(vdir / "rgb.mkv"), count_frames(vdir / "mask.mkv"))
+#     if got != (n, n):
+#         raise RuntimeError(f"encoded rgb={got[0]} mask={got[1]} != {n} (decode verify)")
+#     return w, h
+#
+#

H2O/shared-v1.0/script/provenance/operations/h2o-check-fallback.py

@@ -0,0 +1,137 @@
+"""Full coverage/numeric audit, plus representative persisted-media QC and offline HTML."""
+import os
+for key in ('OMP_NUM_THREADS','OPENBLAS_NUM_THREADS','MKL_NUM_THREADS','NUMEXPR_NUM_THREADS'):
+    os.environ[key]='1'
+os.environ['HAND_WAREHOUSE_FFMPEG_THREADS']='4'
+os.environ.pop('HAND_WAREHOUSE_FFV1_SLICES',None)
+os.environ.pop('HAND_WAREHOUSE_MEDIA_TMPDIR',None)
+import collections,datetime,hashlib,html,json,re,sys,time,zipfile
+from concurrent.futures import ThreadPoolExecutor,as_completed
+from pathlib import Path
+from zoneinfo import ZoneInfo
+import numpy as np
+import torch
+root=Path('/data/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-process16-pipe')
+release=root/'H2O/shared-v1.0'
+sys.path.insert(0,str(root/'frozen-code/scripts'))
+from warehouse_v1_0 import SCHEMA,check_transform,check_valid_policy,check_view_type,safe_member,sha256
+from render_h2o_validation import render
+out=release/'doc/validation/final-quick-check'
+out.mkdir(exist_ok=True)
+torch.set_num_threads(1)
+started=time.monotonic()
+expected={}
+with zipfile.ZipFile('/data/xuzhefeng/Datasets/Hand-Datasets-Raw/H2O/raw/data/label_split.zip') as z:
+    for name in z.namelist():
+        if '/action_' not in name:continue
+        for line in z.read(name).decode().splitlines():
+            fields=line.split()
+            if len(fields)>1 and fields[1]!='path':
+                seq=fields[1].replace('/','_')
+                expected[seq]=max(expected.get(seq,0),int(fields[-1])+1)
+folders={p.name:p for p in (release/'data').iterdir() if p.is_dir()}
+assert set(folders)==set(expected),(sorted(set(expected)-set(folders)),sorted(set(folders)-set(expected)))
+assert len(expected)==184 and sum(expected.values())==114329
+receipts={p.name.removesuffix('.build.json'):p for p in (release/'doc').glob('*.build.json')}
+assert set(receipts)==set(expected)
+assert 'EXIT=0' in (root/'full.status').read_text()
+failures=[]
+rows=[]
+def require(ok,message):
+    if not ok:raise ValueError(message)
+def load(path):
+    def cpu(storage,location):
+        require(location=='cpu',f'non-CPU storage: {path}')
+        return storage
+    obj=torch.load(path,map_location=cpu,weights_only=True)
+    require(type(obj) is torch.Tensor,f'non-tensor: {path}')
+    return obj.numpy()
+def check_sequence(seq):
+    folder=folders[seq];n=expected[seq]
+    require({p.name for p in folder.iterdir() if p.is_dir()}=={f'cam{i}' for i in range(5)},'five-camera coverage')
+    receipt=json.loads(receipts[seq].read_text())
+    require(receipt['sequence_id']==seq and set(receipt['views'])=={f'cam{i}' for i in range(5)},'build receipt')
+    result=[]
+    for i in range(5):
+        cam=f'cam{i}';view=folder/cam
+        meta=json.loads((view/'metainfo.json').read_text())
+        require((meta['dataset_id'],meta['sequence_id'],meta['camera_id'],meta['num_frames'])==('H2O',seq,cam,n),'header/timeline')
+        require(meta['view_type']==('ego' if i==4 else 'exo'),'view role')
+        require(meta['sampling']['selection']=={'max_frames':None,'views':None},'unexpected subsampling')
+        check_view_type(meta)
+        components=meta['components']
+        require(all(components[k].get('available') for k in ('rgb','depth','mano','keypoints_xyz','intrinsic','extrinsic')),'required modality absent')
+        require(not components['segment']['available'] and not components['relations']['available'],'unexpected synthesized component')
+        rgb_sources=[];depth_sources=[];rgb_count=depth_count=0
+        with (view/'camera/frames.jsonl').open() as inp:
+            count=0
+            for index,line in enumerate(inp):
+                f=json.loads(line)
+                require(f['frame_index']==index and f['source_frame_id']==index,'source frame index')
+                require(f['sync_id']==f'{seq.replace("_","/")}:{index}','synchronization ID')
+                sources=f['component_sources']
+                for kind,loc in sources.items():
+                    require(re.fullmatch('[0-9a-f]{64}',loc.get('sha256','')) is not None,f'source SHA256: {kind}')
+                    require(f'/{cam}/' in loc.get('member',''),'wrong source camera')
+                rgb_sources.append('rgb' in sources);depth_sources.append('depth' in sources)
+                rgb_count+=rgb_sources[-1];depth_count+=depth_sources[-1];count+=1
+        require(count==n,'frames.jsonl length')
+        numeric=0
+        object_names=[]
+        for key,c in components.items():
+            if not c['available']:continue
+            for field in ('data_ref','valid_ref','mapping_ref'):
+                if c.get(field):
+                    path=safe_member(view,c[field])
+                    require(path.is_file() and path.stat().st_size>0,f'missing/empty {c[field]}')
+            schema=SCHEMA['components'][c['schema_id']]
+            if c.get('valid_policy'):check_valid_policy(c,n)
+            valid=load(view/c['valid_ref']) if c.get('valid_ref','').endswith('.pt') else None
+            if valid is not None:
+                shape=schema.get('valid_shape')
+                require(valid.dtype==np.bool_ and valid.shape[0]==n,f'{key}: validity dtype/timeline')
+                if shape:require(len(shape)==valid.ndim and all(not isinstance(s,int) or valid.shape[j]==s for j,s in enumerate(shape)),f'{key}: validity shape')
+                if key=='rgb':require(np.array_equal(valid,np.asarray(rgb_sources)), 'RGB existence validity mismatch')
+            if c['data_ref'].endswith('.pt'):
+                arr=load(view/c['data_ref']);numeric+=1
+                require(arr.dtype==np.float32 and arr.shape[0]==n,f'{key}: dtype/timeline')
+                shape=schema.get('shape')
+                if shape:require(len(shape)==arr.ndim and all(not isinstance(s,int) or arr.shape[j]==s for j,s in enumerate(shape)),f'{key}: tensor shape')
+                if valid is not None:
+                    require(arr.shape[:valid.ndim]==valid.shape and np.isfinite(arr[valid]).all() and np.isnan(arr[~valid]).all(),f'{key}: finite/NaN validity semantics')
+                    if c['schema_id'] in ('transform/se3/v1','object/rigid/v1'):check_transform(arr[valid])
+                if key=='intrinsic':require(np.allclose(arr[:,2],[0,0,1]) and (arr[:,0,0]>0).all() and (arr[:,1,1]>0).all(),'camera intrinsics')
+            if key.startswith('mesh_'):object_names.append(key.removeprefix('mesh_'))
+        proof=receipt['views'][cam]
+        require(not proof['unresolved'],'unresolved QC components')
+        media={}
+        for p in proof['validation']:
+            if p.get('profile')==components['rgb']['storage']['profile'] and p.get('all_frames_exact') is True:media['rgb']=p
+            elif p.get('component') in ('depth','depth_valid'):media[p['component']]=p
+        require(set(media)=={'rgb','depth','depth_valid'},'media roundtrip proof missing')
+        for key,p in media.items():
+            require(p.get('all_frames_exact') is True and re.fullmatch('[0-9a-f]{64}',p.get('decoded_sha256','')) is not None,f'{key}: failed original roundtrip proof')
+            require((view/'observation'/f'{key}.mkv').stat().st_size==p['bytes'],f'{key}: encoded file size changed since conversion')
+        result.append({'sequence_id':seq,'camera':cam,'frames':n,'numeric_components':numeric,
+                       'rgb_source_frames':rgb_count,'depth_source_frames':depth_count,'objects':sorted(object_names)})
+    return result
+
+result=check_sequence('subject1_k2_6')
+from warehouse_v1_0 import decoded_digest
+receipt=json.loads(receipts['subject1_k2_6'].read_text())
+view=folders['subject1_k2_6']/'cam0'
+meta=json.loads((view/'metainfo.json').read_text())
+proofs=receipt['views']['cam0']['validation']
+verified=[]
+for name,pix in (('rgb','rgb24'),('depth','gray16le'),('depth_valid','gray')):
+    p=next(x for x in proofs if (x.get('profile')==meta['components']['rgb']['storage']['profile'] if name=='rgb' else x.get('component')==name))
+    digest,size,seconds=decoded_digest(view/'observation'/f'{name}.mkv',pix)
+    assert digest==p['decoded_sha256'],(name,digest,p['decoded_sha256'])
+    expected_bytes=expected['subject1_k2_6']*1280*720*({'rgb':3,'depth':4,'depth_valid':1}[name])
+    assert size==expected_bytes,(name,size,expected_bytes)
+    verified.append({'component':name,'decoded_sha256':digest,'decoded_bytes':size,'seconds':seconds})
+page=render(view,out/'subject1_k2_6_cam0_fallback')
+json_result={'passed':True,'views':result,'full_decode_comparison':verified,'html':str(page.relative_to(out)),
+             'note':'Registered lossless H.264 fallback; all current decoded media digests equal original conversion receipts'}
+(out/'H264_FALLBACK_CHECK.json').write_text(json.dumps(json_result,indent=2)+'\n')
+print('FALLBACK_PASS',json.dumps(json_result),flush=True)

H2O/shared-v1.0/script/provenance/operations/h2o-finalize-quick-validation.py

@@ -0,0 +1,39 @@
+import datetime,html,json,shutil
+from pathlib import Path
+from zoneinfo import ZoneInfo
+release=Path('/data/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-process16-pipe/H2O/shared-v1.0')
+out=release/'doc/validation/final-quick-check'
+report=json.loads((out/'QUICK_VALIDATION.json').read_text())
+assert report['failures']==[{'sequence':'subject1_k2_6','error':"ValueError('rgb: failed original roundtrip proof')"}]
+fallback=json.loads((out/'H264_FALLBACK_CHECK.json').read_text());assert fallback['passed']
+shutil.copy2(out/'QUICK_VALIDATION.json',out/'INITIAL_CHECKER_ALERT.json')
+rows=fallback['views']
+report['checked_sequences']+=1;report['checked_views']+=len(rows)
+report['checked_view_frames']+=sum(r['frames'] for r in rows)
+report['rgb_source_frames']+=sum(r['rgb_source_frames'] for r in rows)
+report['depth_source_frames']+=sum(r['depth_source_frames'] for r in rows)
+report['numeric_components']+=sum(r['numeric_components'] for r in rows)
+report['samples'].append({'sequence_id':'subject1_k2_6','camera':'cam0','object':'espresso',
+                          'frames':678,'passed':True,'html':fallback['html'],
+                          'extra_check':'Full RGB/depth/valid current decoded digests equal original receipts'})
+report['resolved_alerts']=[{'view':'subject1_k2_6/cam0','classification':'Registered lossless H.264 fallback, not missing data or corruption',
+    'original_checker_limitation':'First pass selected only FFV1 proofs; fixed to select the profile actually declared by metainfo',
+    'followup':'All five cameras rechecked; problematic view RGB, depth and validity fully decoded and matched original SHA256/byte length'}]
+report['failures']=[];report['passed']=True
+report['time']=datetime.datetime.now(ZoneInfo('Asia/Shanghai')).isoformat()
+report['sample_views']=len(report['samples'])
+assert report['checked_sequences']==184 and report['checked_views']==920
+assert report['checked_view_frames']==report['rgb_source_frames']==report['depth_source_frames']==571645
+report['rgb_profiles']={}
+for meta in (release/'data').glob('*/cam*/metainfo.json'):
+    name=json.loads(meta.read_text())['components']['rgb']['storage']['profile']
+    report['rgb_profiles'][name]=report['rgb_profiles'].get(name,0)+1
+(out/'QUICK_VALIDATION.json').write_text(json.dumps(report,ensure_ascii=False,indent=2)+'\n')
+summary={k:v for k,v in report.items() if k!='samples'}
+(out/'index.html').write_text('<!doctype html><html lang="zh"><meta charset="utf-8"><title>H2O 全量后抽样验收</title><h1>H2O 全量后抽样验收</h1><p>自动检查通过不等于原始几何标注准确。覆盖四个 subject、八类物体、ego/exo；青色为源手部关节点投影，粉色为源物体点投影，仅用于预览。</p><ul>'+''.join('<li><a href="'+html.escape(x['html'],quote=True)+'">'+html.escape(x['sequence_id']+' '+x['camera']+' '+x['object'])+'</a></li>' for x in sorted(report['samples'],key=lambda x:(x['object'],x['sequence_id'],x['camera'])))+'</ul><pre>'+html.escape(json.dumps(summary,ensure_ascii=False,indent=2))+'</pre></html>')
+operations=release/'script/provenance/operations'
+for src in ('h2o-quick-validation.py','h2o-check-fallback.py','h2o-finalize-quick-validation.py'):
+    shutil.copy2(Path('/tmp')/src,operations/src)
+shutil.copy2('/tmp/h2o-quick-validation.log',out/'INITIAL_RUN.log')
+shutil.copy2('/tmp/h2o-fallback-check.log',out/'FALLBACK_RUN.log')
+print(json.dumps(summary,ensure_ascii=False,indent=2),flush=True)

H2O/shared-v1.0/script/provenance/operations/h2o-pipeline-bench-job.sh

@@ -0,0 +1,25 @@
+#!/usr/bin/env bash
+set -u
+cd /home/xuzhefeng/Projects/uni-hoi-dataset || exit 1
+case "$1" in
+ old) code=/data/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-w4-c10-d10/frozen-code/scripts/convert_h2o_v1.0.py; extra=();;
+ v8) code=/tmp/h2o-stream-v8-recovery/scripts/convert_h2o_v1.0.py; extra=(--view-workers 8 --stream-media);;
+ v8ram) code=/tmp/h2o-stream-v8ram-recovery/scripts/convert_h2o_v1.0.py; extra=(--view-workers 8 --stream-media --ram-stage-dir /dev/shm/xuzhefeng-h2o-pipeline-v8ram-20260930 --ram-stage-gib 32);;
+ s16) code=/tmp/h2o-stream-s16-recovery/scripts/convert_h2o_v1.0.py; extra=(--view-workers 16 --stream-media --ffv1-slices 16 --ram-stage-dir /dev/shm/xuzhefeng-h2o-pipeline-s16-20260930 --ram-stage-gib 32);;
+ pipeline) code=/tmp/h2o-stream-pipeline-recovery/scripts/convert_h2o_v1.0.py; extra=(--view-workers 16 --camera-pipeline --stream-media --ffv1-slices 16 --ram-stage-dir /dev/shm/xuzhefeng-h2o-pipeline-overlap-20260930 --ram-stage-gib 32);;
+ process) code=/tmp/h2o-process-pipeline-recovery/scripts/convert_h2o_v1.0.py; extra=(--view-workers 16 --view-backend processes --camera-pipeline --stream-media --ffv1-slices 16 --ram-stage-dir /dev/shm/xuzhefeng-h2o-pipeline-process-20260930 --ram-stage-gib 32);;
+ v16) code=/tmp/h2o-stream-v8-recovery/scripts/convert_h2o_v1.0.py; extra=(--view-workers 16 --stream-media);;
+ *) exit 2;;
+esac
+root=/data/xuzhefeng/Datasets/uni-hoi-h2o-pipeline-bench-20260930
+caseout="$root/$1"
+mkdir -p "$caseout/scratch"
+export TMPDIR="$caseout/scratch" PYTHONUNBUFFERED=1
+cg=$(awk -F: '$1=="0" {print $3}' /proc/self/cgroup)
+cat "/sys/fs/cgroup$cg/cpu.stat" > "$caseout/cpu.start"
+date -Is > "$caseout/start"
+.venv/bin/python "$code" --root "$root/raw" --out "$caseout" --workers 4 --codec-threads 10 --decode-threads 10 --depth-threads 4 --max-frames 200 --mano-dir /home/xuzhefeng/Projects/WiLoR/mano_data "${extra[@]}" > "$caseout/convert.log" 2>&1
+code=$?
+cat "/sys/fs/cgroup$cg/cpu.stat" > "$caseout/cpu.end"
+printf '%s\n' "$code" > "$caseout/exit"
+exit "$code"

H2O/shared-v1.0/script/provenance/operations/h2o-pipeline-benchmark.py

@@ -0,0 +1,48 @@
+import datetime,json,os,subprocess,time
+from pathlib import Path
+root=Path('/data/xuzhefeng/Datasets/uni-hoi-h2o-pipeline-bench-20260930')
+cpus=os.cpu_count()
+records=[]
+for mode in ('old','v8'):
+    unit=f'h2o-pipeline-{mode}-20260930.scope'
+    begin=time.monotonic()
+    proc=subprocess.Popen(['systemd-run','--user','--scope','--unit='+unit,'-p','CPUAccounting=yes',
+                           '/bin/bash','/tmp/h2o-pipeline-bench-job.sh',mode])
+    path=Path(f'/sys/fs/cgroup/user.slice/user-{os.getuid()}.slice/user@{os.getuid()}.service/app.slice/{unit}')
+    samples=[]
+    while proc.poll() is None:
+        try:
+            stat=dict(line.split() for line in (path/'cpu.stat').read_text().splitlines())
+            samples.append((time.monotonic()-begin,int(stat['usage_usec'])))
+        except (FileNotFoundError,PermissionError):pass
+        time.sleep(0.5)
+    elapsed=time.monotonic()-begin
+    assert proc.returncode==0,(mode,(root/mode/'convert.log').read_text()[-3000:])
+    def usage(name):return int(dict(line.split() for line in (root/mode/name).read_text().splitlines())['usage_usec'])
+    total=usage('cpu.end')-usage('cpu.start')
+    peak=max(((b[1]-a[1])/(b[0]-a[0])/1e6/cpus*100 for a,b in zip(samples,samples[10:])),default=0)
+    row=dict(mode=mode,seconds=elapsed,cpu_seconds=total/1e6,cpu_percent_machine=total/elapsed/1e6/cpus*100,
+             peak_5s_cpu_percent_machine=peak,samples=samples)
+    records.append(row)
+    (root/'performance.json').write_text(json.dumps(records,indent=2)+'\n')
+    print(json.dumps({k:v for k,v in row.items() if k!='samples'}),flush=True)
+# Prove actual media and every numeric tensor match, not just wall-clock speed.
+os.environ['HAND_WAREHOUSE_FFMPEG_THREADS'] = '10'
+import sys
+sys.path.insert(0,'/tmp/h2o-stream-v8-recovery/scripts')
+from warehouse_v1_0 import decoded_digest,validate_view
+import torch
+for view in sorted((root/'old/H2O/shared-v1.0/data').glob('*/cam*')):
+    rel=view.relative_to(root/'old/H2O/shared-v1.0/data')
+    candidate=root/'v8/H2O/shared-v1.0/data'/rel
+    assert validate_view(candidate)['validated']
+    for name,pix in (('rgb','rgb24'),('depth','gray16le'),('depth_valid','gray')):
+        aa=decoded_digest(view/'observation'/f'{name}.mkv',pix)[:2]
+        bb=decoded_digest(candidate/'observation'/f'{name}.mkv',pix)[:2]
+        assert aa==bb,(rel,name)
+    for pt in view.rglob('*.pt'):
+        aa=torch.load(pt,weights_only=True)
+        bb=torch.load(candidate/pt.relative_to(view),weights_only=True)
+        assert aa.shape==bb.shape and aa.dtype==bb.dtype and aa.numpy().tobytes()==bb.numpy().tobytes(),(rel,pt)
+print('PASS all 10 real views: full RGB/depth/valid decoded digests and every numeric tensor match',flush=True)
+(root/'validation.exit').write_text('0\n')

H2O/shared-v1.0/script/provenance/operations/h2o-pipeline-process-benchmark.py

@@ -0,0 +1,48 @@
+import datetime,json,os,subprocess,time
+from pathlib import Path
+root=Path('/data/xuzhefeng/Datasets/uni-hoi-h2o-pipeline-bench-20260930')
+cpus=os.cpu_count()
+records=[]
+for mode in ('process',):
+    unit=f'h2o-pipeline-{mode}-20260930.scope'
+    begin=time.monotonic()
+    proc=subprocess.Popen(['systemd-run','--user','--scope','--unit='+unit,'-p','CPUAccounting=yes',
+                           '/bin/bash','/tmp/h2o-pipeline-bench-job.sh',mode])
+    path=Path(f'/sys/fs/cgroup/user.slice/user-{os.getuid()}.slice/user@{os.getuid()}.service/app.slice/{unit}')
+    samples=[]
+    while proc.poll() is None:
+        try:
+            stat=dict(line.split() for line in (path/'cpu.stat').read_text().splitlines())
+            samples.append((time.monotonic()-begin,int(stat['usage_usec'])))
+        except (FileNotFoundError,PermissionError):pass
+        time.sleep(0.5)
+    elapsed=time.monotonic()-begin
+    assert proc.returncode==0,(mode,(root/mode/'convert.log').read_text()[-3000:])
+    def usage(name):return int(dict(line.split() for line in (root/mode/name).read_text().splitlines())['usage_usec'])
+    total=usage('cpu.end')-usage('cpu.start')
+    peak=max(((b[1]-a[1])/(b[0]-a[0])/1e6/cpus*100 for a,b in zip(samples,samples[10:])),default=0)
+    row=dict(mode=mode,seconds=elapsed,cpu_seconds=total/1e6,cpu_percent_machine=total/elapsed/1e6/cpus*100,
+             peak_5s_cpu_percent_machine=peak,samples=samples)
+    records.append(row)
+    (root/'performance-process.json').write_text(json.dumps(records,indent=2)+'\n')
+    print(json.dumps({k:v for k,v in row.items() if k!='samples'}),flush=True)
+# Prove actual media and every numeric tensor match, not just wall-clock speed.
+os.environ['HAND_WAREHOUSE_FFMPEG_THREADS'] = '10'
+import sys
+sys.path.insert(0,'/tmp/h2o-process-pipeline-recovery/scripts')
+from warehouse_v1_0 import decoded_digest,validate_view
+import torch
+for view in sorted((root/'old/H2O/shared-v1.0/data').glob('*/cam*')):
+    rel=view.relative_to(root/'old/H2O/shared-v1.0/data')
+    candidate=root/'process/H2O/shared-v1.0/data'/rel
+    assert validate_view(candidate)['validated']
+    for name,pix in (('rgb','rgb24'),('depth','gray16le'),('depth_valid','gray')):
+        aa=decoded_digest(view/'observation'/f'{name}.mkv',pix)[:2]
+        bb=decoded_digest(candidate/'observation'/f'{name}.mkv',pix)[:2]
+        assert aa==bb,(rel,name)
+    for pt in view.rglob('*.pt'):
+        aa=torch.load(pt,weights_only=True)
+        bb=torch.load(candidate/pt.relative_to(view),weights_only=True)
+        assert aa.shape==bb.shape and aa.dtype==bb.dtype and aa.numpy().tobytes()==bb.numpy().tobytes(),(rel,pt)
+print('PASS all 10 real views: full RGB/depth/valid decoded digests and every numeric tensor match',flush=True)
+(root/'validation-process.exit').write_text('0\n')

H2O/shared-v1.0/script/provenance/operations/h2o-prepare-pipeline-bench.py

@@ -0,0 +1,28 @@
+import json,os,tarfile,shutil
+from pathlib import Path
+root=Path('/data/xuzhefeng/Datasets/uni-hoi-h2o-pipeline-bench-20260930')
+root.mkdir(exist_ok=True)
+raw=root/'raw'; raw.mkdir(exist_ok=True)
+old=Path('/data/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-w4-c10-d10')
+records=[]
+for subj in ('subject3','subject4'):
+    subject=next((old/'scratch').glob('h2o-source-*/'+subj))
+    sequences=[p for p in subject.glob('*/*') if all((p/f'cam{i}/rgb/000199.png').is_file() for i in range(5))]
+    assert sequences, (subj,subject)
+    seq=sequences[0]
+    snapshot=root/'input-snapshot'/subj/seq.relative_to(subject)
+    for src in seq.glob('cam*/*/*'):
+        if src.is_file() and src.parent.name in ('rgb','depth','cam_pose','hand_pose','hand_pose_mano','obj_pose','obj_pose_rt') and int(src.stem)<200:
+            dest=snapshot/src.relative_to(seq);dest.parent.mkdir(parents=True,exist_ok=True)
+            os.link(src,dest)
+    for src in seq.glob('cam*/cam_intrinsics.txt'):
+        dest=snapshot/src.relative_to(seq);dest.parent.mkdir(parents=True,exist_ok=True);os.link(src,dest)
+    archive=raw/f'{subj}_v1_1.tar.gz'
+    with tarfile.open(archive,'w:gz',compresslevel=0) as tf:
+        tf.add(snapshot,arcname=str(snapshot.relative_to(root/'input-snapshot')))
+    records.append({'sequence':str(snapshot.relative_to(root/'input-snapshot')),'frames':200,'views':5,'archive':str(archive),
+                    'origin':str(seq),'gzip_compression_level':0})
+for name in ('object.zip','label_split.zip'):
+    os.link(Path('/data/xuzhefeng/Datasets/Hand-Datasets-Raw/H2O/raw/data')/name,raw/name)
+(root/'input-manifest.json').write_text(json.dumps(records,indent=2)+'\n')
+print('PREPARED',records,flush=True)

H2O/shared-v1.0/script/provenance/operations/h2o-prepare-process-full.py

@@ -0,0 +1,45 @@
+"""Preserve complete persistent views; frozen process pipeline only changes execution/storage work."""
+import datetime,json,os,shutil
+from pathlib import Path
+old=Path('/data/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-w4-c10-d10')
+new=Path('/data/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-process16-pipe')
+new.mkdir()
+shutil.copytree('/tmp/h2o-process-pipeline-recovery',new/'frozen-code')
+release=new/'H2O/shared-v1.0'
+for directory in ('data','doc/validation','script/provenance/previous-run','script/provenance/operations'):
+    (release/directory).mkdir(parents=True,exist_ok=True)
+history=release/'script/provenance/previous-run'
+shutil.copytree(old/'H2O/shared-v1.0/script',history/'script')
+shutil.copytree(old/'frozen-code/docs',history/'docs')
+for name in ('full.log','full.status','run-full.sh'):
+    shutil.copy2(old/name,history/name)
+records=[]
+for record in sorted((old/'H2O/shared-v1.0/doc').glob('*.build.json')):
+    build=json.loads(record.read_text());seq=build['sequence_id']
+    assert Path(seq).name==seq and seq not in ('.','..')
+    assert set(build['views'])=={f'cam{i}' for i in range(5)}
+    folder=old/'H2O/shared-v1.0/data'/seq
+    for cam in build['views']:assert (folder/cam/'metainfo.json').is_file()
+    shutil.copytree(folder,release/'data'/seq,copy_function=os.link)
+    shutil.copy2(record,release/'doc'/record.name)
+    records.append({'sequence_id':seq,'original_script_sha256':build['scripts']})
+manifest={'time':datetime.datetime.now().astimezone().isoformat(),'source_run':str(old),'destination_run':str(new),
+          'reason':'Native bytes unchanged; process/view pipeline, streaming RAW and bounded RAM staging',
+          'target_cpu_percent_machine':80,'cpu_limit':None,'reused_sequences':records,
+          'reuse':'Read-only hardlinks; original build records retained; source frame/component SHA256 checked on resume',
+          'previous_run_provenance':'script/provenance/previous-run'}
+(release/'doc/RESTART_PROVENANCE.json').write_text(json.dumps(manifest,indent=2)+'\n')
+operations=release/'script/provenance/operations'
+for name in ('h2o-prepare-process-full.py','h2o-switch-process-full.sh','h2o-pipeline-benchmark.py',
+             'h2o-pipeline-process-benchmark.py','h2o-pipeline-bench-job.sh','h2o-prepare-pipeline-bench.py',
+             'h2o-process-final-tests.log','h2o-shared-final-tests.log'):
+    shutil.copy2(Path('/tmp')/name,operations/name)
+bench=Path('/data/xuzhefeng/Datasets/uni-hoi-h2o-pipeline-bench-20260930')
+for p in bench.glob('performance*.json'):shutil.copy2(p,release/'doc/validation'/p.name)
+shutil.copy2(bench/'input-manifest.json',release/'doc/validation/PIPELINE_BENCH_INPUTS.json')
+shutil.copy2('/tmp/h2o-pipeline-process-benchmark.log',release/'doc/validation/PIPELINE_BENCHMARK.log')
+shutil.copy2('/tmp/h2o-runtime-profile.log',release/'doc/validation/BEFORE_RUNTIME_PROFILE.log')
+shutil.copy2('/tmp/h2o-current-iostat.log',release/'doc/validation/BEFORE_IOSTAT.log')
+for destination in (new/'run-full.sh',release/'script/run-full.sh'):
+    shutil.copy2('/tmp/h2o-run-process-full.sh',destination)
+print('REUSED_SEQUENCES',len(records),'REUSED_VIEWS',5*len(records),flush=True)

H2O/shared-v1.0/script/provenance/operations/h2o-process-final-tests.log

@@ -0,0 +1,39 @@
+Converted subject1/h1/0
+Converted subject2/h1/0
+Converted subject3/h1/0
+Converted subject4/h1/0
+Converted subject1/h1/0
+Archive subject1_v1_1.tar.gz: 1 sequences
+Converted subject2/h1/0
+Archive subject2_v1_1.tar.gz: 1 sequences
+Converted subject3/h1/0
+Archive subject3_v1_1.tar.gz: 1 sequences
+Converted subject4/h1/0
+Archive subject4_v1_1.tar.gz: 1 sequences
+Converted subject2/h1/0
+Converted subject3/h1/0
+Converted subject1/h1/0
+Archive subject1_v1_1.tar.gz: 1 sequences
+Archive subject2_v1_1.tar.gz: 1 sequences
+Archive subject3_v1_1.tar.gz: 1 sequences
+Converted subject4/h1/0
+Archive subject4_v1_1.tar.gz: 1 sequences
+Verified existing subject1/h1/0
+Verified existing subject3/h1/0
+Archive subject1_v1_1.tar.gz: 1 sequences
+Verified existing subject2/h1/0
+Verified existing subject4/h1/0
+Archive subject2_v1_1.tar.gz: 1 sequences
+Archive subject3_v1_1.tar.gz: 1 sequences
+Archive subject4_v1_1.tar.gz: 1 sequences
+Camera-order fallback: subject1/h1/0
+Converted subject1/h1/0
+Converted subject2/h1/0
+Converted subject1/h1/0
+Archive subject1_v1_1.tar.gz: 1 sequences
+Archive subject2_v1_1.tar.gz: 1 sequences
+Converted subject3/h1/0
+Archive subject3_v1_1.tar.gz: 1 sequences
+Converted subject4/h1/0
+Archive subject4_v1_1.tar.gz: 1 sequences
+PASS H2O pipe RGB/depth, RAM annotations, shared view pool 1/4/8, four archive workers, codec/decode 1/4/10, ordered depth 1/2/4 and missing-depth preservation, no resume PNG writes, RGB FFV1-only, depth and persistent QC

H2O/shared-v1.0/script/provenance/operations/h2o-quick-validation.py

@@ -0,0 +1,161 @@
+"""Full coverage/numeric audit, plus representative persisted-media QC and offline HTML."""
+import os
+for key in ('OMP_NUM_THREADS','OPENBLAS_NUM_THREADS','MKL_NUM_THREADS','NUMEXPR_NUM_THREADS'):
+    os.environ[key]='1'
+os.environ['HAND_WAREHOUSE_FFMPEG_THREADS']='4'
+os.environ.pop('HAND_WAREHOUSE_FFV1_SLICES',None)
+os.environ.pop('HAND_WAREHOUSE_MEDIA_TMPDIR',None)
+import collections,datetime,hashlib,html,json,re,sys,time,zipfile
+from concurrent.futures import ThreadPoolExecutor,as_completed
+from pathlib import Path
+from zoneinfo import ZoneInfo
+import numpy as np
+import torch
+root=Path('/data/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-process16-pipe')
+release=root/'H2O/shared-v1.0'
+sys.path.insert(0,str(root/'frozen-code/scripts'))
+from warehouse_v1_0 import SCHEMA,check_transform,check_valid_policy,check_view_type,safe_member,sha256
+from render_h2o_validation import render
+out=release/'doc/validation/final-quick-check'
+out.mkdir(exist_ok=True)
+torch.set_num_threads(1)
+started=time.monotonic()
+expected={}
+with zipfile.ZipFile('/data/xuzhefeng/Datasets/Hand-Datasets-Raw/H2O/raw/data/label_split.zip') as z:
+    for name in z.namelist():
+        if '/action_' not in name:continue
+        for line in z.read(name).decode().splitlines():
+            fields=line.split()
+            if len(fields)>1 and fields[1]!='path':
+                seq=fields[1].replace('/','_')
+                expected[seq]=max(expected.get(seq,0),int(fields[-1])+1)
+folders={p.name:p for p in (release/'data').iterdir() if p.is_dir()}
+assert set(folders)==set(expected),(sorted(set(expected)-set(folders)),sorted(set(folders)-set(expected)))
+assert len(expected)==184 and sum(expected.values())==114329
+receipts={p.name.removesuffix('.build.json'):p for p in (release/'doc').glob('*.build.json')}
+assert set(receipts)==set(expected)
+assert 'EXIT=0' in (root/'full.status').read_text()
+failures=[]
+rows=[]
+def require(ok,message):
+    if not ok:raise ValueError(message)
+def load(path):
+    def cpu(storage,location):
+        require(location=='cpu',f'non-CPU storage: {path}')
+        return storage
+    obj=torch.load(path,map_location=cpu,weights_only=True)
+    require(type(obj) is torch.Tensor,f'non-tensor: {path}')
+    return obj.numpy()
+def check_sequence(seq):
+    folder=folders[seq];n=expected[seq]
+    require({p.name for p in folder.iterdir() if p.is_dir()}=={f'cam{i}' for i in range(5)},'five-camera coverage')
+    receipt=json.loads(receipts[seq].read_text())
+    require(receipt['sequence_id']==seq and set(receipt['views'])=={f'cam{i}' for i in range(5)},'build receipt')
+    result=[]
+    for i in range(5):
+        cam=f'cam{i}';view=folder/cam
+        meta=json.loads((view/'metainfo.json').read_text())
+        require((meta['dataset_id'],meta['sequence_id'],meta['camera_id'],meta['num_frames'])==('H2O',seq,cam,n),'header/timeline')
+        require(meta['view_type']==('ego' if i==4 else 'exo'),'view role')
+        require(meta['sampling']['selection']=={'max_frames':None,'views':None},'unexpected subsampling')
+        check_view_type(meta)
+        components=meta['components']
+        require(all(components[k].get('available') for k in ('rgb','depth','mano','keypoints_xyz','intrinsic','extrinsic')),'required modality absent')
+        require(not components['segment']['available'] and not components['relations']['available'],'unexpected synthesized component')
+        rgb_sources=[];depth_sources=[];rgb_count=depth_count=0
+        with (view/'camera/frames.jsonl').open() as inp:
+            count=0
+            for index,line in enumerate(inp):
+                f=json.loads(line)
+                require(f['frame_index']==index and f['source_frame_id']==index,'source frame index')
+                require(f['sync_id']==f'{seq.replace("_","/")}:{index}','synchronization ID')
+                sources=f['component_sources']
+                for kind,loc in sources.items():
+                    require(re.fullmatch('[0-9a-f]{64}',loc.get('sha256','')) is not None,f'source SHA256: {kind}')
+                    require(f'/{cam}/' in loc.get('member',''),'wrong source camera')
+                rgb_sources.append('rgb' in sources);depth_sources.append('depth' in sources)
+                rgb_count+=rgb_sources[-1];depth_count+=depth_sources[-1];count+=1
+        require(count==n,'frames.jsonl length')
+        numeric=0
+        object_names=[]
+        for key,c in components.items():
+            if not c['available']:continue
+            for field in ('data_ref','valid_ref','mapping_ref'):
+                if c.get(field):
+                    path=safe_member(view,c[field])
+                    require(path.is_file() and path.stat().st_size>0,f'missing/empty {c[field]}')
+            schema=SCHEMA['components'][c['schema_id']]
+            if c.get('valid_policy'):check_valid_policy(c,n)
+            valid=load(view/c['valid_ref']) if c.get('valid_ref','').endswith('.pt') else None
+            if valid is not None:
+                shape=schema.get('valid_shape')
+                require(valid.dtype==np.bool_ and valid.shape[0]==n,f'{key}: validity dtype/timeline')
+                if shape:require(len(shape)==valid.ndim and all(not isinstance(s,int) or valid.shape[j]==s for j,s in enumerate(shape)),f'{key}: validity shape')
+                if key=='rgb':require(np.array_equal(valid,np.asarray(rgb_sources)), 'RGB existence validity mismatch')
+            if c['data_ref'].endswith('.pt'):
+                arr=load(view/c['data_ref']);numeric+=1
+                require(arr.dtype==np.float32 and arr.shape[0]==n,f'{key}: dtype/timeline')
+                shape=schema.get('shape')
+                if shape:require(len(shape)==arr.ndim and all(not isinstance(s,int) or arr.shape[j]==s for j,s in enumerate(shape)),f'{key}: tensor shape')
+                if valid is not None:
+                    require(arr.shape[:valid.ndim]==valid.shape and np.isfinite(arr[valid]).all() and np.isnan(arr[~valid]).all(),f'{key}: finite/NaN validity semantics')
+                    if c['schema_id'] in ('transform/se3/v1','object/rigid/v1'):check_transform(arr[valid])
+                if key=='intrinsic':require(np.allclose(arr[:,2],[0,0,1]) and (arr[:,0,0]>0).all() and (arr[:,1,1]>0).all(),'camera intrinsics')
+            if key.startswith('mesh_'):object_names.append(key.removeprefix('mesh_'))
+        proof=receipt['views'][cam]
+        require(not proof['unresolved'],'unresolved QC components')
+        media={}
+        for p in proof['validation']:
+            if p.get('profile')==components['rgb']['storage']['profile'] and p.get('all_frames_exact') is True:media['rgb']=p
+            elif p.get('component') in ('depth','depth_valid'):media[p['component']]=p
+        require(set(media)=={'rgb','depth','depth_valid'},'media roundtrip proof missing')
+        for key,p in media.items():
+            require(p.get('all_frames_exact') is True and re.fullmatch('[0-9a-f]{64}',p.get('decoded_sha256','')) is not None,f'{key}: failed original roundtrip proof')
+            require((view/'observation'/f'{key}.mkv').stat().st_size==p['bytes'],f'{key}: encoded file size changed since conversion')
+        result.append({'sequence_id':seq,'camera':cam,'frames':n,'numeric_components':numeric,
+                       'rgb_source_frames':rgb_count,'depth_source_frames':depth_count,'objects':sorted(object_names)})
+    return result
+with ThreadPoolExecutor(max_workers=4) as pool:
+    jobs={pool.submit(check_sequence,seq):seq for seq in sorted(expected)}
+    for job in as_completed(jobs):
+        seq=jobs[job]
+        try:rows.extend(job.result())
+        except Exception as exc:failures.append({'sequence':seq,'error':repr(exc)})
+        if len(rows)%100==0:print('STRUCTURAL_VIEWS',len(rows),'FAILURES',len(failures),flush=True)
+print('STRUCTURAL_COMPLETE',len(rows),'FAILURES',failures,flush=True)
+# Balance subjects while covering all eight source object classes; include matching exo views.
+objects=('book','espresso','lotion','spray','milk','cocoa','chips','cappuccino')
+selected=[]
+for j,name in enumerate(objects):
+    subject=f'subject{j%4+1}_'
+    candidates=[r for r in rows if r['camera']=='cam4' and name in r['objects']]
+    require(candidates,f'no output for object {name}')
+    own=[r for r in candidates if r['sequence_id'].startswith(subject)]
+    r=min(own or candidates,key=lambda x:(x['frames'],x['sequence_id']))
+    selected.append((r['sequence_id'],'cam4',name))
+    selected.append((r['sequence_id'],'cam0',name))
+samples=[]
+def full_sample(item):
+    seq,cam,name=item
+    view=folders[seq]/cam
+    destination=out/f'{seq}_{cam}'
+    page=render(view,destination) # Reuse persisted contract QC, depth semantics, frame snapshots, native mesh preview.
+    return {'sequence_id':seq,'camera':cam,'object':name,'frames':expected[seq],
+            'passed':True,'html':str(page.relative_to(out))}
+with ThreadPoolExecutor(max_workers=4) as pool:
+    jobs={pool.submit(full_sample,item):item for item in selected}
+    for job in as_completed(jobs):
+        item=jobs[job]
+        try:samples.append(job.result());print('SAMPLE_PASS',item,flush=True)
+        except Exception as exc:failures.append({'sample':item,'error':repr(exc)});print('SAMPLE_FAIL',item,repr(exc),flush=True)
+report={'time':datetime.datetime.now(ZoneInfo('Asia/Shanghai')).isoformat(),'elapsed_seconds':time.monotonic()-started,
+        'release':str(release),'passed':not failures,'scope':'All sequence/camera coverage, frame/source manifests, all numeric tensors and original media receipts; 16 representative full persisted-view QC checks and offline previews. Not a new all-video source-pixel comparison.',
+        'expected_sequences':len(expected),'expected_sync_frames':sum(expected.values()),
+        'checked_sequences':len(rows)//5,'checked_views':len(rows),'checked_view_frames':sum(r['frames'] for r in rows),
+        'rgb_source_frames':sum(r['rgb_source_frames'] for r in rows),'depth_source_frames':sum(r['depth_source_frames'] for r in rows),
+        'numeric_components':sum(r['numeric_components'] for r in rows),'samples':samples,'failures':failures,
+        'next_step':'Human inspection of representative offline HTML; automatic QC is not proof of source geometric accuracy.'}
+(out/'QUICK_VALIDATION.json').write_text(json.dumps(report,ensure_ascii=False,indent=2)+'\n')
+(out/'index.html').write_text('<!doctype html><html lang="zh"><meta charset="utf-8"><title>H2O 全量后抽样验收</title><h1>H2O 全量后抽样验收</h1><p>自动检查通过不等于源几何标注正确；以下为四个 subject、八类物体、ego/exo 的源数据预览。</p><ul>'+''.join('<li><a href="'+html.escape(x['html'],quote=True)+'">'+html.escape(x['sequence_id']+' '+x['camera']+' '+x['object'])+'</a></li>' for x in sorted(samples,key=lambda x:(x['object'],x['camera'])))+'</ul><pre>'+html.escape(json.dumps({k:v for k,v in report.items() if k not in ('samples',)},ensure_ascii=False,indent=2))+'</pre></html>')
+print('FINAL_REPORT',json.dumps({k:v for k,v in report.items() if k!='samples'},ensure_ascii=False),flush=True)
+raise SystemExit(0 if report['passed'] else 1)

H2O/shared-v1.0/script/provenance/operations/h2o-shared-final-tests.log

@@ -0,0 +1 @@
+PASS: RGB full/prefix hashes, source depth frames, depth/valid/segment bits and missing-frame rejection

H2O/shared-v1.0/script/provenance/operations/h2o-switch-process-full.sh

@@ -0,0 +1,12 @@
+#!/usr/bin/env bash
+set -euo pipefail
+cd /home/xuzhefeng/Projects/uni-hoi-dataset
+test "$(cat /tmp/h2o-process-final-tests.exit)" = 0
+test "$(cat /tmp/h2o-shared-final-tests.exit)" = 0
+test "$(cat /tmp/h2o-pipeline-process-benchmark.exit)" = 0
+test ! -e /data/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-process16-pipe
+old=/data/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-w4-c10-d10
+printf '\nRECONFIGURE=%s REASON=process16_camera_pipeline_ram32_streaming\n' "$(date -Is)" >> "$old/full.status"
+systemctl --user stop h2o-full-d10-20260930.scope
+.venv/bin/python /tmp/h2o-prepare-process-full.py
+systemd-run --user --scope --unit=h2o-full-process16-20260930 -p CPUAccounting=yes /bin/bash /data/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-process16-pipe/run-full.sh

H2O/shared-v1.0/script/provenance/operations/watch-steady-cpu.py

@@ -0,0 +1,38 @@
+"""Measure the whole process tree only after conversion FFmpeg is active; no CPU-fill work."""
+import datetime,json,os,time
+from pathlib import Path
+unit='h2o-full-process16-20260930.scope'
+cg=Path(f'/sys/fs/cgroup/user.slice/user-{os.getuid()}.slice/user@{os.getuid()}.service/app.slice/{unit}')
+root=Path('/data/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-process16-pipe')
+report=root/'H2O/shared-v1.0/doc/validation/STEADY_CPU_5MIN.json'
+start_wait=time.monotonic()
+while time.monotonic()-start_wait<7200:
+    if not (cg/'cgroup.procs').exists():
+        print('Scope exited before a steady conversion sample',flush=True);raise SystemExit(1)
+    ffmpeg=False
+    for pid in (cg/'cgroup.procs').read_text().split():
+        try:ffmpeg |= Path('/proc',pid,'comm').read_text().strip()=='ffmpeg'
+        except FileNotFoundError:pass
+    if ffmpeg:break
+    time.sleep(5)
+else:
+    print('No conversion FFmpeg within two hours; no steady CPU claim made',flush=True);raise SystemExit(2)
+def sample():
+    values=dict(line.split() for line in (cg/'cpu.stat').read_text().splitlines())
+    return time.monotonic(),int(values['usage_usec'])
+begin=datetime.datetime.now().astimezone().isoformat()
+samples=[sample()]
+for _ in range(60):
+    time.sleep(5)
+    try:samples.append(sample())
+    except FileNotFoundError:
+        print('Scope exited during sample; inspect full.status',flush=True);raise SystemExit(1)
+cpus=os.cpu_count()
+intervals=[(b[1]-a[1])/(b[0]-a[0])/1e6/cpus*100 for a,b in zip(samples,samples[1:])]
+result={'start':begin,'end':datetime.datetime.now().astimezone().isoformat(),'logical_cpus':cpus,
+        'whole_scope_cpu_percent_machine':(samples[-1][1]-samples[0][1])/(samples[-1][0]-samples[0][0])/1e6/cpus*100,
+        'highest_5sec_cpu_percent_machine':max(intervals),'lowest_5sec_cpu_percent_machine':min(intervals),
+        'target_percent':80,'intervals_percent_machine':intervals,'samples':samples,
+        'scope':unit,'includes':'All scope threads and child processes, including children already exited'}
+report.write_text(json.dumps(result,indent=2)+'\n')
+print(json.dumps({k:v for k,v in result.items() if k not in ('samples','intervals_percent_machine')}),flush=True)

H2O/shared-v1.0/script/provenance/previous-run/docs/CONTRACT.md

@@ -0,0 +1,457 @@
+# 手物数据仓库统一约定 v1.0
+
+格式版本：`hand-warehouse/1.0`。存储约定：`lossless-video-2026-09-18`。
+
+范围限于手物交互重建。按本任务纳入的源模态转换，不要求每个数据集填满所有组件。缺失模态、尚未下载、语义未核实分别报告，不能互相替代。
+
+本文定义各字段的含义与使用规则，机器可读注册表见 [schema.json](schema.json)。两者互相补充；如有冲突，应先核实并统一定义。
+
+## 1. 目录
+
+```text
+hand-datasets/
+  <dataset_name>/
+    raw/
+      doc/                         # 官方说明、来源与校验记录
+      script/                      # 官方下载脚本、必要包装脚本
+      data/                        # 保留原始包，不原地解压或改写
+    <dataset_version>/
+      doc/                         # 发布说明、schema、来源清单
+      script/                      # 转换、校验脚本及运行配置
+      data/
+        <sequence_id>/
+          <camera_id>/
+            metainfo.json
+            observation/
+              rgb.mkv
+              rgb_valid.pt
+              depth.mkv           # float32 深度的可逆视频编码
+              depth_valid.mkv     # 逐像素 bool 的无损视频编码
+            camera/
+              frames.jsonl
+              intrinsic.pt
+              intrinsic_valid.pt
+              extrinsic.pt
+              extrinsic_valid.pt
+            annotation/
+              mano/
+                motion.pt
+                motion_valid.pt
+                visibility.pt
+              keypoints/
+                uv.pt
+                uv_valid.pt
+                xyz.pt
+                xyz_valid.pt
+                visibility.pt    # 源数据有时提供
+              object/
+                <object_id>/
+                  assets/          # 该物体的网格、材质、纹理及部件资产
+                    <mesh_id>.obj
+                  pose.pt
+                  pose_valid.pt
+                  visibility.pt
+                  metainfo.json    # 物体身份、类别、mesh、尺寸与部件引用
+                  articulation.json
+                  keypoints/
+                  vertex_part_ids.pt
+                  joint_state.pt
+                  joint_state_valid.pt
+                  parts/<part_id>/pose.pt
+                  parts/<part_id>/pose_valid.pt
+              segment/
+                mask.mkv
+                mask_valid.pt
+                instances.jsonl
+              contact/
+                relations.jsonl
+                <subject_id>__<object_id>/  # 对应实体间的距离/接触区域
+```
+
+目录展示可能存在的组件，不要求填满。具体文件、字段和缺省规则分别见后续对应小节。原始包的解压发生在临时工作区，raw 内原始包保持不变。
+
+## 2. 相机视图的 metainfo.json
+
+### 2.1 顶层字段与发布信息
+
+每个相机视图必须有 `metainfo.json`。以下字段除另行说明外均为必填：
+
+| 字段 | 填写内容 |
+|---|---|
+| schema_version | 结构版本，本版为 `hand-warehouse/1.0` |
+| dataset_id | 数据集的逻辑标识 |
+| release_id | 本次数据发布版本，对应目录中的 dataset_version |
+| sequence_id | 所属序列或独立图片样本的标识 |
+| camera_id | 当前相机视图标识 |
+| sample_type | `image` 或 `video`，规则见 2.2 |
+| num_frames | 当前视图帧数 T，所有随帧组件按同一时间轴组织 |
+| hand_slots | 固定为 `["left", "right"]`；解剖学左右的使用规则见 5.1 |
+| sampling | 采样类型、源 FPS、帧索引引用及选择策略，见 2.2 |
+| camera | 当前相机模型与图像尺寸，见 2.3 |
+| components | 实际提供的组件及其文件、schema、有效性和来源，见 2.5 |
+| split | `train/val/test/unknown`，划分协议见 2.4 |
+| view_type | 可选，`ego/exo/unknown` |
+
+`dataset_version` 是数据发布版本；`schema_version` 是结构版本，两者独立。原始包校验摘要、官方来源和转换配置应记录在发布清单中。解压发生在临时工作区，原始包保持不变。
+
+### 2.2 图片、视频与 sampling
+
+`sample_type=image`：允许单图 T=1、`sampling.type=single`，也允许将独立图片批量装入一个视频容器，T>1 时 `sampling.type=independent`、`source_fps=null`。分片大小按数据集选择；FreiHAND 默认每片最多 4096 张，训练与评估分开，尾片不补帧。该数量不是其他数据集的强制要求；同批分辨率、相机模型类型及划分协议一致。每行 frames 必须有稳定且唯一的 `sample_id`、原始文件定位；内参、手形状等仍逐帧保留，不能因打包而共享错误的标定或身份。`sequence_id` 在这里仅是批次标识，不代表真实序列；不得按相邻帧推断跟踪、同步或接触延续。源 camera_id、image_variant、split 等随样本变化时放在 frames 的对应字段中。真实采集时间未知则 timestamp_ns=null。合成背景变体不视为真实多相机。
+
+`sample_type=video`：`sampling.type=dense/sparse`；`source_fps` 是真实采集 FPS，未知为 null。非连续取样必须保留源帧号和已知源时间，不可用采样后索引除以 FPS 冒充源时间。可另记采样策略与步长。
+
+shared 可以仅存所选 RGB 帧对应的数据。OakInk2 图像/动捕不同采样率时用 annotation_matches 明确映射；不要求将未采样动捕点硬塞进 T。多视角身份和源同步关系要保留；选择了哪些帧/相机及原因写在 sampling 中。
+
+`sampling.frames_ref` 指向 `camera/frames.jsonl`。帧号、逐帧时间、同步与标注匹配记录的结构见 4.1；MKV 封装时间的约定见 3.1。
+
+### 2.3 camera 模型声明
+
+`camera` 至少包含 `model`、`schema_id`、`image_size:[W,H]`。模型未知使用 `unknown` 和 `null`，不猜成针孔。实际内参、畸变及外参数组放在 `camera/`，各模型的固定结构见第 4 节。
+
+### 2.4 split 与划分协议
+
+保留上游划分，不将同一采集序列的相邻帧随意拆到不同集合。
+
+有官方划分时必填 split_scheme 和 split_source_locator，split_scheme 区分如 Dex-YCB s0/s1、ARCTIC P1/P2 等选定协议；不知道时 split=unknown 并给理由，不凭空分配。每发布可选择一种协议；存在逐帧差异时 frames 中的 split 覆盖视图默认值，并登记其来源。
+
+### 2.5 components 组件登记
+
+组件表的键是本发布内唯一名称，例如 `rgb`、`depth`、`mano`、`keypoints_uv`。每个已提供组件声明以下信息；模态特有参数在第 3–5 节对应小节中定义。
+
+| 字段 | 约定 |
+|---|---|
+| available | 本发布是否提供该组件；缺省也表示未提供 |
+| schema_id | 已登记的具体存储与语义类型，不得临时创造未登记名称 |
+| data_ref | 数据文件路径，相对当前相机视图根目录，使用 POSIX 路径 |
+| axes | 数组各维度的含义 |
+| valid_ref / valid_axes | 有效性文件及其维度；语义由所属组件规定 |
+| storage / valid_storage | 视频数据／视频有效性的存储配置；包含已登记 profile、实际 codec、pixel_format、encoded_size，见 3.1；非视频组件按原数组规则 |
+| provenance | source / format_conversion / derived，以及转换或生成方法 |
+| source_locator | 原文件、压缩包成员和记录／帧的定位信息 |
+
+不存在的组件允许缺省或写作 `{ "available": false, "reason": "..." }`，不用引用不存在的文件。静态网格可声明标量有效性；具体见 5.4。观测的 valid 见第 3 节，标注的 valid 与 visibility 区分见 5.1。
+
+`provenance.kind` 使用 source/format_conversion/derived，`method` 说明转换或生成方法；发布清单补充脚本版本、参数、输入摘要。`source_locator` 定位到原始文件/压缩包成员和记录、帧号；允许引用有完整溯源清单的前一发布，但链条最终应能回到 raw。不得仅写一个无法定位记录的数据集名字。
+
+`.pt` 仅保存 CPU Tensor，不保存 Python 自定义对象；浮点默认 float32，有效性为 bool，浮点无效位置填 NaN。所有随帧数组保留首维 T，单帧也不 squeeze；静态资产按其 schema 单独声明。单位、坐标及固定形状见各组件小节。
+
+机器注册表 `schema.json` 中，`components` 固定解码后的数据类型、形状、单位和语义；`storage_profiles` 固定视频内的编码／还原方式。不得临时创造未登记的 schema_id 或 storage.profile。同一语义允许选择其登记的不同无损存储 profile。
+
+## 3. observation 观测数据
+
+### 3.1 rgb.mkv 与 rgb_valid.pt
+
+RGB、深度和栅格 mask 在 shared 中统一使用 MKV 视频容器，不展开为逐帧图片或图像 tar。MANO、关键点、位姿等数值标注仍用数组；分割的重叠多边形仍用 JSONL，不为了视频化而有损栅格化。单图可为一帧视频，独立图批量封装规则见 2.2。
+
+**无损指共享层不增加信息损失。** 源 JPEG／H.264 等本身可能已是有损编码，原样保留并不恢复已丢失信息；不得重新以有损 JPEG、H.264、H.265、AV1 参数编码。源视频／图片的色彩、位深、分辨率和几何关系不得被静默改变。RGB schema `rgb/uint8/v1` 的统一读取结果是 `[T,H,W,3] uint8`、RGB 通道，有效性仍为 `rgb_valid.pt [T] bool`。超出该 schema 的高位深或 HDR 源必须报告并扩展定义，不能降到 8 位冒充无损。
+
+#### 3.1.1 允许的 RGB 存储 profile
+
+| storage.profile | 编码方式 | 适用规则 |
+|---|---|---|
+| mkv/mjpeg_copy/v1 | 源 JPEG 压缩帧直接封装成 MJPEG/MKV | 不解码重编码；保持每张 JPEG 码流字节；普通 8 位 JPEG 且 MKV/解码器支持时使用 |
+| mkv/source_copy/v1 | 已有视频的 RGB 对应码流直接封装为 MKV | 仅源编码能被 MKV 和选定解码器支持时使用，不伪造帧；需维持正确的呈现顺序与依赖帧 |
+| mkv/h264_rgb_lossless/v1 | libx264rgb 的真正无损模式 | 必须 RGB 路径、CRF=0，并通过全帧 RGB 字节比较；不能替换成默认 libx264 YUV420 路径 |
+| mkv/ffv1_rgb/v1 | FFV1 的 RGB 可逆编码路径 | 作为通用候选；解码回 RGB 后必须逐字节相同，不做色度降采样 |
+
+优先尝试源码流直接封装；不能封装或需要选择稀疏帧时，比较 FFV1 与无损 RGB H.264。对完整待发布片段，只在通过无损验收的候选中选择文件较小者，不假设某个编码器总最小；编码时间和随机读取代价记入构建记录。中间候选不作为 shared 多份常驻。H.264 跨帧压缩允许较长 GOP，必须记录 GOP／关键帧间隔，随机取帧需从前一关键帧正确解码，不能把 seek 返回的近似帧当目标帧。
+
+普通源视频可以保留它已有的有损码流，但不允许新增有损重编码。源是 VRS 等记录容器时，先查明图像流实际编码：已有兼容 JPEG／视频包可直接提取封装，否则解码后使用真正无损编码；不把整个 VRS 改后缀当 MKV。PNG 不要求直接塞进 MKV，可选上表两种无损重编码方式。
+
+#### 3.1.2 视频配置与校验
+
+每个视频组件在 metainfo 中登记 `storage`，至少包含：
+
+| 字段 | 内容 |
+|---|---|
+| profile | 上述登记的存储 profile；语义仍由组件 schema_id 指定 |
+| codec | 文件中实际 codec，例如 mjpeg、h264、ffv1；source_copy 必须写实际 codec |
+| pixel_format | 实际编码像素格式；不得仅凭编码器参数猜测 |
+| encoded_size | 视频物理像素尺寸 `[width,height]`；深度可能不同于逻辑尺寸，见 3.2 |
+| decoder | 解码器名称、版本，以及颜色矩阵／范围／旋转处理等实际像素解释配置；未知的源色彩声明 unknown，不擅自称为 sRGB |
+| encoder_parameters | 重编码时的编码器、版本、无损模式和 GOP 等参数；直接封装时记录 stream_copy |
+| source_lossy | 源是否已有损：true / false / unknown；不是共享层再有损编码的许可 |
+
+无损重新编码的 RGB 必须使用同一固定解码配置，比较**全部帧**的统一 RGB 字节、数量及顺序。MJPEG-copy 不重新编码，生产校验采用 `jpeg-packets-all-rgb-sample16/v1`：从 MKV 逐包取回，对全部源 JPEG 的 SHA256、数量、顺序和图像尺寸进行精确核对；每视图另用同一固定解码配置比较最多 16 帧的源与目标 RGB 字节（含首尾与四分位位置，其余由固定种子抽取）。记录策略、实际索引、种子与结果；不足 16 帧时全查。试点与最终完整序列随机抽查，对选中视图的全部帧复核解码 RGB。历史已做全帧像素检查的证明继续保留。
+
+逐包一致与像素抽查的范围必须分别声明，不能将抽查写成全量像素检查。不能用 PSNR 高、视觉相似、编码器参数称为 lossless 等替代精确检查。同时保留源编码的色彩标记和必要元数据；原图有 EXIF/色彩/方向等元数据时保留 raw 和必要来源信息；不能静默自动旋转图像而不更新像素对应与标定。以上加速策略仅适用于原 JPEG 码流直接封装，不适用于深度、标签或重新编码的 RGB。
+
+所有共享视频的呈现帧顺序与统一 frame_index 一一对应，每个栅格组件恰好 T 帧。MKV 的 `storage_fps` 只服务封装，声明 `timebase_role=index_only`；source_copy 可以保留源 PTS 以避免改坏码流，但真实时间与标注对齐仍以 frames 为准。未知真实时间保留 null。缺失某帧组件时保留索引位置并使用该组件的无效占位／valid 标识，不删帧；不能以重复相邻图像冒充已知观测。
+
+### 3.2 depth.mkv 与 depth_valid.mkv
+
+深度语义仍是 `depth/z_float32/v1`：统一读取结果 `[T,H,W] float32`，米制相机 Z 分量，`depth_type=z`。有效深度有限且大于 0，无效位置为 NaN。`depth_origin=measured/rendered/unknown` 必填，来源不明时 unknown。默认与 RGB 共用像素网格及相机参考系，不处理未注册 RGB-D。射线距离必须依据源定义转换成 z，不能只改声明。
+
+#### 3.2.1 float32 的可逆视频编码
+
+采用 `mkv/depth_float32_bits/v1`：**不是把深度量化成 uint16**，而是将规范化后每个 IEEE-754 float32 的完整 32 位拆成两个 uint16，使用 FFV1 gray16le 保存。视频物理尺寸为 `[2W,H]`：每行前 W 列存原 float32 位串的低 16 位，后 W 列存高 16 位。读取时先合并为 uint32 位串，再按 float32 重新解释，不能数值 cast 成 float32。
+
+```text
+u = reinterpret_uint32(depth_float32)
+encoded[:, :, :W] = uint16(u & 65535)
+encoded[:, :, W:] = uint16(u >> 16)
+u_restored = uint32(low) | (uint32(high) << 16)
+depth_float32 = reinterpret_float32(u_restored)
+```
+
+统一采用 little-endian 原始字节接口。`camera.image_size` 始终是逻辑 RGB 尺寸 `[W,H]`，不能改成编���视频的 2W；`storage.encoded_size=[2W,H]`。包括 NaN 在内的 float32 位串必须全帧完全恢复，不允许改成 float16、整数毫米、归一化灰度或 RGB 伪彩色。这里的无损验收基准是完成既定米制／坐标转换后的标准 float32 数组，不声称消除 float64→float32 等原约定转换本身的舍入。
+
+#### 3.2.2 逐像素有效性
+
+`depth_valid.mkv` 使用 `mkv/bool_ffv1/v1`，FFV1 gray8（FFmpeg pixel_format=`gray`），物理尺寸 `[W,H]`。像素只能为 0 或 1，统一解码为 `[T,H,W] bool`，逐像素含义不变。这样不再把庞大的 bool 数组保存为未压缩 `.pt`；标注的帧级／手级 valid 仍保留原有小数组。
+
+登记示例（W、H 以下取 640、480；decoder/encoder_parameters 的实际版本由转换器填写）：
+
+```json
+{
+  "available": true,
+  "schema_id": "depth/z_float32/v1",
+  "data_ref": "observation/depth.mkv",
+  "axes": ["T", "height", "width"],
+  "valid_ref": "observation/depth_valid.mkv",
+  "valid_axes": ["T", "height", "width"],
+  "unit": "m",
+  "depth_type": "z",
+  "depth_origin": "measured",
+  "invalid_value": "NaN",
+  "storage": {
+    "profile": "mkv/depth_float32_bits/v1",
+    "codec": "ffv1",
+    "pixel_format": "gray16le",
+    "encoded_size": [1280, 480],
+    "source_lossy": false,
+    "decoder": {"name": "ffmpeg", "version": "填写实际版本"},
+    "encoder_parameters": {"name": "ffv1", "level": 3, "version": "填写实际版本"}
+  },
+  "valid_storage": {
+    "profile": "mkv/bool_ffv1/v1",
+    "codec": "ffv1",
+    "pixel_format": "gray",
+    "encoded_size": [640, 480],
+    "source_lossy": false,
+    "decoder": {"name": "ffmpeg", "version": "填写实际版本"},
+    "encoder_parameters": {"name": "ffv1", "version": "填写实际版本"}
+  },
+  "provenance": {"kind": "format_conversion", "method": "源深度转米制 z float32；逐位打包为 FFV1"},
+  "source_locator": {"file": "raw/data/depth.zip", "member": "depth/0001.png"}
+}
+```
+
+深度和有效性视频都必须逐帧解码验收：float32 位串相同、bool 相同、T 相同，并检查 valid=true 的正有限值和 valid=false 的 NaN。读取器必须按 profile 还原，不能把 depth.mkv 的编码像素直接当深度值。
+
+## 4. camera 相机、帧索引与标定
+
+### 4.1 frames.jsonl：索引、时间与对应关系
+
+`camera/frames.jsonl` 每行对应统一帧号 `frame_index=0..T-1`，是所有组件的唯一行索引。至少包含 `frame_index`、`timestamp_ns`、`clock_domain`、`source_locator`；原始帧号/文件名保存在 `source_frame_id`，不能被新帧号替代。每行还可有 `annotation_matches`，按组件键保存 `source_frame_id`、`source_timestamp_ns`、`clock_domain`、`match_method` 和 `delta_ns`；方法为 exact/index_map/nearest/interpolated，索引映射无时间依据时 delta_ns=null。nearest 必须记录允许误差；interpolated 必须记录两端源标识和权重。delta_ns 仅在同一时间域中计算为源标注时间减图像时间。多相机共享 `sequence_id`；已知同步关系必须用 `sync_id` 或明确的源帧/时间映射表达，不能仅凭相同数组索引认定同步。
+
+每个视图必须提供此文件。源时间使用整数纳秒及 `clock_domain`，未知写 JSON `null`；不同时间域不能直接比较。逐帧 split 可以覆盖 metainfo 的视图默认划分，并保留来源，见 2.4。
+
+### 4.2 坐标系与标定通则
+
+| 项目 | 统一约定 |
+|---|---|
+| 相机坐标 | 右手系，X 向右、Y 向下、Z 向前，长度单位米 |
+| 齐次变换 | 列向量，`p_target = T_target_from_source @ p_source` |
+| 图像坐标 | `(u,v)`；左上像素中心为 `(0,0)`，u 向右、v 向下，单位 pixel |
+| 固定标定 | 本版展开为 T 份，同一 schema 不混用静态与动态形状 |
+| 分辨率或模型变化 | 分开存储为视图／片段，不在同一组件中混装 |
+| 未知标定 | 相机内外参允许缺省，不填单位阵冒充已知标定 |
+
+### 4.3 intrinsic.pt 与畸变参数
+
+| model / schema_id | intrinsic.pt 的固定结构 |
+|---|---|
+| pinhole / camera/pinhole/v1 | `[T,3,3]`，标准 K，最后一行为 `[0,0,1]`；限无畸变或已校正图像 |
+| fisheye624 / camera/fisheye624/v1 | `[T,15]`，`[f,cx,cy,k0,k1,k2,k3,k4,k5,p0,p1,s0,s1,s2,s3]`，采用 Project Aria Fisheye624 投影定义 |
+| opencv_fisheye4 / camera/opencv_fisheye4/v1 | `intrinsic.pt [T,3,3]` 保存 K；同组件 `distortion_ref` 指向 `[T,4]`，顺序 `[k1,k2,k3,k4]`，使用 OpenCV fisheye 等距角度多项式；已由 ReInterHand 实际原始相机参数验证 |
+| rational8 / camera/rational8/v1 | `intrinsic.pt [T,3,3]` 保存 K；同组件必需 `distortion_ref` 指向 `[T,8]`，顺序 `[k1,k2,p1,p2,k3,k4,k5,k6]` |
+| unknown / null | 不提供可用 intrinsic，不声称可以投影 |
+
+有畸变的其他模型必须注册新 schema，固定参数名、顺序和投影公式；不能塞进针孔 K 后遗失畸变项。`metainfo.camera.schema_id` 必须与 intrinsic 组件一致。同一模型的别名不可隐式改变参数布局。
+
+`opencv_fisheye4`：对前方点令 `x=X/Z, y=Y/Z, r=sqrt(x²+y²), theta=atan(r)`，`theta_d=theta*(1+k1*theta²+k2*theta⁴+k3*theta⁶+k4*theta⁸)`；`u=fx*x*theta_d/r+cx, v=fy*y*theta_d/r+cy`。光轴 r=0 时使用连续极限 `theta_d/r=1`。四系数不能按普通径向畸变解释，也不能补零冒充 Fisheye624。
+
+rational8 对归一化坐标 x=X/Z、y=Y/Z，令 r²=x²+y²，径向项为 `(1+k1*r²+k2*r⁴+k3*r⁶)/(1+k4*r²+k5*r⁴+k6*r⁶)`；再加 p1/p2 切向项并经 K 得到像素。`distortion_ref` 与 intrinsic 共用 `[T]` valid，保存到 camera/distortion.pt。不得把这种畸变三维坐标当作真实相机系 xyz 保存。
+
+内参有效性为 `intrinsic_valid.pt [T] bool`；有畸变时使用对应 `distortion_ref`。同一模型的参数名和顺序固定。
+
+### 4.4 extrinsic.pt 与 extrinsic_valid.pt
+
+`extrinsic.pt` 为 `[T,4,4]`，schema `transform/se3/v1`，方向固定 `T_camera_from_world`，有效性 `[T]`。世界坐标含义及源坐标转换记录在 provenance 中；只有相机坐标标注不要求虚构世界坐标。旋转块须正交且 determinant≈1。
+
+## 5. annotation 标注数据
+
+### 5.1 标注通则：手槽、valid 与 visibility
+
+手槽固定为 0=left、1=right，指解剖学左右，不按画面位置或可见性重排。本阶段只处理左右手两槽，不处理更多手。实体引用使用 `hand_left`、`hand_right`；手槽不证明跨采集片段的个人身份。
+
+#### 5.1.1 valid：标注是否可用于监督
+
+标注的 `valid=true` 表示：有对应实际标注、通过本 schema 数值检查，且通过来源明确提供并适用于该组件的有效性/QA 条件。`valid=false` 可由缺标注、无法建立帧对应、源无效占位或官方 QA 不通过导致；不等于画面没有手或手不可见。各组件独立判断，MANO 按 `[T,2]` 标记左右手，关键点按 `[T,2,K]` 标记。
+
+**2026-09-20 将标注 valid 从“仅表示存在”修订为“可用于监督”，字段形状不变。** 公式为 `source_annotation_present AND numerical_checks_pass AND applicable_official_QA_pass`。没有官方 QA 时省略最后一项，不伪造 QA；不能把未知当成已通过已存在的 QA。适配器在组件 `valid_policy` 中记录具体规则、来源、聚合/广播作用域。缺失官方必需 QA 的记录须明确处理为未知或暂停验收，不能静默通过。无效浮点项仍填 NaN，原始数据保留在 raw；被排除原值可用于 doc 中明确标识的诊断渲染，不能冒充 shared 有效监督。
+
+本地筛选须依据明确的规则，并在 `valid_policy.local_excluded_frame_indices`、排除原因和正式报告中记录。单个可疑投影作为问题记录，不直接硬编码排除该帧。已确认的曝光异常须明确作用范围，不能把一个序列的“前两帧有问题”套到所有序列。RGB 编码完好与几何监督可用是两件事，允许保留原图及帧索引而屏蔽对应标注；未发现问题不等于完成全部人工审核，试点待确认时不能作为已验收发布使用。
+
+源字段必须按官方语义映射，不能直接按名字复制：
+
+| 来源 | 官方含义 | 映射原则 |
+|---|---|---|
+| [H2O](https://github.com/taeinkwon/h2odataset) 的手标注首位 | 0 未标注、1 已标注 | 可作为对应手标注存在性的依据，不推断未标注原因 |
+| [HOT3D](https://github.com/facebookresearch/hot3d#using-mask-files) 的 pose_available、visible、mask_qa_pass | 分别表达姿态存在、可见性、对画面中可见手物位姿的人工 QA | 逐手/物体记录决定存在性，再 AND 当前相机当前帧 QA；QA 是整帧聚合判定，保守排除该帧全部手物几何监督，不声称每个实体均有错误；源可见性基于位姿推导时，其自身 valid 同样受 QA 约束 |
+| [Dex-YCB](https://github.com/NVlabs/dex-ycb-toolkit#loading-dataset-and-visualizing-samples) 的全零 pose_m、全 -1 joints | 不可见的手或缺标注共用占位表示 | 对应组件 valid=false；不能从该值判定 visibility=false，也不能声称原因是动捕缺失 |
+| [ARCTIC](https://github.com/zc-alexfan/arctic/blob/master/docs/data/data_doc.md) 的 is_valid/left_valid/right_valid | 是否有效用于官方评估 | 核实每项筛选依据；不把可见性/评估选择直接等同 QA。本次原始 mano.npy 不含这些派生评估标记，按原记录与数值检查生成 valid，并声明没有额外官方 QA 输入 |
+
+源 valid=false 若代表评估筛选或其他条件，应检查实际标注和源约定后映射；不能一律清空坐标。反之，占位零或 -1 只有在源官方明确规定时才判为缺失，不泛化成所有数据的判断规则。
+
+`valid=true` 的标注应通过其 schema 的数值与形状检查；检查失败必须报告或修复，不能静默隐藏。QA 不通过应报告排除数量与原因，并保留可追溯信息；不能把这类无效称为缺标注。RGB、深度等观测有效性仍按各自 schema；手物 QA 不自动否定 RGB、静态相机内参或物体资产。相机外参是否过滤须依据其自身来源规则；HOT3D 本次保留源外参用于溯源和诊断，不能用其存在性证明 QA 失败帧的几何监督正确。
+
+#### 5.1.2 visibility：可见性属性
+
+附加质量属性只保留 `visibility`，存放在对应模态或标注集内，不设 quality 目录。同一份可见性仅存一次；uv 与 xyz 共用时引用同一个文件。
+
+visibility 通过所属组件 attributes 登记 data_ref、axes、dtype、visibility_type、meaning、provenance、source_locator；未知值通过属性 valid_ref 标识，全已知可声明 valid=true。比例与布尔值不相互强转，也不按任意阈值二值化。没有 MANO 时，整手可见性挂在 keypoints 组件，使用独立的 whole_hand_visibility.pt，明确 meaning=whole_hand，不与逐关键点可见性混用。
+
+`valid` 回答“这项标注能否用于监督”，`visibility` 回答“在当前图像中是否可见（或可见多少）”。不可见的手或关键点仍可有有效标注；不能因 visibility=0 就清空坐标。源没有 visibility 时缺省或标为未知，不由 valid 推断。
+
+| 情况 | 手标注 valid | visibility |
+|---|---|---|
+| RGB 已开始录制，动捕尚未开始，没有其他对应手标注 | false | 以独立源可见性标注为准；没有则未知 |
+| 手被遮挡，但源有标注且通过适用有效性检查 | true | 若源声明不可见，则 false 或源可见比例 |
+| 手可见且有对应标注，通过适用 QA | true | 若源有此信息，则 true 或源可见比例 |
+| 手可见，但该帧漏标或动捕中断 | false | 可以为 true，但必须有独立来源依据 |
+| 有手部参数，但官方 QA 明确不通过 | false | 不据此断言不可见；源可见比例若基于不可靠位姿，其属性 valid 也为 false |
+
+visibility 自己的 valid_ref 表示该可见性标注是否可用，不复用 MANO 或关键点坐标的 valid_ref；明确记录源存在性及适用的 QA 条件。上述例子不要求凭图像自动生成可见性。
+
+不设 quality 组件或目录，不另存拟合误差、曝光、截断比例等通用质量数组。适用的官方 QA 合并进对应 valid，在 `valid_policy` 记录来源与规则；正式校验报告保留原始 QA 时间区间及排除统计。原始 QA 文件仍在 raw，官方 split 保留。
+
+### 5.2 mano：参数与整手可见性
+
+MANO schema `mano/axis_angle61/v1`：`motion.pt [T,2,61]`，顺序为完整 16 关节 axis-angle 48、shape betas 10、全局平移 3。姿态使用弧度，平移米，位于当前相机系。有效性 `[T,2]`；缺少一侧整槽 NaN。PCA 必须先按对应模型展开，不能将 PCA 系数直接填入 axis-angle。
+
+`model_profile` 必须固定 MANO 版本、`flat_hand_mean=true`、左手 shapedirs 修正状态。平移表示模型输出顶点所加的全局平移，不是某个腕关节的坐标；源语义不同须转换。shape 无量纲。模型权重不作为共享数据附带上传。
+
+`mano/visibility.pt [T,2]` 保存源整手可见比例时使用 float32 [0,1]，`visibility_type=ratio`；只有实际采用 0–100 的百分比才除以100，不能只凭表头判断。HOT3D 的 `visibility_ratio[%]` 实值已经是0–1，官方读取器原样读取，必须保留原值。通过 MANO 组件 attributes 登记，使用独立有效性，不复用 motion_valid。
+
+手部共享表示仅保留 MANO 参数与选定关键点，不发布独立手表面几何。需要渲染或接触计算时由 MANO 参数生成，几何引用使用固定 MANO 模型配置与顶点顺序；不影响 object 下的物体模型和材质资产。
+
+### 5.3 keypoints：二维点、三维点与逐点可见性
+
+关键点独立于 MANO 参数存在；没有 motion 也可有关键点。`uv.pt [T,2,K,2]` 与 `xyz.pt [T,2,K,3]`，有效性 `[T,2,K]`，分别采用 `keypoints/uv/v1` 和 `keypoints/xyz/v1`。K 由 `keypoint_schema` 固定，不能在同一数组中逐帧变化；同体系统一顺序。数值使用像素坐标或米制相机系。
+
+| keypoint_schema | K 与固定索引 |
+|---|---|
+| mano21/v1 | 21；wrist 后依次 thumb/index/middle/ring/little，每指由近端至指尖 |
+| rhd21/v1 | 21；保持 wrist，然后五指由近端至指尖；原始每指 tip→palm 反序。只统一顺序，不宣称解剖学位置等价于 MANO |
+
+具体名称与源索引映射见 `schema.json`。当前只登记 MANO21 与 RHD21；不登记 MediaPipe、ReInterHand capture21 或 fit24 中间关键点。
+
+每个发布中的每个视图只提供一套标准手关键点，直接放在 keypoints 下，不设 annotation_set_id 字段或目录。uv 与 xyz 是这套关键点的二维与三维表示，不是两套标注。provenance 分别记录二者来自原始标注还是计算结果；不要求已有 2D 必须由当前 xyz 重新投影覆盖。
+
+ReInterHand 的 Original keypoints���Original fits keypoints 及带前臂关键点不纳入 shared；如需手关键点，从选定 MANO 参数生成 MANO21 并记录模型/指尖回归约定。其他数据集默认使用选定的官方关键点，不额外生成平行的 MANO 派生集合。后续主动更换生成方法或结果时发布新的 dataset_version，不在同一版本内增设方法分支。
+
+manual/projection/fitting 只记录在 provenance，不作为目录或类型分支。不得通过坐标为零、是否出画或 visibility 判断标注是否存在。派生标注仅在实际生成并通过数值检查后声明 valid=true；没有生成的项为 false，不伪造坐标。
+
+`keypoints/visibility.pt [T,2,K]` 保存源逐点可见性，布尔表示使用 `visibility_type=boolean`。uv 与 xyz 共用时引用同一份文件。整手比例属性没有 MANO 可挂载时使用 `whole_hand_visibility.pt [T,2]`，`meaning=whole_hand`；它与逐点可见性分开，通用登记规则见 5.1。
+
+不使用 annotation_set_id；ReInterHand 中间关键点集合不进入 shared，raw 保留来源文件。
+
+### 5.4 object：身份、资产、位姿与几何
+
+#### 5.4.1 metainfo.json：物体身份与跟踪作用域
+
+手用 `hand_left`、`hand_right` 及槽位 0、1 表示，不再另存 entities.json。物体身份登记在 `annotation/object/<object_id>/metainfo.json`，必填 `object_id`、`identity_scope`、`scope_id`、`source_locator`；可含 `category_id`、`source_object_id`、`mesh_ref`、`has_model`、部件引用。即使只有分割、没有 pose/mesh，也可只有这份物体元信息。
+
+`identity_scope` 为 frame/subsequence/sequence，`scope_id` 指定具体作用域；跨相机共享身份只在已有源对应证据时使用。VISOR 按官方 subsequence 与实体名关联时记录 `association_basis=official_subsequence_entity_name`，不能声称这是全视频物理实例跟踪。无需与 object_id 一一对应的标准 track_id；官方独立轨迹标识存在时保留为 source_track_id，并声明作用域。手槽不证明跨采集片段的个人身份。
+
+#### 5.4.2 assets：网格、材质与纹理
+
+资产属于物体，统一存放在 `annotation/object/<object_id>/` 内，可用 `assets/` 子目录组织网格、材质、纹理和部件资产。序列层不设共享 assets。所有引用采用视图根目录为基准；其中物体资产组件的 `data_ref` 与部件的 `mesh_ref` 采用相对于当前视图目录的 POSIX 路径，例如 `annotation/object/cup_01/assets/cup.obj`，且必须指向所属物体目录内的资产。OBJ 内部的材质、纹理引用遵循文件格式自身的相对路径规则，依赖文件一并保留。不同相机视图允许各自保存物体资产，不要求跨视图去重；下载物体时包含其资产依赖。
+
+网格 schema 为 `mesh/obj/v1`，顶点米制，静态有效性可直接声明标量 `valid`，或通过 `valid_ref` 指向标量 bool Tensor；不按 T 重复静态资产。由物体／部件通过引用关联，不仅凭文件名猜测。
+
+#### 5.4.3 pose.pt、pose_valid.pt 与 visibility.pt
+
+刚体 schema 为 `object/rigid/v1`，`pose.pt [T,4,4]` 固定为 `T_camera_from_object`，平移米制，有效性 `[T]`。物体坐标原点、轴向必须与 mesh 一致；旋转块为正交矩阵且 determinant≈1。
+
+`object/<id>/visibility.pt [T]` 保存源物体可见比例，float32 [0,1]，`visibility_type=ratio`。通过所属物体组件 attributes 登记，独立表达其有效性。
+
+#### 5.4.4 articulation.json、部件位姿与 joint_state.pt
+
+关节物体：schema `object/articulated/v1`，`articulation.json` 登记 `parts` 与 `joints`。每个 part 有 `part_id`、`parent_id`（根为 null）、`has_model`、`mesh_ref`；has_model=false 时 mesh_ref=null，结构节点允许没有网格。父链不得成环。每个 joint 有唯一 `joint_id`、`parent_part`、`child_part`、`type`；fixed 单位 null、revolute 弧度、prismatic 米。有源限位时保存 `limits:[lower,upper]`；没有时 null，不伪造上下限。部件姿态统一为 `T_camera_from_part [T,4,4]`，不与父局部变换混装。
+
+若提供 `joint_state.pt [T,J]`，必须登记固定 joint_order、轴在父部件系的表达、零位变换与正方向，valid 为 `[T,J]`；缺少这些运动学定义时只发布部件位姿与已有来源信息，不伪造可驱动的关节模型。关节值与部件位姿同时存在时需验证一致性。
+
+部件层级应声明语义：纯语义包含树允许 `joints=[]`，不将包含关系推断为机械关节。关节状态 schema 为 `object/joint_state/v1`；转动用弧度、移动用米，混合类型按 joint_order 对应的关节定义解释。
+
+#### 5.4.5 dimensions、几何关键点与顶点部件标签
+
+`dimensions.pt [T,3]` 采用 `object/dimensions/v1`，有效性为 `[T]`。
+
+物体/部件 `dimensions` 为 `[length,width,height]` 米，轴顺序与该 mesh 的局部坐标关联，登记 `dimensions_semantics` 和来源。HOI4D dimensions 与 mesh 的具体尺度关���未验证前标为 source_unspecified，保留源值并暂停依赖该关系的几何验收，不能默认为待施加缩放。需要施加缩放时明确 scale_applied_to_mesh，禁止重复应用。
+
+源物体关键点放 `object/<id>/keypoints/xyz.pt [T,K,3]`（相机系）或 `template_xyz.pt [K,3]`（物体系），schema 分别为 object_keypoints/camera/v1 和 object_keypoints/template/v1。明确 keypoint_schema、keypoint_names、单位及有效性；模板点配合 pose 可投影，无须重复生成全部帧。H2O 中心/角点/棱中点，以及 ARCTIC 部件点分别登记具体源顺序，不套用手的 21 点 schema。每个发布选择一套源几何点并记录选择，不另建方法分支。
+
+`vertex_part_ids.pt [V] int64` 绑定 `mesh_ref` 与 topology_id，登记整数到 part_id 的映射；未知用 -1。ARCTIC 不同拓扑的资产分别绑定，不能共用未经转换的顶点标签。部件层级中 is_instance/has_model 等源属性保留，几何文件仍全部归 object。
+
+### 5.5 segment：分割类别、编码与覆盖范围
+
+每个发布选择一套任务使用的分割表示，不建立方法／标注集分支。没有分割就不创建 segment 目录，不为其他模态创建空的实体表。
+
+| schema_id | 文件与数值语义 |
+|---|---|
+| segment/uint16_labels/v1 | mask.mkv；整数区域编号，通过 instances.jsonl 解释 |
+| segment/instance_polygons/v1 | polygons.jsonl；实例多边形，可重叠；当前不支持孔洞拓扑 |
+| segment/foreground_intensity/v1 | mask.mkv；0–255 前景强度，不是区域编号或经过校准的概率 |
+
+#### 5.5.1 uint16 标签图与 instances.jsonl
+
+组件声明 `segmentation_type=semantic/instance/part`；同一组件固定一种类型。`mask.mkv` 解码为 `[T,H,W] uint16`，schema `segment/uint16_labels/v1`，使用 `mkv/uint16_ffv1/v1`，FFV1 gray16le 无损；只有所有区域编码均 ≤255 时允许 gray8（pixel_format=gray），读取后转 uint16，数值不变。禁止调色板和有损编码。有效性 `[T]`，表示该帧有可用分割标注。数值 0 固定为 unassigned，65535 固定为 ignore，1–65534 为区域编码，不能直接等同 entity_id。
+
+`instances.jsonl` 每帧一行，包含 `frame_index` 与 `labels`。每条标签包含 `mask_value`、`category_id`、`entity_id`、`part_id`；无意义或未知字段为 null。semantic 允许 entity_id=null；instance 必须引用 hand_left/hand_right 或登记的 object_id；part 还须关联登记部件。源实例 ID 和其他属性分别保存在 `source_instance_id`、`source_attributes`。mask 中每个非保留编码必须有且仅有一个映射。
+
+FreiHAND 的试点按本任务需要，统一到手／背景类别，保留每个原生区域编码；不要求给原生细分区域猜测手指名称。需要细部件监督时，尚未核实的部件名称必须保持未知，不能因粗粒度手类别已核实便声称部件语义也已统一。
+
+#### 5.5.2 重叠实例多边形
+
+每发布选择一套任务使用的分割表示，不建立方法/标注集分支。VISOR 的实际试点存在重叠实例多边形，单层标签图不能无损表达。对此补充 `segment/instance_polygons/v1`：`segment/polygons.jsonl` 每帧包含 `frame_index` 和 `instances`，每实例包含 `mask_value` 与 `polygons`（列表，每个多边形为源图像像素坐标顶点序列，保留原始精度）。身份和类别仍由同一 `instances.jsonl` 映射，有效性仍为 `[T]`，通过 schema 区分物理编码。允许多实例重叠，不采用后画覆盖前画。消费者按需栅格化为独立实例二值 mask；不能声称合并的单层 mask 无损。
+
+#### 5.5.3 灰度前景强度
+
+ReInterHand 的官方渲染 mask 为双手合并的灰度前景掩码，实测存在 0–255 的中间值，不是左右手实例编号。登记 `segment/foreground_intensity/v1`：仍存 `segment/mask.mkv`，解码为 `[T,H,W] uint16`，使用 `mkv/uint16_ffv1/v1`，此类数据优先编码为 FFV1 gray8（pixel_format=gray），读取后转 uint16。有效值范围 0–255，归一化值为 `float32(value)/255`；0 为背景端点、255 为前景端点。组件声明 `foreground_category=hand`、`entity_id=null` 和双手合并语义，不使用区域编号映射。这里保留原始渲染 mask 强度，不把它称为经过校准的概率。若任务需要二值 mask，阈值化属于任务层派生操作，记录阈值；共享层不丢弃中间灰度。不同 schema 的数值含义不同，不能将该文件按整数类别图解码。
+
+FreiHAND 训练集的 JPEG mask 使用 `segment/foreground_intensity/v1`，逐像素保留 JPEG 解码后的灰度，`entity_id=hand_right`。该 schema 不限定双手合并：单手或合并范围由 `entity_id` 与 `coverage` 明确声明。评估集整数区域图仍用 `segment/uint16_labels/v1`，两者不能仅凭文件名使用同一种解码语义。
+
+#### 5.5.4 coverage 与有效性
+
+`coverage.kind=full/selected_entities/partial/unknown`。full 指声明���标注范围完整；selected_entities 还应列出目标范围。无论哪种，0 本身不自动成为背景；背景要有明确类别映射，未标注区域不能用于构造负样本。保留源逐类/逐实例 exhaustive 信息，未知不可改写为完整。
+
+三种表示的帧级有效性均为 `[T]`，表示该帧是否提供此分割标注；对应 valid_ref 在组件中登记，仍用小型 bool `.pt` 数组。栅格 mask 必须登记 storage，格式见 3.1.2，所有帧的整数编码或强度逐值完全相同。编码语义由 schema_id 决定，读取器不能仅依据 mask.mkv 文件名选择解码方式。
+
+### 5.6 contact：接触、抓取、区域与距离
+
+#### 5.6.1 relations.jsonl 与 relations_valid.pt
+
+`annotation/contact/relations.jsonl`，schema contact/relations/v1，每帧一行含 frame_index、relations。每条记录含 relation_type（contact/grasp）、subject_id（手槽或登记实体）、object_id、status、source_locator、provenance。status 固定为 positive/negative/unknown/target_unannotated：positive 必须有可解析目标，negative 表示源明确否定该关系，unknown 为不确定，target_unannotated 为确有接触但目标未标注。非 positive 的 object_id 可 null，不能用 null 冒充“不接触”。稀疏源无记录的帧不自动补 negative。
+
+VISOR 的对象 mask ID 必须先按同帧 source_instance_id 解析成 object_id；hand-not-in-contact、none-of-the-above、inconclusive 分别映射 negative、target_unannotated、unknown，保留原始字符串。Dex ycb_grasp_ind 通过 ycb_ids 映射到登记物体，以 relation_type=grasp 保存；不得当作逐顶点接触。
+
+relations 组件的 valid_ref 指向 `relations_valid.pt [T] bool`，表示该帧是否有可用源关系标注。有效帧中可以存在 status=unknown 的明确不确定标注；无源标注的帧为 valid=false、relations=[]。组件仍声明 available/provenance/source_locator。
+
+#### 5.6.2 实体对的顶点区域与距离
+
+已发布的接触区域/距离按参与关系的 subject_id 与 object_id 组织，每对实体保留选定的一套结果，以精确 geometry_ref/topology_id 关联到手或物体/部件；不要求每套数据都有。顶点区域 schema contact/vertex_region/v1：bool `[T,V]` 加同形 valid；顶点距离 schema contact/vertex_distance/v1：float32 `[T,V]`、米、同形 valid，声明 source/target geometry_ref、distance_definition（signed/unsigned、测量到点集或表面等明确源定义）。geometry_ref 必须指向已登记 mesh 或由固定 MANO 模型配置确定的拓扑；拓扑/顶点数变化时按拓扑拆分片段。其他源区域表示只有在实际接入并登记 schema 后才可转换，不能随意栅格化宣称等价。
+
+顶点区域／距离已有存储定义，但尚未完成真实输入试点；接入时必须提供源语义与几何验收证据，不因已经登记 schema 就视为通过。
+
+## 附录：任务范围
+
+不强制收录 2D bbox、完整高频动捕、动作程序、身体 SMPL-X、点云或网络特征。范围内的接触、visibility、选定手关键点、物体几何不能因不便处理而静默丢弃。分割采用的物理编码由第 5.5 节决定，存在重叠时不能以单层标签图声称无损。
+
+不设置 entities_ref/segment/entities.json，不要求与 object_id 重复的 track_id。MediaPipe 不在当前关键点定义中；旧版本迁移必须保留可追溯关系，不能直接删表。手部共享表示不含独立 hand_mesh。
+
+schema.json 是普通注册表，不是 JSON Schema 文档或可执行验收器；定义存在不代表真实数据已验证。验收须保留实测范围与未决项。
+
+本存储约定只更新标准，不代表既有 FFV1／depth.pt 试点已迁移。新发布必须显式写 storage 配置；旧产物按其旧定义读取，不能只改文件后缀或版本标识伪装为合规。

H2O/shared-v1.0/script/provenance/previous-run/docs/h2o-RAW_TO_SHARED.md

@@ -0,0 +1,48 @@
+# H2O raw → shared-v1.0
+
+## 范围与来源
+
+本版本仅转换 `raw/data/subject{1,2,3,4}_v1_1.tar.gz` 的真实五视角 (`cam0`–`cam4`) 序列；`h2o_CASA.tar` 是单独的扩展，不与标准序列混合。每个原始序列存于 `data/subjectN_<scene>_<sequence>/cam{0..4}/`。源压缩包保持不变，不合成源不存在的标注。输出仅在本地制作；完整序列试点及人工检查通过、更新全量预估之前，不启动全量。
+
+| 源包/字段 | shared 文件/处理 |
+| --- | --- |
+| `subjectN_v1_1.tar.gz`: `camX/rgb/<frame>.png` | `observation/rgb.mkv`：逐帧源 RGB 解码为 RGB24，FFV1 直接编码；输出**每帧**解码与源 RGB 字节摘要完全一致。`rgb_valid.pt` 记录存在性，源帧号和每组件的包成员及 SHA256 写入 `camera/frames.jsonl`。不再额外编码未采用的 H.264 RGB 候选；FFV1 编码器进程失败时才尝试 H.264。 |
+| 同包 `camX/depth/<frame>.png` | `observation/depth.mkv` 和 `depth_valid.mkv`；按源毫米单位除以 1000 转米，保留无效深度与全帧时间轴，深度位串以 FFV1 保真封装并逐帧校验。既不做 3 米裁剪也不造出缺失的深度。 |
+| `camX/cam_intrinsics.txt`、`camX/cam_pose/<frame>.txt` | `camera/intrinsic.pt`、`extrinsic.pt`：保留源内参；逐帧源 `T_world_from_camera` 求逆后存 `T_camera_from_world`；声明变换约定与有效性。 |
+| `camX/hand_pose_mano/<frame>.txt` | `annotation/mano/motion.pt`、`motion_valid.pt`：源左/右手标志各自判断有效性；姿态与平移保持源数值，声明 MANO 模型配置，缺模型则明确标依赖缺失，不伪造网格或掩码。 |
+| `camX/hand_pose/<frame>.txt` | `annotation/keypoints/xyz.pt`、`xyz_valid.pt`：源双手 21 点，相机坐标系米制，与 MANO 有效性彼此独立。HTML 中的投影仅用于人工预览，不写回训练数据。 |
+| `camX/obj_pose/<frame>.txt`、`obj_pose_rt/<frame>.txt` | `annotation/object/<name>/keypoints/xyz.pt`、`pose.pt`：按源类号关联 21 个物体点及物体→相机位姿；仅源存在时登记组件及有效性。 |
+| `object.zip` | `annotation/object/<name>/assets/` 与源 OBJ、MTL、纹理；静态形状预览不等于所有帧网格位姿的投影证明。 |
+| `label_split.zip` 的 `action_{train,val,test}.txt` | 序列元数据的官方划分及来源；未知映射明确记 `unknown`，不自行编排。 |
+
+`cam4` 是 ego，`cam0`–`cam3` 是 exo。真实时间/FPS 缺乏可靠来源时使用 null，不把 MKV 封装帧率冒充采集时钟。源 `action_label`、`verb_label`、pose 任务帧列表和 `manolabel_v1.1.tar.gz` 保留在 raw；当前契约没有相应动作组件，本发布不宣称它们已转换。源没有像素分割，本版本不投影网格生成替代分割。
+
+## 源资产文件名修复
+
+源 `object/cocoa/cocoa.obj` 引用 `ovomaltine.mtl`，材质内写的是 `map_Kd cocoa.png`，但原 ZIP 只提供 `ovomaltine.png`。仅在临时解压目录为该已确认的纹理增加**字节相同**的 `cocoa.png` 别名，保留原 PNG、OBJ、MTL 不变；在网格 `source_locator.texture_aliases` 中记录别名、原始包成员和 SHA256。别名与已保留的旧 cocoa 资产一致，不替换图像、标注或网格。全部八类物体的材质依赖在读取大 subject 包之前预检，其他缺失/越界依赖仍报错。
+
+## 加速与验收
+
+四个互不依赖的 subject 压缩包可由 `--workers 1..4` 并行顺序读取，各 worker 一次只处理一个序列。数值库保持单线程；FFV1 编码支持 `--codec-threads 1|2|4|8|10`，当前全量按用户要求使用 10；`--decode-threads 10` 通过现有 `HAND_WAREHOUSE_FFMPEG_THREADS` 控制源 RGB PNG→RGB24、完整视频解码摘要、深度/有效性语义校验与 RGB 索引抽取的解码线程，完整校验仍保留。深度 PNG 使用 `--depth-threads 4` 有序分批读取，每批最多读取 4 帧，避免缓存整条序列，缺帧仍按原时间轴写 NaN；不并行改写姿态、标注或原始包。并发数由相同输入实测吞吐量决定，记录 OS 线程峰值和磁盘占用，不以 CPU 线程数本身作为提速证据。RGB 源为 PNG，不适用 ARCTIC/HOT3D 的 JPEG 包直封和 16 帧抽查：RGB、深度、有效性与数值模态仍逐帧精确校验。临时文件在 `/data` 上隔离放置，历史 62 条目录不复用、不覆盖；调整并发时，在新的运行目录复用本次全量已经原子发布的完整序列，并从源重新核对帧号及逐组件 SHA256；已发布视图只读源包并核对摘要，不再把源 PNG/标注写入临时目录，待转换视图才正常暂存。未完成的临时序列不复用，原运行的脚本、配置及 QC 保留用于溯源。
+
+先运行不截帧的代表性完整序列试点，保存自动 QC、真实耗时和 `doc/validation/index.html`（RGB、深度、源 3D 关键点投影、原生物体网格和全组件数值/有效性）。试点由人工核查之后再决定全量；全量后再次抽样人工核查。报告须区分数据完整性、源未提供的模态与仅供人工判断的几何对齐。
+
+2026-09-29 完整序列试点已获人工确认，用户已授权本地全量。2026-09-30 用户取消 CPU 使用量上限，指定 4 个独立压缩包 worker、每个 FFV1 编码器 10 线程、每个校验解码器 10 线程、每 worker 深度 PNG 读取 4 线程；Python worker、读取及校验子进程不计入这最多 40 个编码线程，因此总线程数不固定为 40。该配置不代表已证明加速倍数。源包和旧产物保留；代码变化使用独立运行目录，已完成序列保留原始构建记录及生产脚本快照。
+
+
+## 多线程解码测速（2026-09-30）
+
+在 101 上以真实 `subject4/h1/0/cam0` 的完整 RGB、深度和有效性视频，以及源头 100 帧 PNG 测试；采用 1、10、10、1 的解码线程顺序，源深度 PNG 另比较 1/2/4 线程。逐字节摘要和输出长度必须相同，测速脚本、全部轮次及结果保存在 `script/provenance/operations/` 和 `doc/validation/`。测速期间旧全量任务仍运行，磁盘负载和缓存会影响结果；不得将阶段加速比冒充整套转换加速比。四个 gzip 包已经独立并行读取，暂不引入多进程重构或增加数值库线程；小矩阵及标注处理没有已测得的进一步并行收益。
+
+
+本次同输入结果（解码后长度和 SHA256 全部相同）：
+
+| 阶段 | 1 线程（秒） | 新配置（秒） | 范围 |
+| --- | ---: | ---: | --- |
+| FFV1 RGB 解码与摘要 | 39.46 | 8.70（10 线程） | 完整 501 帧，中位数 |
+| FFV1 深度位串解码与摘要 | 34.91 | 11.56（10 线程） | 完整 501 帧，中位数 |
+| 有效性视频解码与摘要 | 2.24 | 2.27（10 线程） | 完整 501 帧；未获提速 |
+| 源 RGB PNG 解码 | 2.67 | 1.49（10 线程） | 100 帧，1 线程使用热缓存轮次，不使用冷缓存较慢轮次来夸大提速 |
+| 源深度 PNG 读取及毫米→米 | 4.67 | 2.61（4 线程） | 100 帧，中位数；帧序与数值相同 |
+
+这些结果只代表相应阶段；全流程还包括顺序 gzip 扫描、文件创建、源 SHA256、资产复制和磁盘读写，实际全量 ETA 待稳定吞吐后再更新。

H2O/shared-v1.0/script/provenance/previous-run/docs/schema.json

@@ -0,0 +1,869 @@
+{
+  "schema_version": "hand-warehouse/1.0",
+  "validity_revision": "usable-annotation-2026-09-20",
+  "annotation_validity": {
+    "meaning": "annotation usable for supervision, not merely record presence",
+    "formula": "source_annotation_present AND numerical_checks_pass AND applicable_official_QA_pass AND NOT explicitly_reviewed_local_exclusion",
+    "without_official_qa": "omit QA term; declare source rules; do not fabricate QA",
+    "policy_field": "components.<name>.valid_policy",
+    "visibility_is_independent": true,
+    "invalid_float_value": "NaN",
+    "hot3d": "broadcast per-camera frame manual QA conservatively to hand/object labels and pose-derived visibility; retain RGB and source calibration; keep native QA in raw and review report",
+    "hot3d_visibility": "native visibility_ratio[%] already ranges from 0 to 1; do not divide by 100"
+  },
+  "hand_slots": [
+    "left",
+    "right"
+  ],
+  "keypoints": {
+    "mano21/v1": {
+      "count": 21,
+      "names": [
+        "wrist",
+        "thumb_1",
+        "thumb_2",
+        "thumb_3",
+        "thumb_4",
+        "index_1",
+        "index_2",
+        "index_3",
+        "index_4",
+        "middle_1",
+        "middle_2",
+        "middle_3",
+        "middle_4",
+        "ring_1",
+        "ring_2",
+        "ring_3",
+        "ring_4",
+        "little_1",
+        "little_2",
+        "little_3",
+        "little_4"
+      ],
+      "source_index_map": null,
+      "note": "adapter maps official MANO joint and fingertip convention"
+    },
+    "rhd21/v1": {
+      "count": 21,
+      "names": [
+        "wrist",
+        "thumb_1",
+        "thumb_2",
+        "thumb_3",
+        "thumb_4",
+        "index_1",
+        "index_2",
+        "index_3",
+        "index_4",
+        "middle_1",
+        "middle_2",
+        "middle_3",
+        "middle_4",
+        "ring_1",
+        "ring_2",
+        "ring_3",
+        "ring_4",
+        "little_1",
+        "little_2",
+        "little_3",
+        "little_4"
+      ],
+      "source_index_map_per_hand": [
+        0,
+        4,
+        3,
+        2,
+        1,
+        8,
+        7,
+        6,
+        5,
+        12,
+        11,
+        10,
+        9,
+        16,
+        15,
+        14,
+        13,
+        20,
+        19,
+        18,
+        17
+      ]
+    }
+  },
+  "camera": {
+    "opencv_fisheye4": {
+      "K": [
+        3,
+        3
+      ],
+      "distortion": [
+        "k1",
+        "k2",
+        "k3",
+        "k4"
+      ],
+      "projection": "theta=atan(r); theta_d=theta*(1+k1*theta^2+k2*theta^4+k3*theta^6+k4*theta^8); source ReInterHand"
+    },
+    "pinhole": {
+      "K": [
+        3,
+        3
+      ]
+    },
+    "fisheye624": {
+      "parameters": [
+        "f",
+        "cx",
+        "cy",
+        "k0",
+        "k1",
+        "k2",
+        "k3",
+        "k4",
+        "k5",
+        "p0",
+        "p1",
+        "s0",
+        "s1",
+        "s2",
+        "s3"
+      ]
+    },
+    "rational8": {
+      "K": [
+        3,
+        3
+      ],
+      "distortion": [
+        "k1",
+        "k2",
+        "p1",
+        "p2",
+        "k3",
+        "k4",
+        "k5",
+        "k6"
+      ]
+    }
+  },
+  "segmentation": {
+    "segment/uint16_labels/v1": {
+      "storage": "mask.mkv",
+      "dtype": "uint16",
+      "mapping": "instances.jsonl",
+      "valid_shape": [
+        "T"
+      ],
+      "required_metadata": [
+        "segmentation_type",
+        "coverage",
+        "mapping_ref",
+        "storage"
+      ],
+      "shape": [
+        "T",
+        "H",
+        "W"
+      ],
+      "reserved_codes": {
+        "0": "unassigned",
+        "65535": "ignore"
+      },
+      "region_codes": [
+        1,
+        65534
+      ],
+      "allowed_storage_profiles": [
+        "mkv/uint16_ffv1/v1"
+      ]
+    },
+    "segment/instance_polygons/v1": {
+      "storage": "polygons.jsonl",
+      "mapping": "instances.jsonl",
+      "frame_keys": [
+        "frame_index",
+        "instances"
+      ],
+      "instance_keys": [
+        "mask_value",
+        "polygons"
+      ],
+      "coordinate_unit": "pixel",
+      "overlap_allowed": true,
+      "valid_shape": [
+        "T"
+      ],
+      "required_metadata": [
+        "segmentation_type",
+        "coverage",
+        "mapping_ref"
+      ],
+      "polygon_shape": [
+        "N",
+        2
+      ],
+      "holes_supported": false
+    },
+    "segment/foreground_intensity/v1": {
+      "storage": "mask.mkv",
+      "dtype": "uint16",
+      "value_range": [
+        0,
+        255
+      ],
+      "normalization_scale": 0.00392156862745098,
+      "axes": [
+        "T",
+        "height",
+        "width"
+      ],
+      "meaning": "native hand foreground intensity; individual hand or union declared in metadata; not instance codes or calibrated probability",
+      "zero": "background",
+      "full_scale": "foreground",
+      "mapping": "foreground_category in component metadata",
+      "valid_shape": [
+        "T"
+      ],
+      "required_metadata": [
+        "segmentation_type",
+        "coverage",
+        "value_range",
+        "normalization_scale",
+        "foreground_category",
+        "entity_id",
+        "storage"
+      ],
+      "shape": [
+        "T",
+        "H",
+        "W"
+      ],
+      "allowed_storage_profiles": [
+        "mkv/uint16_ffv1/v1"
+      ]
+    }
+  },
+  "note": "Registry for hand-warehouse/1.0. Read CONTRACT.md for field semantics. This is not a JSON Schema document or executable validator.",
+  "components": {
+    "mano/axis_angle61/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        2,
+        61
+      ],
+      "valid_shape": [
+        "T",
+        2
+      ],
+      "unit": "m",
+      "angle_unit": "rad",
+      "coordinate_frame": "camera",
+      "layout": {
+        "pose_axis_angle": [
+          0,
+          48
+        ],
+        "shape": [
+          48,
+          58
+        ],
+        "translation": [
+          58,
+          61
+        ]
+      }
+    },
+    "keypoints/uv/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        2,
+        21,
+        2
+      ],
+      "valid_shape": [
+        "T",
+        2,
+        21
+      ],
+      "unit": "pixel",
+      "required_metadata": [
+        "keypoint_schema"
+      ]
+    },
+    "keypoints/xyz/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        2,
+        21,
+        3
+      ],
+      "valid_shape": [
+        "T",
+        2,
+        21
+      ],
+      "unit": "m",
+      "required_metadata": [
+        "keypoint_schema"
+      ]
+    },
+    "depth/z_float32/v1": {
+      "storage": "mkv",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        "H",
+        "W"
+      ],
+      "valid_shape": [
+        "T",
+        "H",
+        "W"
+      ],
+      "unit": "m",
+      "depth_type": "z",
+      "pixel_grid": "current_rgb",
+      "coordinate_frame": "current_rgb_camera",
+      "valid_storage": "mkv",
+      "allowed_storage_profiles": [
+        "mkv/depth_float32_bits/v1"
+      ],
+      "allowed_valid_storage_profiles": [
+        "mkv/bool_ffv1/v1"
+      ],
+      "required_metadata": [
+        "storage",
+        "valid_storage",
+        "depth_origin"
+      ],
+      "depth_origin_values": [
+        "measured",
+        "rendered",
+        "unknown"
+      ]
+    },
+    "transform/se3/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        4,
+        4
+      ],
+      "valid_shape": [
+        "T"
+      ],
+      "unit": "m",
+      "required_metadata": [
+        "convention"
+      ]
+    },
+    "object/rigid/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        4,
+        4
+      ],
+      "valid_shape": [
+        "T"
+      ],
+      "unit": "m",
+      "required_metadata": [
+        "convention"
+      ]
+    },
+    "camera/pinhole/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        3,
+        3
+      ],
+      "valid_shape": [
+        "T"
+      ]
+    },
+    "camera/rational8/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        3,
+        3
+      ],
+      "valid_shape": [
+        "T"
+      ]
+    },
+    "camera/opencv_fisheye4/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        3,
+        3
+      ],
+      "valid_shape": [
+        "T"
+      ]
+    },
+    "camera/fisheye624/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        15
+      ],
+      "valid_shape": [
+        "T"
+      ]
+    },
+    "camera/distortion8/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        8
+      ],
+      "valid_shape": [
+        "T"
+      ]
+    },
+    "camera/fisheye4_distortion/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        4
+      ],
+      "valid_shape": [
+        "T"
+      ]
+    },
+    "object/joint_state/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        "J"
+      ],
+      "valid_shape": [
+        "T",
+        "J"
+      ],
+      "required_metadata": [
+        "joint_order"
+      ],
+      "unit_rule": "revolute radians; prismatic metres; mixed types declared per joint"
+    },
+    "object_keypoints/camera/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        "K",
+        3
+      ],
+      "valid_shape": [
+        "T",
+        "K"
+      ],
+      "unit": "m",
+      "coordinate_frame": "camera",
+      "required_metadata": [
+        "keypoint_schema"
+      ]
+    },
+    "object/dimensions/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        3
+      ],
+      "valid_shape": [
+        "T"
+      ],
+      "unit": "m",
+      "required_metadata": [
+        "dimensions_semantics"
+      ],
+      "axis_order": [
+        "length",
+        "width",
+        "height"
+      ],
+      "scale_rule": "do not apply to mesh without source-supported size/scale relation"
+    },
+    "object_keypoints/template/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "K",
+        3
+      ],
+      "valid_shape": [
+        "K"
+      ],
+      "unit": "m",
+      "coordinate_frame": "object",
+      "static": true,
+      "required_metadata": [
+        "keypoint_schema"
+      ]
+    },
+    "mesh/obj/v1": {
+      "storage": "obj",
+      "unit": "m",
+      "static": true,
+      "valid_shape": [],
+      "dependencies": "MTL and textures retained under owning object"
+    },
+    "rgb/uint8/v1": {
+      "storage": "mkv",
+      "dtype": "uint8",
+      "shape": [
+        "T",
+        "H",
+        "W",
+        3
+      ],
+      "valid_shape": [
+        "T"
+      ],
+      "color_order": "RGB",
+      "allowed_storage_profiles": [
+        "mkv/mjpeg_copy/v1",
+        "mkv/source_copy/v1",
+        "mkv/h264_rgb_lossless/v1",
+        "mkv/ffv1_rgb/v1"
+      ],
+      "required_metadata": [
+        "storage"
+      ]
+    },
+    "contact/relations/v1": {
+      "storage": "jsonl",
+      "valid_shape": [
+        "T"
+      ],
+      "frame_keys": [
+        "frame_index",
+        "relations"
+      ],
+      "absence_semantics": "no record is not a negative contact annotation"
+    },
+    "contact/vertex_region/v1": {
+      "storage": "pt",
+      "dtype": "bool",
+      "shape": [
+        "T",
+        "V"
+      ],
+      "valid_shape": [
+        "T",
+        "V"
+      ],
+      "required_metadata": [
+        "source_geometry_ref",
+        "target_geometry_ref",
+        "topology_id"
+      ],
+      "pilot_status": "defined_not_empirically_validated"
+    },
+    "contact/vertex_distance/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        "V"
+      ],
+      "valid_shape": [
+        "T",
+        "V"
+      ],
+      "unit": "m",
+      "required_metadata": [
+        "source_geometry_ref",
+        "target_geometry_ref",
+        "topology_id",
+        "distance_definition"
+      ],
+      "pilot_status": "defined_not_empirically_validated"
+    },
+    "object/articulated/v1": {
+      "storage": "json",
+      "required_keys": [
+        "parts",
+        "joints"
+      ],
+      "part_keys": [
+        "part_id",
+        "parent_id",
+        "has_model",
+        "mesh_ref"
+      ],
+      "hierarchy_semantics": "declare mechanical articulation or semantic containment; containment alone does not imply joints",
+      "acyclic": true
+    },
+    "segment/uint16_labels/v1": {
+      "storage": "mkv",
+      "dtype": "uint16",
+      "mapping": "instances.jsonl",
+      "valid_shape": [
+        "T"
+      ],
+      "required_metadata": [
+        "segmentation_type",
+        "coverage",
+        "mapping_ref",
+        "storage"
+      ],
+      "shape": [
+        "T",
+        "H",
+        "W"
+      ],
+      "reserved_codes": {
+        "0": "unassigned",
+        "65535": "ignore"
+      },
+      "region_codes": [
+        1,
+        65534
+      ],
+      "allowed_storage_profiles": [
+        "mkv/uint16_ffv1/v1"
+      ]
+    },
+    "segment/instance_polygons/v1": {
+      "storage": "jsonl",
+      "mapping": "instances.jsonl",
+      "frame_keys": [
+        "frame_index",
+        "instances"
+      ],
+      "instance_keys": [
+        "mask_value",
+        "polygons"
+      ],
+      "coordinate_unit": "pixel",
+      "overlap_allowed": true,
+      "valid_shape": [
+        "T"
+      ],
+      "required_metadata": [
+        "segmentation_type",
+        "coverage",
+        "mapping_ref"
+      ],
+      "polygon_shape": [
+        "N",
+        2
+      ],
+      "holes_supported": false
+    },
+    "segment/foreground_intensity/v1": {
+      "storage": "mkv",
+      "dtype": "uint16",
+      "value_range": [
+        0,
+        255
+      ],
+      "normalization_scale": 0.00392156862745098,
+      "axes": [
+        "T",
+        "height",
+        "width"
+      ],
+      "meaning": "native hand foreground intensity; individual hand or union declared in metadata; not instance codes or calibrated probability",
+      "zero": "background",
+      "full_scale": "foreground",
+      "mapping": "foreground_category in component metadata",
+      "valid_shape": [
+        "T"
+      ],
+      "required_metadata": [
+        "segmentation_type",
+        "coverage",
+        "value_range",
+        "normalization_scale",
+        "foreground_category",
+        "entity_id",
+        "storage"
+      ],
+      "shape": [
+        "T",
+        "H",
+        "W"
+      ],
+      "allowed_storage_profiles": [
+        "mkv/uint16_ffv1/v1"
+      ]
+    }
+  },
+  "storage_revision": "lossless-video-2026-09-18",
+  "sampling": {
+    "image": {
+      "single": {
+        "T": 1
+      },
+      "independent": {
+        "T": ">=2",
+        "required_frame_fields": [
+          "sample_id",
+          "source_locator"
+        ],
+        "temporal_relationship": false
+      },
+      "source_fps": null
+    },
+    "video": {
+      "sampling_types": [
+        "dense",
+        "sparse"
+      ]
+    }
+  },
+  "storage_profiles": {
+    "mkv/mjpeg_copy/v1": {
+      "container": "matroska",
+      "codec": "mjpeg",
+      "method": "stream_copy_from_source_jpeg",
+      "logical_dtype": "uint8",
+      "decoded_pixel_format": "rgb24",
+      "validation": [
+        "source_jpeg_packet_sha256_exact",
+        "all_packet_count_order_and_dimensions_exact",
+        "deterministic_decoded_rgb_sample_exact",
+        "pilot_and_final_selected_whole_sequences_decoded_rgb_exact"
+      ],
+      "validation_policy": {
+        "id": "jpeg-packets-all-rgb-sample16/v1",
+        "sample_count_per_view": 16,
+        "short_views": "all_frames_if_T_less_than_16",
+        "fixed_indices": "0,T-1,T//4,T//2,3*T//4, deduplicated",
+        "remaining_indices": "sample without replacement to min(T,16) using Python random.Random(seed)",
+        "seed": "first 8 bytes of SHA256(sequence_id/camera_id/20260920), big-endian integer",
+        "report": "policy, exact selected frame indices, seed, separate packet and pixel coverage",
+        "legacy_all_frame_pixel_proofs": "still accepted; preserve historical evidence"
+      }
+    },
+    "mkv/source_copy/v1": {
+      "container": "matroska",
+      "codec": "source_codec_declared_in_metadata",
+      "method": "stream_copy_only",
+      "logical_dtype": "uint8",
+      "decoded_pixel_format": "rgb24",
+      "validation": [
+        "no_reencoding",
+        "all_decoded_rgb_frames_exact",
+        "presentation_frame_order_exact"
+      ]
+    },
+    "mkv/h264_rgb_lossless/v1": {
+      "container": "matroska",
+      "codec": "h264",
+      "encoder": "libx264rgb",
+      "crf": 0,
+      "decoded_pixel_format": "rgb24",
+      "logical_dtype": "uint8",
+      "gop": "declared_in_encoder_parameters",
+      "validation": [
+        "all_decoded_rgb_frames_exact"
+      ],
+      "forbidden": [
+        "lossy_rate_control",
+        "rgb_to_yuv420_subsampling"
+      ]
+    },
+    "mkv/ffv1_rgb/v1": {
+      "container": "matroska",
+      "codec": "ffv1",
+      "decoded_pixel_format": "rgb24",
+      "logical_dtype": "uint8",
+      "validation": [
+        "all_decoded_rgb_frames_exact"
+      ]
+    },
+    "mkv/depth_float32_bits/v1": {
+      "container": "matroska",
+      "codec": "ffv1",
+      "pixel_format": "gray16le",
+      "logical_dtype": "float32",
+      "logical_shape": [
+        "T",
+        "H",
+        "W"
+      ],
+      "encoded_shape": [
+        "T",
+        "H",
+        "2*W"
+      ],
+      "packing": "IEEE754 little-endian float32 bits; low uint16 in first W columns, high uint16 in last W columns",
+      "decode": "reinterpret_float32(uint32(low) | (uint32(high) << 16))",
+      "validation": [
+        "all_float32_bits_exact_including_nan"
+      ]
+    },
+    "mkv/bool_ffv1/v1": {
+      "container": "matroska",
+      "codec": "ffv1",
+      "pixel_format": "gray",
+      "logical_dtype": "bool",
+      "stored_values": [
+        0,
+        1
+      ],
+      "validation": [
+        "all_bool_values_exact"
+      ]
+    },
+    "mkv/uint16_ffv1/v1": {
+      "container": "matroska",
+      "codec": "ffv1",
+      "pixel_formats": [
+        "gray",
+        "gray16le"
+      ],
+      "logical_dtype": "uint16",
+      "gray_condition": "all values <=255; decode by exact uint8-to-uint16 promotion",
+      "validation": [
+        "all_integer_values_exact"
+      ]
+    }
+  },
+  "video_metadata": {
+    "required": [
+      "profile",
+      "codec",
+      "pixel_format",
+      "encoded_size",
+      "decoder",
+      "encoder_parameters",
+      "source_lossy"
+    ],
+    "source_lossy_values": [
+      true,
+      false,
+      "unknown"
+    ],
+    "timebase_role": "index_only",
+    "lossless_reference": "source compressed packets or fixed-decoder source pixels; depth uses normalized standard float32 bits",
+    "selection": "choose smaller complete candidate only after exact roundtrip validation"
+  }
+}

H2O/shared-v1.0/script/provenance/previous-run/full.log

@@ -0,0 +1,61 @@
+Verified existing subject1/h1/0
+Verified existing subject4/h1/0
+Verified existing subject3/h1/0
+Verified existing subject1/h1/6
+Verified existing subject2/h1/0
+Verified existing subject1/h1/3
+Verified existing subject2/h1/6
+Verified existing subject1/h1/7
+Converted subject3/h1/6
+Converted subject4/h1/6
+Verified existing subject1/h1/4
+Verified existing subject1/h1/2
+Verified existing subject2/h1/3
+Verified existing subject1/h1/1
+Verified existing subject1/h1/5
+Verified existing subject1/o1/0
+Converted subject3/h1/3
+Verified existing subject2/h1/7
+Verified existing subject1/o1/6
+Converted subject4/h1/3
+Verified existing subject1/o1/3
+Verified existing subject2/h1/4
+Verified existing subject1/o1/7
+Verified existing subject1/o1/4
+Verified existing subject2/h1/2
+Verified existing subject1/o1/2
+Converted subject3/h1/7
+Converted subject4/h1/7
+Verified existing subject1/o1/1
+Verified existing subject2/h1/1
+Verified existing subject1/o1/5
+Verified existing subject1/o2/0
+Converted subject3/h1/4
+Verified existing subject2/h1/5
+Verified existing subject1/o2/6
+Verified existing subject1/o2/3
+Verified existing subject2/o1/0
+Converted subject4/h1/4
+Verified existing subject1/o2/7
+Verified existing subject1/o2/4
+Converted subject3/h1/2
+Verified existing subject2/o1/6
+Converted subject4/h1/2
+Verified existing subject1/o2/2
+Verified existing subject2/o1/3
+Converted subject3/h1/1
+Verified existing subject1/o2/1
+Verified existing subject1/o2/5
+Verified existing subject1/k1/0
+Verified existing subject1/k1/6
+Verified existing subject2/o1/4
+Verified existing subject1/k1/3
+Converted subject3/h1/5
+Converted subject4/h1/1
+Verified existing subject1/k1/7
+Verified existing subject2/o1/2
+Converted subject4/h1/5
+Converted subject3/o1/0
+Converted subject1/k1/4
+Converted subject4/o1/0
+Verified existing subject2/o1/1

H2O/shared-v1.0/script/provenance/previous-run/full.status

@@ -0,0 +1,7 @@
+START=2026-09-30T11:03:05+08:00
+WORKERS=4 CODEC_THREADS=10 DECODE_THREADS=10 DEPTH_THREADS=4 SUBJECTS=1,2,3,4 CASA=false
+CPU_LIMIT=none
+CGROUP=/user.slice/user-25044.slice/user@25044.service/app.slice/h2o-full-d10-20260930.scope
+CPU_MAX=max controller_disabled
+
+RECONFIGURE=2026-09-30T15:47:55+08:00 REASON=process16_camera_pipeline_ram32_streaming

H2O/shared-v1.0/script/provenance/previous-run/run-full.sh

@@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+set -u
+base="$HOME/Datasets/uni-hoi-h2o-full-20260930-w4-c10-d10"
+raw="$HOME/Datasets/Hand-Datasets-Raw/H2O"
+cd "$HOME/Projects/uni-hoi-dataset" || exit 1
+mkdir -p "$base/scratch"
+export TMPDIR="$base/scratch" PYTHONUNBUFFERED=1
+for n in 1 2 3 4; do test -s "$raw/raw/data/subject${n}_v1_1.tar.gz" || exit 2; done
+test ! -e "$base/full.status" || exit 3
+cg=$(awk -F: '$1=="0" {print $3}' /proc/self/cgroup)
+if [[ -f "/sys/fs/cgroup$cg/cpu.max" ]]; then
+    read -r quota period < "/sys/fs/cgroup$cg/cpu.max"
+else
+    quota=max
+    period=controller_disabled
+fi
+printf 'START=%s\nWORKERS=4 CODEC_THREADS=10 DECODE_THREADS=10 DEPTH_THREADS=4 SUBJECTS=1,2,3,4 CASA=false\nCPU_LIMIT=none\nCGROUP=%s\nCPU_MAX=%s %s\n' "$(date -Is)" "$cg" "$quota" "$period" > "$base/full.status"
+if [[ "$quota" != max ]]; then
+    printf 'EXIT=4 ERROR=UNEXPECTED_CPU_QUOTA\n' >> "$base/full.status"
+    exit 4
+fi
+/usr/bin/time -v .venv/bin/python "$base/frozen-code/scripts/convert_h2o_v1.0.py" --root "$raw" --out "$base" --release-id shared-v1.0 --workers 4 --codec-threads 10 --decode-threads 10 --depth-threads 4 --mano-dir "$HOME/Projects/WiLoR/mano_data" > "$base/full.log" 2>&1
+code=$?
+printf 'EXIT=%s END=%s\n' "$code" "$(date -Is)" >> "$base/full.status"
+if [[ -f "/sys/fs/cgroup$cg/cpu.stat" ]]; then cat "/sys/fs/cgroup$cg/cpu.stat" >> "$base/full.status"; fi
+exit "$code"

H2O/shared-v1.0/script/provenance/previous-run/script/check_dexycb_projection.py

@@ -0,0 +1,31 @@
+#!/usr/bin/env python3
+"""Check saved Dex-YCB source 3D joints against saved source 2D joints in sampled views."""
+import json
+from pathlib import Path
+import sys
+
+import numpy as np
+import torch
+
+root = Path(sys.argv[1])
+report = json.loads((root / 'doc/validation/STRUCTURAL_AUDIT.json').read_text())
+checks = []
+for name in report['sampled_views']:
+    view = root / name
+    load = lambda path: torch.load(view / path, weights_only=True, map_location='cpu').numpy()
+    k = load('camera/intrinsic.pt')
+    xyz = load('annotation/keypoints/xyz.pt')
+    uv = load('annotation/keypoints/uv.pt')
+    valid = load('annotation/keypoints/xyz_valid.pt') & load('annotation/keypoints/uv_valid.pt')
+    z = xyz[..., 2]
+    valid &= np.isfinite(xyz).all(axis=-1) & np.isfinite(uv).all(axis=-1) & (z > 0)
+    z = np.where(valid, z, 1)
+    pred = np.stack((k[:, 0, 0, None, None] * xyz[..., 0] / z + k[:, 0, 2, None, None],
+                     k[:, 1, 1, None, None] * xyz[..., 1] / z + k[:, 1, 2, None, None]), axis=-1)
+    errors = np.linalg.norm(pred[valid] - uv[valid], axis=-1)
+    checks.append(dict(view=name, points=int(errors.size), median_px=float(np.median(errors)) if errors.size else None,
+                       p95_px=float(np.percentile(errors, 95)) if errors.size else None))
+(root / 'doc/validation/KEYPOINT_PROJECTION_QC.json').write_text(json.dumps({
+    'scope': 'Saved source 3D joints projected with saved color intrinsics vs saved source 2D joints; independent of hand mesh rendering.',
+    'views': checks}, ensure_ascii=False, indent=2) + '\n')
+print(json.dumps(checks, ensure_ascii=False), flush=True)

H2O/shared-v1.0/script/provenance/previous-run/script/config.json

@@ -0,0 +1,17 @@
+{
+  "root": "/home/xuzhefeng/Datasets/Hand-Datasets-Raw/H2O",
+  "out": "/home/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-w4-c10-d10",
+  "release_id": "shared-v1.0",
+  "limit": null,
+  "max_frames": null,
+  "mano_dir": "/home/xuzhefeng/Projects/WiLoR/mano_data",
+  "views": null,
+  "jpeg_full_check": false,
+  "workers": 4,
+  "codec_threads": 10,
+  "decode_threads": 10,
+  "depth_threads": 4,
+  "subjects": null,
+  "include_casa": false,
+  "archive": null
+}

H2O/shared-v1.0/script/provenance/previous-run/script/convert_dexycb_v1.0.py

@@ -0,0 +1,349 @@
+#!/usr/bin/env python3
+"""Dex-YCB source annotations → CONTRACT.md v1.0. No derived annotations."""
+import io
+import re
+import shutil
+import tarfile
+import tempfile
+from pathlib import Path
+
+import numpy as np
+from PIL import Image
+import yaml
+
+from warehouse_v1_0 import (View, build_record, find_file, locator, mano_models, mesh_asset,
+    model_profile, parser, pca_to_axis_angle, release, roots, safe_member, sequence_dir)
+
+IMG_W, IMG_H = 640, 480
+
+class _Loader(yaml.SafeLoader):
+    pass
+
+
+_Loader.add_constructor("tag:yaml.org,2002:python/tuple",
+                        lambda l, n: tuple(l.construct_sequence(n)))
+
+
+def yaml_load_bytes(data: bytes):
+    return yaml.load(io.BytesIO(data), Loader=_Loader)
+
+
+class Calibration:
+    def __init__(self, dexycb_root: Path):
+        self.intrinsics = {}   # serial -> {fx,fy,cx,cy,w,h}
+        self.extrinsics = {}   # group  -> {serial: [12] 行优先 3×4 cam→world}
+        self.betas = {}        # mano_calib_name -> [10]
+        calib_dir = dexycb_root / "calibration"
+        if calib_dir.is_dir():
+            self._load((p.as_posix(), p.read_bytes())
+                       for p in sorted(calib_dir.rglob("*")) if p.is_file())
+        else:
+            tar_path = dexycb_root / "calibration.tar.gz"
+            if not tar_path.exists():
+                raise FileNotFoundError(f"no calibration/ nor calibration.tar.gz in {dexycb_root}")
+            with tarfile.open(tar_path, "r:gz") as tf:
+                self._load((m.name, tf.extractfile(m).read()) for m in tf if m.isfile())
+
+    _RE_INTR = re.compile(r"(?:^|/)intrinsics/(\d+)_\d+x\d+\.yml$")
+    _RE_EXTRA = re.compile(r"(?:^|/)extrinsics_(\d{8}_\d{6})/extrinsics\.yml$")
+    _RE_MANO = re.compile(r"(?:^|/)mano_(\d{8}_\d{6}_subject-\d{2}_right)/mano\.yml$")
+
+    def _load(self, files):
+        for rel, data in files:
+            m = self._RE_INTR.search(rel)
+            if m:
+                c = yaml_load_bytes(data)["color"]
+                self.intrinsics[m.group(1)] = {"fx": float(c["fx"]), "fy": float(c["fy"]),
+                                               "cx": float(c["ppx"]), "cy": float(c["ppy"]),
+                                               "w": IMG_W, "h": IMG_H}
+                continue
+            m = self._RE_EXTRA.search(rel)
+            if m:
+                raw = yaml_load_bytes(data)["extrinsics"]
+                self.extrinsics[m.group(1)] = {
+                    str(s): [float(x) for x in np.asarray(v).ravel()] for s, v in raw.items()}
+                continue
+            m = self._RE_MANO.search(rel)
+            if m:
+                self.betas[m.group(1)] = [float(b) for b in yaml_load_bytes(data)["betas"]]
+
+
+
+def calibration_locator(path, member, **extra):
+    if path.is_dir():
+        return locator(path / member.removeprefix('calibration/'), **extra)
+    return locator(path, member, **extra)
+
+
+def prepare_models(root, destination):
+    archive = find_file(root, 'models.tar.gz')
+    if archive is None:
+        return {}, None
+    pattern = re.compile(r'^models/(\d{3}_[a-z0-9_]+)/(textured_simple\.obj(?:\.mtl)?|texture_map\.png)$')
+    names = set()
+    with tarfile.open(archive, 'r|gz') as tf:
+        for member in tf:
+            match = pattern.fullmatch(member.name)
+            if member.isfile() and match:
+                names.add(match[1])
+                path = safe_member(destination, '/'.join(match.groups()))
+                path.parent.mkdir(parents=True, exist_ok=True)
+                path.write_bytes(tf.extractfile(member).read())
+    if len(names) != 21:
+        raise ValueError('expected all 21 source YCB model directories')
+    return {i: destination / name for i, name in enumerate(sorted(names), 1)}, archive
+
+
+def convert_view(path, args, seq, serial, frames, meta, meta_loc, calib, calib_path, models, assets, asset_archive):
+    total = int(meta.get('num_frames', max(frames) + 1))
+    ids = list(range(min(total, args.max_frames) if args.max_frames else total))
+    for f in ids:
+        frames.setdefault(f, {})
+    if not ids:
+        raise ValueError('no RGB frames')
+    source_rows = []
+    for f in ids:
+        row = frames[f]
+        source_rows.append(dict(source_frame_id=f, sync_id=f'{seq}:{f}',
+            source_locator=row['color'][1] if 'color' in row else (meta_loc | {'source_frame_id': f, 'missing_rgb': True}), annotation_matches={
+                'labels': {'source_frame_id': f, 'source_timestamp_ns': None, 'clock_domain': None,
+                           'match_method': 'index_map', 'delta_ns': None, 'source_locator': row['labels'][1]}
+                for _ in [0] if 'labels' in row},
+            component_sources={k: v[1] for k, v in row.items()}))
+    view = View(path, 'dex-ycb', args.release_id, seq, serial, source_rows,
+                sampling='dense' if ids == list(range(ids[0], ids[-1] + 1)) else 'sparse')
+    view.valid_rules = {
+        'mano': 'Source pose_m present and not all zero; side model/calibration available',
+        'keypoints': 'Corresponding joint_2d/joint_3d present; whole-hand all -1 is missing, broadcast to its 21 points',
+        'pose': 'Source pose_y for this ycb_ids slot present and not an all-zero matrix',
+        'segment': 'Source seg array provided for this frame; region codes and entity references validated',
+        'relations': 'Source sequence ycb_grasp_ind resolves through ycb_ids; broadcast sequence grasp to selected frames',
+    }
+    view.full_jpeg = getattr(args, 'jpeg_full_check', False)
+    view.meta['view_type'] = 'exo'
+    view.meta['sampling']['selection'] = {'max_frames': args.max_frames, 'views': args.views}
+    subject = re.search(r'subject-(\d+)', seq)
+    if subject and 1 <= int(subject[1]) <= 10:
+        index = int(subject[1])
+        view.meta.update(split='val' if index == 7 else 'test' if index in (8, 9) else 'train',
+            split_scheme='Dex-YCB/s1', split_source_locator={
+                'url': 'https://github.com/NVlabs/dex-ycb-toolkit/blob/master/dex_ycb_toolkit/dex_ycb.py',
+                'rule': 's1 unseen-subject: 01-06/10 train, 07 val, 08-09 test'})
+        view.meta.pop('split_reason', None)
+    view.rgb([frames[f]['color'][0] if 'color' in frames[f] else None for f in ids], {'frames_ref': 'camera/frames.jsonl'})
+    n = len(ids)
+    intr = calib.intrinsics.get(serial)
+    K = None if intr is None else np.array([[intr['fx'], 0, intr['cx']], [0, intr['fy'], intr['cy']], [0, 0, 1]])
+    ext = calib.extrinsics.get(str(meta['extrinsics']), {}).get(serial)
+    E = None
+    if ext is not None:
+        E = np.eye(4)
+        E[:3] = np.asarray(ext).reshape(3, 4)
+        E = np.linalg.inv(E)
+    view.camera(K, calibration_locator(calib_path, f'calibration/intrinsics/{serial}_640x480.yml'), E,
+        calibration_locator(calib_path, f'calibration/extrinsics_{meta["extrinsics"]}/extrinsics.yml',
+                conversion='inverse of source T_world_from_camera'))
+    sides = meta.get('mano_sides')
+    if sides not in (['left'], ['right']):
+        raise ValueError(f'expected exactly one anatomical source hand: {sides}')
+    side = sides[0]
+    slot = ['left', 'right'].index(side)
+    cal_name = meta['mano_calib'][0] if isinstance(meta['mano_calib'], list) else meta['mano_calib']
+    labels = [dict(np.load(frames[f]['labels'][0])) if 'labels' in frames[f] else {} for f in ids]
+    label_loc = {'frames_ref': 'camera/frames.jsonl', 'source_field': 'component_sources.labels'}
+    motion = np.full((n, 2, 61), np.nan, np.float32)
+    mv = np.zeros((n, 2), bool)
+    if side in models and any('pose_m' in lab for lab in labels):
+        for i, lab in enumerate(labels):
+            if 'pose_m' not in lab:
+                continue
+            pose = np.asarray(lab['pose_m']).reshape(51)
+            if not np.any(pose):  # Official Dex-YCB missing-pose sentinel, independent of joints/seg.
+                continue
+            motion[i, slot] = np.r_[pose[:3], pca_to_axis_angle(pose[3:48], models[side]),
+                                     calib.betas[cal_name], pose[48:51]]
+            mv[i, slot] = True
+        view.tensor('mano', 'mano/axis_angle61/v1', 'annotation/mano/motion.pt', motion, mv,
+                    ['T', 'hand', 'parameter'], label_loc, model_profile=model_profile(models),
+                    calibration_source=calibration_locator(calib_path, f'calibration/mano_{cal_name}/mano.yml'),
+                    provenance={'kind': 'format_conversion', 'method': 'source PCA @ hands_components + hands_mean; explicit axis-angle consumed with flat_hand_mean=true'})
+    elif any('pose_m' in lab for lab in labels):
+        view.absent('mano', 'Source PCA exists; provide --mano-dir to expand it', 'dependency_missing')
+    else:
+        view.absent('mano', 'No source pose_m field')
+    for source_key, key, dim, unit in [('joint_2d', 'uv', 2, 'pixel'), ('joint_3d', 'xyz', 3, 'm')]:
+        if not any(source_key in x for x in labels):
+            view.absent('keypoints_' + key, 'No source ' + source_key)
+            continue
+        points = np.full((n, 2, 21, dim), np.nan, np.float32)
+        valid = np.zeros((n, 2, 21), bool)
+        for i, lab in enumerate(labels):
+            if source_key not in lab:
+                continue
+            p = np.asarray(lab[source_key]).reshape(21, dim)
+            points[i, slot] = p
+            valid[i, slot] = not np.all(p == -1)  # Official sentinel covers the whole source hand.
+            # An individual off-image point (-1, -1) is not evidence of missing annotation.
+        view.tensor('keypoints_' + key, f'keypoints/{key}/v1', f'annotation/keypoints/{key}.pt',
+                    points, valid, ['T', 'hand', 'keypoint', 'coordinate'], label_loc,
+                    keypoint_schema='mano21/v1', unit=unit)
+    ycb_ids = [int(x) for x in meta['ycb_ids']]
+    for j, oid in enumerate(ycb_ids):
+        entity = f'ycb_{oid:02d}'
+        view.object(entity, meta_loc, source_object_id=oid, category_id=oid)
+        poses = np.full((n, 4, 4), np.nan, np.float32)
+        valid = np.zeros(n, bool)
+        for i, lab in enumerate(labels):
+            if 'pose_y' in lab:
+                pose = np.asarray(lab['pose_y'])
+                if pose.shape != (len(ycb_ids), 3, 4):
+                    raise ValueError('pose_y shape differs from ycb_ids')
+                if np.any(pose[j]):  # Official all-zero missing object pose sentinel.
+                    poses[i] = np.eye(4)
+                    poses[i, :3] = pose[j]
+                    valid[i] = True
+        if any('pose_y' in lab for lab in labels):
+            view.tensor('pose_' + entity, 'object/rigid/v1', f'annotation/object/{entity}/pose.pt',
+                    poses, valid, ['T', 'row', 'column'], label_loc,
+                    convention='T_camera_from_object', unit='m')
+        else:
+            view.absent('pose_' + entity, 'No source pose_y field')
+        if oid in assets:
+            mesh_asset(view, entity, assets[oid], assets[oid] / 'textured_simple.obj',
+                locator(asset_archive, f'models/{assets[oid].name}/textured_simple.obj'))
+        else:
+            view.absent('mesh_' + entity, 'models.tar.gz unavailable', 'not_downloaded')
+    if any('seg' in lab for lab in labels):
+        mask_rows = []
+        width, height = view.meta['camera']['image_size']
+        for i, lab in enumerate(labels):
+            values = np.unique(lab.get('seg', np.zeros((height, width), np.uint8)))
+            mappings = []
+            for value in values:
+                value = int(value)
+                if value == 0:
+                    continue
+                if value != 255 and value not in ycb_ids:
+                    raise ValueError(f'unknown Dex-YCB region value {value}')
+                mappings.append(dict(mask_value=value, category_id='hand' if value == 255 else value,
+                    entity_id='hand_' + side if value == 255 else f'ycb_{value:02d}', part_id=None,
+                    source_instance_id=value))
+            mask_rows.append(dict(frame_index=i, labels=mappings))
+        view.raster('segment', (lab.get('seg', np.zeros((height, width), np.uint16)) for lab in labels),
+                    ['seg' in lab for lab in labels], label_loc, mask_rows)
+        view.meta['components']['segment'].update(coverage={'kind': 'selected_entities',
+            'entities': ['hand_' + side] + [f'ycb_{o:02d}' for o in ycb_ids]}, source_zero_meaning='background')
+    else:
+        view.absent('segment', 'No source seg field')
+    if any('aligned_depth_to_color' in frames[f] for f in ids):
+        def depths():
+            width, height = view.meta['camera']['image_size']
+            for f in ids:
+                item = frames[f].get('aligned_depth_to_color')
+                if item is None:
+                    yield np.full((height, width), np.nan, np.float32)
+                else:
+                    with Image.open(item[0]) as im:
+                        yield np.asarray(im).astype(np.float32) / 1000.0
+        view.raster('depth', depths(), None, {'frames_ref': 'camera/frames.jsonl',
+            'source_field': 'component_sources.aligned_depth_to_color', 'conversion': 'millimetres to metres'})
+    else:
+        view.absent('depth', 'No source aligned_depth_to_color frames')
+    grasp = meta.get('ycb_grasp_ind')
+    if grasp is not None:
+        if not 0 <= int(grasp) < len(ycb_ids):
+            raise ValueError('invalid ycb_grasp_ind')
+        relation = dict(relation_type='grasp', subject_id='hand_' + side,
+            object_id=f'ycb_{ycb_ids[int(grasp)]:02d}', status='positive', source_locator=meta_loc,
+            provenance={'kind': 'source', 'method': 'sequence ycb_grasp_ind resolved through ycb_ids; not per-frame contact'})
+        view.relations([{'frame_index': i, 'relations': [relation]} for i in range(n)], np.ones(n, bool), meta_loc)
+    else:
+        view.absent('relations', 'No source grasp/contact relation')
+    return view.finish()
+
+
+MEMBER = re.compile(r'^([^/]+)/(\d{8}_\d{6})/(meta\.yml|(\d{12})/(color|labels|aligned_depth_to_color)_(\d+)\.(jpg|npz|png))$')
+
+
+def convert_archive(archive, args, out, calib, calib_path, models, assets, asset_archive):
+    selected_views = set(args.views.split(',')) if args.views else None
+    done = 0
+    current, meta, meta_loc, rows = None, None, None, {}
+    with tempfile.TemporaryDirectory(prefix='dex-source-') as wd:
+        work = Path(wd)
+        def flush():
+            nonlocal done
+            if not rows:
+                return
+            seq = archive.name.split('.tar')[0] + '_' + current
+            reports = {}
+            with sequence_dir(out, seq) as target:
+                for serial, frames in sorted(rows.items()):
+                    reports[serial] = convert_view(target / serial, args, seq, serial, frames, meta,
+                                                   meta_loc, calib, calib_path, models, assets, asset_archive)
+            build_record(out, seq, reports)
+            for staged in list(work.iterdir()):
+                if staged.is_dir():
+                    shutil.rmtree(staged)
+                else:
+                    staged.unlink()
+            done += 1
+            print(f'Converted {seq}', flush=True)
+        with tarfile.open(archive, 'r|gz') as tf:
+            for member in tf:
+                match = MEMBER.fullmatch(member.name)
+                if not member.isfile() or not match:
+                    continue
+                seq = match[2]
+                if seq != current:
+                    flush()
+                    if args.limit and done >= args.limit:
+                        break
+                    current, meta, meta_loc, rows = seq, None, None, {}
+                data_kind = match[3]
+                if data_kind == 'meta.yml':
+                    data = tf.extractfile(member).read()
+                    meta = yaml_load_bytes(data)
+                    meta_loc = locator(archive, member.name, data)
+                    continue
+                serial, kind, f = match[4], match[5], int(match[6])
+                if selected_views and serial not in selected_views or args.max_frames and f >= args.max_frames:
+                    continue
+                data = tf.extractfile(member).read()
+                dest = safe_member(work, member.name)
+                dest.parent.mkdir(parents=True, exist_ok=True)
+                dest.write_bytes(data)
+                row = rows.setdefault(serial, {}).setdefault(f, {})
+                if kind in row:
+                    raise ValueError('duplicate archive frame')
+                row[kind] = (dest, locator(archive, member.name, data))
+            else:
+                flush()
+    return done
+
+
+def main(argv=None):
+    p = parser('dex-ycb', __doc__)
+    p.add_argument('--subjects')
+    args = p.parse_args(argv)
+    archives = sorted({a for r in roots(args.root) for a in r.glob('*-subject-*.tar.gz')})
+    if args.subjects:
+        subjects = {s.strip().zfill(2) for s in args.subjects.split(',')}
+        archives = [a for a in archives if re.search(r'subject-(\d+)', a.name)[1] in subjects]
+    if not archives:
+        p.error('no Dex-YCB subject archives found')
+    cp = find_file(args.root, 'calibration.tar.gz') or find_file(args.root, 'calibration')
+    if cp is None:
+        p.error('calibration source missing')
+    calib = Calibration(cp.parent)
+    models = mano_models(args.mano_dir)
+    out = release(args, 'dex-ycb')
+    with tempfile.TemporaryDirectory(prefix='dex-models-') as wd:
+        assets, asset_archive = prepare_models(args.root, Path(wd))
+        for archive in archives:
+            convert_archive(archive, args, out, calib, cp, models, assets, asset_archive)
+    return 0
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())

H2O/shared-v1.0/script/provenance/previous-run/script/convert_h2o_v1.0.py

@@ -0,0 +1,545 @@
+#!/usr/bin/env python3
+"""H2O native RGB-D, MANO, joints and objects → CONTRACT.md v1.0 (no rendered masks)."""
+from concurrent.futures import ThreadPoolExecutor
+import json
+import os
+import re
+import shutil
+import tarfile
+import tempfile
+from pathlib import Path
+import zipfile
+
+# Bound native libraries before loading NumPy/Torch; each worker still uses independent FFmpeg processes.
+for key in ('OMP_NUM_THREADS', 'OPENBLAS_NUM_THREADS', 'MKL_NUM_THREADS', 'NUMEXPR_NUM_THREADS'):
+    os.environ[key] = '1'
+
+import numpy as np
+from PIL import Image
+import torch
+
+from warehouse_v1_0 import (View, build_record, check_mesh, find_file, locator, mano_models, mesh_asset,
+    model_profile, parser, release, roots, safe_member, sequence_dir, sha256)
+
+OBJECTS = {1: 'book', 2: 'espresso', 3: 'lotion', 4: 'spray',
+           5: 'milk', 6: 'cocoa', 7: 'chips', 8: 'cappuccino'}
+MEMBER = re.compile(r'^(?:h2o_CASA/)?(subject\d+)/([^/]+)/(\d+)/(cam\d)/(.+)$')
+FRAME = re.compile(r'(rgb|depth|cam_pose|hand_pose|hand_pose_mano|obj_pose|obj_pose_rt)/(\d+)\.(png|jpg|txt)$')
+
+
+def load_splits(root):
+    archive = find_file(root, 'label_split.zip')
+    result = {}
+    if archive is not None:
+        with zipfile.ZipFile(archive) as z:
+            for split in ('train', 'val', 'test'):
+                member = f'label_split/action_{split}.txt'
+                for line_no, line in enumerate(z.read(member).decode().splitlines(), 1):
+                    fields = line.split()
+                    if len(fields) < 2 or fields[1] == 'path':
+                        continue
+                    key = fields[1]
+                    if key in result and result[key][0] != split:
+                        raise ValueError('conflicting official splits')
+                    result[key] = (split, locator(archive, member, record=line_no))
+    return result
+
+
+def load_assets(root, work):
+    archive = find_file(root, 'object.zip')
+    if archive is None:
+        return {}, None
+    with zipfile.ZipFile(archive) as z:
+        for member in z.infolist():
+            if member.is_dir():
+                continue
+            path = safe_member(work, member.filename)
+            path.parent.mkdir(parents=True, exist_ok=True)
+            path.write_bytes(z.read(member))
+    result = {}
+    for oid, name in OBJECTS.items():
+        directory = work / 'object' / name
+        aliases = []
+        # Source cocoa.obj -> ovomaltine.mtl -> cocoa.png, but the ZIP names the
+        # supplied texture ovomaltine.png. Keep all originals and add a byte-identical alias.
+        if name == 'cocoa' and not (directory / 'cocoa.png').exists():
+            mtl = directory / 'ovomaltine.mtl'
+            if mtl.is_file() and 'map_Kd cocoa.png' in mtl.read_text().splitlines():
+                texture = directory / 'ovomaltine.png'
+                shutil.copyfile(texture, directory / 'cocoa.png')
+                aliases.append({'file': 'cocoa.png', 'sha256': sha256(texture),
+                                'source': locator(archive, 'object/cocoa/ovomaltine.png')})
+        candidates = list(directory.glob('*.obj'))
+        if len(candidates) != 1:
+            raise ValueError(f'expected one source OBJ for {name}')
+        check_mesh(candidates[0], directory)  # Fail before decoding any large subject archive.
+        result[oid] = (directory, candidates[0], aliases)
+    return result, archive
+
+
+def convert_view(path, args, seq, cam, rows, intr, splits, models, assets, asset_archive):
+    codec_threads = getattr(args, 'codec_threads', 1)
+    ids = list(range(min(rows), max(rows) + 1))
+    for f in ids:
+        rows.setdefault(f, {})
+    n = len(ids)
+    if not n:
+        raise ValueError('no RGB frames')
+    split, split_loc = splits.get(seq, ('unknown', None))
+    frame_rows = [dict(source_frame_id=f, sync_id=f'{seq}:{f}', source_locator=rows[f]['rgb'][1] if 'rgb' in rows[f] else {'sequence': seq, 'camera': cam, 'source_frame_id': f, 'missing_rgb': True},
+                       component_sources={k: item[1] for k, item in rows[f].items()},
+                       annotation_matches={k: dict(source_frame_id=f, source_timestamp_ns=None,
+                           clock_domain=None, match_method='index_map', delta_ns=None)
+                           for k in rows[f] if k != 'rgb'}) for f in ids]
+    view = View(path, 'H2O', args.release_id, seq.replace('/', '_'), cam, frame_rows, split=split)
+    view.valid_rules = {
+        'mano': 'Per-hand hand_pose_mano flag equals 1; corresponding side model available',
+        'keypoints': "Per-hand hand_pose flag equals 1, broadcast to that hand's 21 source points independently of MANO",
+        'pose': 'Source obj_pose_rt class equals this nonzero object class',
+        'points': 'Source obj_pose class equals this nonzero object class, broadcast to its 21 points',
+    }
+    view.full_jpeg = getattr(args, 'jpeg_full_check', False)
+    view.meta['view_type'] = 'ego' if cam == 'cam4' else 'exo' if cam in ('cam0', 'cam1', 'cam2', 'cam3') else 'unknown'
+    view.meta['sampling']['selection'] = {'max_frames': args.max_frames, 'views': args.views}
+    if split_loc:
+        view.meta.update(split_scheme='official_action_split', split_source_locator=split_loc)
+    # In the measured 275 H2O views, exact FFV1 was always smaller than exact H.264.
+    view.rgb([rows[f]['rgb'][0] if 'rgb' in rows[f] else None for f in ids],
+             {'frames_ref': 'camera/frames.jsonl'}, ffv1_first=True, codec_threads=codec_threads)
+    loc = {'frames_ref': 'camera/frames.jsonl', 'source_field': 'component_sources'}
+    if intr is not None:
+        fx, fy, cx, cy, width, height = np.loadtxt(intr[0]).ravel()
+        if [int(width), int(height)] != view.meta['camera']['image_size']:
+            raise ValueError('source intrinsics resolution mismatch')
+        K = np.array([[fx, 0, cx], [0, fy, cy], [0, 0, 1]])
+        view.camera(K, intr[1])
+    camps = np.full((n, 4, 4), np.nan, np.float32)
+    cv = np.zeros(n, bool)
+    data = []
+    for i, f in enumerate(ids):
+        lab = {k: np.loadtxt(item[0]).ravel() for k, item in rows[f].items() if k not in ('rgb', 'depth')}
+        data.append(lab)
+        if 'cam_pose' in lab:
+            camps[i] = np.linalg.inv(lab['cam_pose'].reshape(4, 4))
+            cv[i] = True
+    if cv.any():
+        view.tensor('extrinsic', 'transform/se3/v1', 'camera/extrinsic.pt', camps, cv,
+                    ['T', 'row', 'column'], loc, convention='T_camera_from_world',
+                    source_convention='T_world_from_camera')
+    motion, mv = np.full((n, 2, 61), np.nan, np.float32), np.zeros((n, 2), bool)
+    points, pv = np.full((n, 2, 21, 3), np.nan, np.float32), np.zeros((n, 2, 21), bool)
+    object_ids = set()
+    for i, lab in enumerate(data):
+        if 'hand_pose_mano' in lab:
+            hands = lab['hand_pose_mano'].reshape(2, 62)
+            for side, row in enumerate(hands):
+                if row[0] not in (0, 1):
+                    raise ValueError('unexpected H2O MANO presence flag')
+                if row[0] == 1:
+                    motion[i, side] = np.r_[row[4:52], row[52:62], row[1:4]]
+                    mv[i, side] = True
+        if 'hand_pose' in lab:
+            hands = lab['hand_pose'].reshape(2, 64)
+            for side, row in enumerate(hands):
+                if row[0] not in (0, 1):
+                    raise ValueError('unexpected H2O keypoint presence flag')
+                if row[0] == 1:
+                    points[i, side] = row[1:].reshape(21, 3)
+                    pv[i, side] = True
+        for key in ('obj_pose', 'obj_pose_rt'):
+            if key in lab and lab[key][0] != 0:
+                object_ids.add(int(lab[key][0]))
+    if any('hand_pose_mano' in lab for lab in data):
+        if len(models) == 2:
+            view.tensor('mano', 'mano/axis_angle61/v1', 'annotation/mano/motion.pt', motion, mv,
+                        ['T', 'hand', 'parameter'], loc, model_profile=model_profile(models))
+        else:
+            view.absent('mano', 'Source parameters exist; --mano-dir must identify both model profiles', 'dependency_missing')
+    if any('hand_pose' in lab for lab in data):
+        view.tensor('keypoints_xyz', 'keypoints/xyz/v1', 'annotation/keypoints/xyz.pt', points, pv,
+                    ['T', 'hand', 'keypoint', 'coordinate'], loc, keypoint_schema='mano21/v1', unit='m')
+    for oid in sorted(object_ids):
+        if oid not in OBJECTS:
+            raise ValueError(f'unknown H2O object class {oid}')
+        entity = OBJECTS[oid]
+        view.object(entity, loc, category_id=oid, source_object_id=oid)
+        poses, ov = np.full((n, 4, 4), np.nan, np.float32), np.zeros(n, bool)
+        pts, kv = np.full((n, 21, 3), np.nan, np.float32), np.zeros((n, 21), bool)
+        for i, lab in enumerate(data):
+            if 'obj_pose_rt' in lab and lab['obj_pose_rt'][0] == oid:
+                poses[i] = lab['obj_pose_rt'][1:].reshape(4, 4)
+                ov[i] = True
+            if 'obj_pose' in lab and lab['obj_pose'][0] == oid:
+                pts[i] = lab['obj_pose'][1:].reshape(21, 3)
+                kv[i] = True
+        if ov.any():
+            view.tensor('pose_' + entity, 'object/rigid/v1', f'annotation/object/{entity}/pose.pt',
+                        poses, ov, ['T', 'row', 'column'], loc, convention='T_camera_from_object', unit='m')
+        if kv.any():
+            view.tensor('points_' + entity, 'object_keypoints/camera/v1',
+                        f'annotation/object/{entity}/keypoints/xyz.pt', pts, kv,
+                        ['T', 'keypoint', 'coordinate'], loc, keypoint_schema='H2O_official_object21',
+                        keypoint_names=['center'] + [f'corner_{i}' for i in range(8)] + [f'edge_midpoint_{i}' for i in range(12)],
+                        source_order='official obj_point_order.jpg: center, 8 corners, 12 edge midpoints', unit='m')
+        if oid in assets:
+            directory, obj, aliases = assets[oid]
+            asset_loc = locator(asset_archive, f'object/{entity}/{obj.name}')
+            if aliases:
+                asset_loc['texture_aliases'] = aliases
+            mesh_asset(view, entity, directory, obj, asset_loc)
+        else:
+            view.absent('mesh_' + entity, 'Source object.zip unavailable', 'not_downloaded')
+    if any('depth' in rows[f] for f in ids):
+        width, height = view.meta['camera']['image_size']
+        depth_threads = getattr(args, 'depth_threads', 1)
+        def read_depth(f):
+            item = rows[f].get('depth')
+            if item:
+                with Image.open(item[0]) as im:
+                    return np.asarray(im).astype(np.float32) / 1000.0
+            return np.full((height, width), np.nan, np.float32)
+        def depths():
+            if depth_threads == 1:
+                yield from map(read_depth, ids)
+            else:
+                with ThreadPoolExecutor(max_workers=depth_threads) as pool:
+                    # Ordered batches bound decoded arrays to depth_threads frames per view.
+                    for start in range(0, len(ids), depth_threads):
+                        yield from pool.map(read_depth, ids[start:start + depth_threads])
+        view.raster('depth', depths(), None, loc | {'source_field': 'component_sources.depth',
+            'definition': 'https://github.com/taeinkwon/h2oplayer/blob/main/utils.py',
+            'conversion': 'registered millimetre camera-Z to metres; no viewer range clipping'},
+            codec_threads=codec_threads)
+    else:
+        view.absent('depth', 'No source depth frames')
+    view.absent('segment', 'H2O provides no native segmentation; rendering deferred to v2.0')
+    view.absent('relations', 'No source contact/grasp annotation')
+    return view.finish()
+
+
+def convert_archive(archive, args, out, splits, models, assets, asset_archive):
+    selected_views = set(args.views.split(',')) if args.views else None
+    subjects = {'subject' + s.strip() for s in args.subjects.split(',')} if args.subjects else None
+    current, rows, intrinsics, done = None, {}, {}, 0
+    published_views = set()
+    with tempfile.TemporaryDirectory(prefix='h2o-source-') as wd:
+        work = Path(wd)
+        def flush():
+            nonlocal done
+            if not rows:
+                return
+            seq = '/'.join(current)
+            reports, pending = {}, {}
+            for cam, frames in sorted(rows.items()):
+                existing = out / 'data' / seq.replace('/', '_') / cam
+                if (existing / 'metainfo.json').is_file():
+                    old_frames = [json.loads(line) for line in (existing / 'camera/frames.jsonl').read_text().splitlines()]
+                    if [f['source_frame_id'] for f in old_frames] != list(range(min(frames), max(frames) + 1)):
+                        raise ValueError('existing view has different sampling; choose a new release-id')
+                    for frame in old_frames:
+                        incoming = frames.get(frame['source_frame_id'], {})
+                        previous = frame.get('component_sources', {})
+                        if set(previous) != set(incoming) or any(previous[k].get('sha256') != v[1].get('sha256') for k, v in incoming.items()):
+                            raise ValueError('duplicate H2O view differs from the published source; choose a new release-id')
+                else:
+                    pending[cam] = frames
+            if pending:
+                with sequence_dir(out, seq.replace('/', '_')) as target:
+                    for cam, frames in pending.items():
+                        reports[cam] = convert_view(target / cam, args, seq, cam, frames,
+                            intrinsics.get(cam), splits, models, assets, asset_archive)
+                build_record(out, seq.replace('/', '_'), reports)
+            for staged in list(work.iterdir()):
+                if staged.is_dir():
+                    shutil.rmtree(staged)
+                else:
+                    staged.unlink()
+            done += 1
+            print(f'{"Converted" if pending else "Verified existing"} {seq}', flush=True)
+        # Uncompressed CASA supports seeking over large frames; gzip is single-pass.
+        mode = 'r:' if archive.suffix == '.tar' else 'r|gz'
+        with tarfile.open(archive, mode) as tf:
+            for member in tf:
+                match = MEMBER.fullmatch(member.name)
+                if not member.isfile() or not match:
+                    continue
+                subj, scene, seq, cam, tail = match.groups()
+                if subjects and subj not in subjects:
+                    continue
+                key = (subj, scene, seq)
+                if key != current:
+                    flush()
+                    if args.limit and done >= args.limit:
+                        break
+                    current, rows, intrinsics = key, {}, {}
+                    published_views = {view.name for view in
+                        (out / 'data' / '_'.join(key)).glob('cam*')
+                        if (view / 'metainfo.json').is_file()}
+                if selected_views and cam not in selected_views:
+                    continue
+                frame = FRAME.fullmatch(tail)
+                if tail != 'cam_intrinsics.txt' and not frame:
+                    continue
+                if frame and args.max_frames and int(frame[2]) >= args.max_frames:
+                    continue
+                data = tf.extractfile(member).read()
+                dest = safe_member(work, member.name)
+                if cam not in published_views:
+                    dest.parent.mkdir(parents=True, exist_ok=True)
+                    dest.write_bytes(data)
+                # Published views still verify source bytes, but never read their staging paths.
+                item = (dest, locator(archive, member.name, data))
+                if tail == 'cam_intrinsics.txt':
+                    intrinsics[cam] = item
+                else:
+                    kind, f = frame[1], int(frame[2])
+                    row = rows.setdefault(cam, {}).setdefault(f, {})
+                    if kind in row:
+                        raise ValueError('duplicate H2O frame')
+                    row[kind] = item
+            else:
+                flush()
+    return done
+
+
+def main(argv=None):
+    p = parser('H2O', __doc__)
+    p.set_defaults(release_id='shared-v1.0')
+    p.add_argument('--workers', type=int, choices=range(1, 5), default=1,
+                   help='Independent subject archives in parallel, 1..4; benchmark before using >2')
+    p.add_argument('--codec-threads', type=int, choices=(1, 2, 4, 8, 10), default=4,
+                   help='FFV1 slice threads per encode; default 4, use 10 for the current full run')
+    p.add_argument('--decode-threads', type=int, choices=(1, 2, 4, 8, 10), default=1,
+                   help='FFmpeg threads for source RGB decoding and decoded-media checks')
+    p.add_argument('--depth-threads', type=int, choices=(1, 2, 4), default=1,
+                   help='Bounded ordered source depth PNG loading per worker')
+    p.add_argument('--subjects')
+    p.add_argument('--include-casa', action='store_true')
+    p.add_argument('--archive', type=Path, help='Convert one explicit source archive')
+    args = p.parse_args(argv)
+    os.environ['HAND_WAREHOUSE_FFMPEG_THREADS'] = str(args.decode_threads)
+    if args.include_casa and args.workers != 1:
+        p.error('CASA can share sequence IDs with subject archives; convert it sequentially in a separate release')
+    archives = [args.archive.expanduser()] if args.archive else sorted(
+        {a for root in roots(args.root) for a in root.glob('subject*_v1_1.tar.gz')})
+    if args.subjects and not args.archive:
+        requested = {'subject' + s.strip() for s in args.subjects.split(',')}
+        archives = [arc for arc in archives if arc.name.split('_', 1)[0] in requested]
+    if args.include_casa:
+        casa = find_file(args.root, 'h2o_CASA.tar')
+        if casa:
+            archives.append(casa)
+    if not archives:
+        p.error('no H2O archives found')
+    out = release(args, 'H2O')
+    source_mapping = next(p for p in (Path(__file__).resolve().parents[1] / 'docs/h2o-RAW_TO_SHARED.md',
+                                     Path(__file__).resolve().parents[1] / 'doc/RAW_TO_SHARED.md',
+                                     out / 'doc/RAW_TO_SHARED.md') if p.is_file())
+    mapping = out / 'doc/RAW_TO_SHARED.md'
+    if mapping.exists() and mapping.read_bytes() != source_mapping.read_bytes():
+        raise ValueError('RAW_TO_SHARED.md changed; choose a new release-id')
+    if not mapping.exists():
+        shutil.copyfile(source_mapping, mapping)
+    torch.set_num_threads(1)
+    models, splits = mano_models(args.mano_dir), load_splits(args.root)
+    with tempfile.TemporaryDirectory(prefix='h2o-assets-') as wd:
+        assets, asset_archive = load_assets(args.root, Path(wd))
+        def convert_one(archive):
+            return convert_archive(archive, args, out, splits, models, assets, asset_archive)
+        if args.workers == 1:
+            for archive in archives:
+                convert_one(archive)
+        else:
+            with ThreadPoolExecutor(max_workers=args.workers) as pool:
+                for archive, count in zip(archives, pool.map(convert_one, archives), strict=True):
+                    print(f'Archive {archive.name}: {count} sequences', flush=True)
+    return 0
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())
+
+# TODO(v2.0): 以下旧 MANO 网格和渲染分割代码仅保留备查；v1.0 不生成标注。
+# def load_mano_model(path: Path):
+#     """MPI MANO pkl → numpy 字典。shapedirs 在 pkl 内为 chumpy Select（对 (778,3,20) 基
+#     底 flat 选取前 10 个 shape 分量；idxs 已核实等价于 base[:,:,:10]）。"""
+#     with open(path, "rb") as f:
+#         d = _ManoUnpickler(f, encoding="latin1").load()
+#     sd = d["shapedirs"]
+#     base = np.asarray(sd.a.x, np.float64)                       # (778, 3, 20)
+#     shapedirs = base.reshape(-1)[np.asarray(sd.idxs)].reshape(base.shape[0], 3, -1)
+#     jreg = d["J_regressor"]
+#     jreg = np.asarray(jreg.todense()) if hasattr(jreg, "todense") else np.asarray(jreg)
+#     return {"v_template": np.asarray(d["v_template"], np.float64),
+#             "shapedirs": shapedirs,
+#             "posedirs": np.asarray(d["posedirs"], np.float64),
+#             "weights": np.asarray(d["weights"], np.float64),
+#             "J_regressor": np.asarray(jreg, np.float64),
+#             "kintree": np.asarray(d["kintree_table"], np.int64),
+#             "faces": np.asarray(d["f"], np.int64)}
+#
+#
+# def rodrigues(r):
+#     """axis-angle (N,3) → 旋转矩阵 (N,3,3)。公式 I + sinθ·K + (1−cosθ)·K²（已对 scipy 核验）。"""
+#     r = np.asarray(r, np.float64).reshape(-1, 3)
+#     th = np.linalg.norm(r, axis=1)
+#     R = np.zeros((len(r), 3, 3))
+#     nz = th > 1e-12
+#     R[~nz] = np.eye(3)
+#     k = r[nz] / th[nz, None]
+#     K = np.zeros((len(k), 3, 3))
+#     K[:, 0, 1], K[:, 0, 2], K[:, 1, 2] = -k[:, 2], k[:, 1], -k[:, 0]
+#     K[:, 1, 0], K[:, 2, 0], K[:, 2, 1] = k[:, 2], -k[:, 1], k[:, 0]
+#     s, c = np.sin(th[nz])[:, None, None], np.cos(th[nz])[:, None, None]
+#     R[nz] = np.eye(3) + s * K + (1 - c) * (K @ K)
+#     return R
+#
+#
+# def _mano_chain(model, betas, pose):
+#     """返回 (v_shaped, J, Rs, G)：G 为 16 个关节的全局变换（平移列 = 关节位置）。"""
+#     v = model["v_template"] + np.einsum("nkc,c->nk", model["shapedirs"], betas)
+#     J = model["J_regressor"] @ v
+#     Rs = rodrigues(np.asarray(pose, np.float64).reshape(16, 3))
+#     kt = model["kintree"]
+#     G = np.zeros((16, 4, 4))
+#     for j in range(16):
+#         p = kt[0, j] if j else -1
+#         loc = np.eye(4)
+#         loc[:3, :3] = Rs[j]
+#         loc[:3, 3] = J[j] - (J[p] if j else 0.0)
+#         G[j] = loc if not j else G[p] @ loc
+#     return v, J, Rs, G
+#
+#
+# def mano_forward(model, betas, pose, transl):
+#     """MANO LBS 前向（经典蒙皮 M_j = G_j·G0_j⁻¹；对 hand_pose 21 关节实证 0.0 mm）。
+#     pose: 48 axis-angle（[0:3] 全局）；返回相机系米顶点 (778,3)。"""
+#     v, J, Rs, G = _mano_chain(model, betas, pose)
+#     lrot = (Rs[1:] - np.eye(3)).ravel()
+#     v_posed = v + np.einsum("nkc,c->nk", model["posedirs"], lrot)
+#     kt = model["kintree"]
+#     G0 = np.zeros((16, 4, 4))
+#     for j in range(16):
+#         p = kt[0, j] if j else -1
+#         loc = np.eye(4)
+#         loc[:3, 3] = J[j] - (J[p] if j else 0.0)
+#         G0[j] = loc if not j else G0[p] @ loc
+#     M = np.einsum("nj,jab->nab", model["weights"], G @ np.linalg.inv(G0))
+#     out = np.einsum("nab,nb->na", M[:, :3, :3], v_posed) + M[:, :3, 3]
+#     return out + np.asarray(transl, np.float64)
+#
+#
+# def mano_chain_joints(model, betas, pose, transl):
+#     """16 个关节位置（相机系米）= 链平移列；自检用。"""
+#     _, _, _, G = _mano_chain(model, betas, pose)
+#     return G[:, :3, 3] + np.asarray(transl, np.float64)
+#
+#
+# def rasterize(verts_cam, faces, labels, W, H, fx, fy, cx, cy):
+#     z = verts_cam[:, 2]
+#     P = faces
+#     ok = (z[P] > 0.02).all(1)                 # 任一顶点在相机后 → 丢弃该三角形
+#     P, labels = P[ok], labels[ok]
+#     if len(P) == 0:
+#         return np.zeros((H, W), np.uint16), np.zeros(0, np.int64)
+#     x = verts_cam[:, 0] / z * fx + cx
+#     y = verts_cam[:, 1] / z * fy + cy
+#     xa = np.maximum(0, np.minimum(np.floor(x[P[:, 0]]), np.minimum(
+#         np.floor(x[P[:, 1]]), np.floor(x[P[:, 2]]))).astype(np.int64))
+#     xb = np.minimum(W - 1, np.maximum(np.ceil(x[P[:, 0]]), np.maximum(
+#         np.ceil(x[P[:, 1]]), np.ceil(x[P[:, 2]]))).astype(np.int64))
+#     ya = np.maximum(0, np.minimum(np.floor(y[P[:, 0]]), np.minimum(
+#         np.floor(y[P[:, 1]]), np.floor(y[P[:, 2]]))).astype(np.int64))
+#     yb = np.minimum(H - 1, np.maximum(np.ceil(y[P[:, 0]]), np.maximum(
+#         np.ceil(y[P[:, 1]]), np.ceil(y[P[:, 2]]))).astype(np.int64))
+#     xa, ya = np.minimum(xa, xb), np.minimum(ya, yb)   # 完全出画的三角形 → 空 bbox
+#     wdt, hgt = xb - xa + 1, yb - ya + 1
+#     cnt = wdt * hgt
+#     tot = int(cnt.sum())
+#     tri = np.repeat(np.arange(len(P)), cnt)
+#     k = np.arange(tot) - (np.cumsum(cnt) - cnt)[tri]
+#     px = xa[tri] + k % wdt[tri] + 0.5
+#     py = ya[tri] + k // wdt[tri] + 0.5
+#     ax, ay = x[P[tri, 0]], y[P[tri, 0]]
+#     bx, by = x[P[tri, 1]], y[P[tri, 1]]
+#     cx_, cy_ = x[P[tri, 2]], y[P[tri, 2]]
+#     area = ((x[P[:, 1]] - x[P[:, 0]]) * (y[P[:, 2]] - y[P[:, 0]])
+#             - (y[P[:, 1]] - y[P[:, 0]]) * (x[P[:, 2]] - x[P[:, 0]]))
+#     areat = area[tri]
+#     e0 = (bx - ax) * (py - ay) - (by - ay) * (px - ax)
+#     e1 = (cx_ - bx) * (py - by) - (cy_ - by) * (px - bx)
+#     e2 = (ax - cx_) * (py - cy_) - (ay - cy_) * (px - cx_)
+#     s = np.sign(areat)
+#     inside = (np.abs(areat) > 1e-9) & ((e0 * s >= 0) & (e1 * s >= 0) & (e2 * s >= 0))
+#     px, py, tri = px[inside], py[inside], tri[inside]
+#     at = areat[inside]
+#     w0, w1, w2 = e1[inside] / at, e2[inside] / at, e0[inside] / at
+#     zz = w0 * z[P[tri, 0]] + w1 * z[P[tri, 1]] + w2 * z[P[tri, 2]]
+#     pix = py.astype(np.int64) * W + px.astype(np.int64)
+#     order = np.lexsort((zz, pix))             # 同像素按深度升序 → 首条即最近
+#     pix, tri = pix[order], tri[order]
+#     first = np.unique(pix, return_index=True)[1]
+#     mask = np.zeros(H * W, np.uint16)
+#     win = tri[first]
+#     mask[pix[first]] = labels[win] + 1        # 临时标签 +1（0=背景）
+#     return mask.reshape(H, W), np.bincount(labels[win] + 1)
+#
+#
+# # ============================================================================
+# # ffmpeg 写出与核验（与 convert_dexycb.py 同构）
+# # ============================================================================
+#             verts, faces_all, labels = [], [], []
+#             for ti, (tid, kind) in enumerate(tracks):
+#                 if kind == "hand_left" and b["mano"][0] > 0.5:
+#                     verts.append(mano_forward(mano_models["left"], b["mano"][52:62],
+#                                               b["mano"][4:52], b["mano"][1:4]))
+#                     faces_all.append(mano_models["left"]["faces"])
+#                 elif kind == "hand_right" and b["mano"][62] > 0.5:
+#                     verts.append(mano_forward(mano_models["right"], b["mano"][114:124],
+#                                               b["mano"][66:114], b["mano"][63:66]))
+#                     faces_all.append(mano_models["right"]["faces"])
+#                 elif isinstance(kind, tuple) and int(b["objp"][0]) == kind[1]:
+#                     M = np.asarray(b["objrt"][1:17], np.float64).reshape(4, 4)
+#                     V0, F0 = meshes[kind[1]]
+#                     verts.append(V0 @ M[:3, :3].T + M[:3, 3])
+#                     faces_all.append(F0)
+#                 else:
+#                     continue
+#                 labels.append(np.full(len(faces_all[-1]), ti))
+#             if verts:
+#                 V = np.concatenate(verts)
+#                 offs = np.cumsum([0] + [len(v) for v in verts[:-1]])
+#                 F = np.concatenate([fa + o for fa, o in zip(faces_all, offs)])
+#                 mask, counts = rasterize(V, F, np.concatenate(labels),
+#                                          IMG_W, IMG_H, fx, fy, cx, cy)
+#             else:
+#                 mask, counts = np.zeros((IMG_H, IMG_W), np.uint16), np.zeros(0, np.int64)
+#             lut = np.zeros(n_tracks + 1, np.uint16)
+#             for ti in range(n_tracks):
+#                 present = ti + 1 < len(counts) and counts[ti + 1] > 0
+#                 if present:
+#                     presence[tracks[ti][0]][f] = True
+#                     if active_id[ti] is None:
+#                         pid = next(i for i in range(1, 65535) if i not in used)
+#                         used.add(pid)
+#                         active_id[ti], seg_start[ti] = pid, f
+#                         frame_maps[ti].append({"id": pid})
+#                     lut[ti + 1] = active_id[ti]
+#                 elif active_id[ti] is not None:
+#                     frame_maps[ti][-1]["frames"] = [seg_start[ti], f - 1]
+#                     segments[ti].append([seg_start[ti], f - 1])
+#                     used.discard(active_id[ti])
+#                     active_id[ti] = None
+#             mw.write(lut[mask])
+
+
+# TODO(v2.0): 旧物体尺寸派生代码；v1.0 仅保留源物体关键点。
+#                 pts = np.asarray(b["objp"][1:64], np.float64).reshape(21, 3)
+#                 half = np.abs((pts[1:9] - pts[0]) @ M[:3, :3]).max(0)
+#                 obj_recs.append({
+#                     "track_id": f"o{classes.index(c)}", "frame": f,
+#                     "mesh_ref": f"h2o/{OBJ_CLASSES[c]}.obj",
+#                     "pose": {"quat": rotmat_to_quat_wxyz(M[:3, :3])[0].tolist(),
+#                              "trans": M[:3, 3].tolist()},
+#                     "bbox_center3d": pts[0].tolist(),
+#                     "bbox_size": (2 * half).tolist()})

H2O/shared-v1.0/script/provenance/previous-run/script/convert_hoi4d_v1.0.py

@@ -0,0 +1,595 @@
+#!/usr/bin/env python3
+"""HOI4D native observations and annotations → CONTRACT.md v1.0."""
+from collections import Counter
+from concurrent.futures import ThreadPoolExecutor
+from contextlib import ExitStack
+from functools import lru_cache
+import gzip
+import io
+import json
+import os
+from pathlib import Path
+import pickle
+import re
+import shutil
+import shlex
+import posixpath
+import subprocess
+import tempfile
+import tarfile
+import threading
+import zipfile
+
+# Each worker owns one Python thread; prevent native libraries spawning CPU pools.
+for key in ('OMP_NUM_THREADS', 'OPENBLAS_NUM_THREADS', 'MKL_NUM_THREADS', 'NUMEXPR_NUM_THREADS'):
+    os.environ[key] = '1'
+
+import numpy as np
+from PIL import Image
+from scipy.spatial.transform import Rotation
+
+import torch
+
+from warehouse_v1_0 import (View, build_record, check_mesh, decode_args, find_file, json_write, locator,
+    mano_models, model_profile, parser, probe, release, run, safe_member, sequence_dir)
+
+CATEGORIES = {
+    1:  dict(cad="rigid/ToyCar", groups=[("main", [1], "Toy Car", 1)],
+             hands={2: "right"}, objmap={"Toycar": "main"}),
+    2:  dict(cad="rigid/Mug", groups=[("a", [1], "Mug", 2), ("b", [3], "Mug", 2)],
+             hands={2: "right"}, objmap={"Watercup": "a"}),
+    3:  dict(cad="articulated/Laptop", groups=[("main", [1, 3], "Laptop", 3)],
+             hands={2: "right"},
+             objmap={"Laptopdisplay": "main", "Laptopkeyboard": "main"}),
+    4:  dict(cad="articulated/StorageFurniture",
+             groups=[("main", [1, 3, 4], "Storage Furniture", 4),
+                     ("bottle", [5], "Bottle", 5)],
+             hands={2: "right"},
+             objmap={"Lockerbody": "main", "Lockersldingdoor": "main",
+                     "Lockerdrawer": "main", "Safebox": "main"}),
+    5:  dict(cad="rigid/Bottle", groups=[("main", [1], "Bottle", 5),
+                                         ("mug", [3], "Mug", 2)],
+             hands={2: "right"}, objmap={"bottleddrinks": "main"}),
+    6:  dict(cad="articulated/Safe", groups=[("main", [1, 3], "Safe", 6),
+                                              ("bottle", [4], "Bottle", 5)],
+             hands={2: "right"}, objmap={"Safebox": "main", "Safedoor": "main"}),
+    7:  dict(cad="rigid/Bowl", groups=[("main", [1], "Bowl", 7),
+                                       ("ball", [3], "Ball", None)],
+             hands={2: "right"}, objmap={"bowl": "main"}),
+    8:  dict(cad="articulated/Bucket",
+             groups=[("a", [1, 3], "Bucket", 8), ("b", [4, 5], "Bucket", 8)],
+             hands={2: "right", 6: "left"},
+             objmap={"bucket": "a", "Buckethandle": "a"}),
+    9:  dict(cad="articulated/Scissors", groups=[("main", [1, 3], "Scissors", 9)],
+             hands={2: "right"},
+             objmap={"scissors": "main", "scissors2": None}),   # scissors2：场景道具，无掩码
+    11: dict(cad="articulated/Pliers", groups=[("main", [1, 3], "Pliers", 11)],
+             hands={2: "right"}, objmap={"Pliersleft": "main", "Pliersright": "main"}),
+    12: dict(cad="rigid/Kettle", groups=[("main", [1], "Kettle", 12),
+                                         ("mug", [3], "Mug", 2)],
+             hands={2: "right"}, objmap={"kettle": "main"}),
+    13: dict(cad="rigid/Knife", groups=[("main", [1], "Knife", 13),
+                                        ("apple", [3], "Apple", None)],
+             hands={2: "right"}, objmap={"knife": "main"}),
+    14: dict(cad="articulated/TrashCan",
+             groups=[("main", [1, 3], "Trash Can", 14), ("bottle", [4], "Bottle", 5)],
+             hands={2: "right"}, objmap={"Dustbinbase": "main", "Dustbincover": "main"}),
+    17: dict(cad="articulated/Lamp", groups=[("main", [1, 3, 4, 5], "Lamp", 17)],
+             hands={2: "right", 6: "left"},
+             objmap={"Lampbase": "main", "Tablelampfirstjoint": "main",
+                     "Tablelampsecondjoint": "main", "Thirdjointoftablelamp": "main"}),
+    18: dict(cad="articulated/Stapler",
+             groups=[("main", [1, 3, 4], "Stapler", 18), ("paper", [5], "Paper", None)],
+             hands={2: "right", 6: "left"},       # 6=左手：官方 label.csv 缺列，实测修正
+             objmap={"Staplerbase": "main", "Staplercover": "main"}),
+    20: dict(cad="rigid/Chair", groups=[("main", [1], "Chair", 20)],
+             hands={2: "right", 3: "left"}, objmap={"chair": "main"}),
+}
+
+
+# Official HOI4D definitions/motion segmentation/label.csv; no spatial side inference.
+REGION_NAMES = {1: {1: 'Toy Car', 2: 'Right Hand'}, 2: {1: 'Mug', 2: 'Right Hand', 3: 'Mug'}, 3: {1: 'Screen', 2: 'Right Hand', 3: 'Keyboard'}, 4: {1: 'Door', 2: 'Right Hand', 3: 'Body of the Cabinet', 4: 'Drawer', 5: 'Bottle'}, 5: {1: 'Bottle', 2: 'Right Hand', 3: 'Mug'}, 6: {1: 'Body of the Safe', 2: 'Right Hand', 3: 'Door', 4: 'Bottle'}, 7: {1: 'Bowl', 2: 'Right Hand', 3: 'Ball'}, 8: {1: 'Body of the Bucket', 2: 'Right Hand', 3: 'Handle', 4: 'Body of the Bucket', 5: 'Handle', 6: 'Left Hand'}, 9: {1: 'Left Part of the Scissors', 2: 'Right Hand', 3: 'Right Part of the Scissors'}, 11: {1: 'Left Part of the Pliers', 2: 'Right Hand', 3: 'Right Part of the Pliers'}, 12: {1: 'Kettle', 2: 'Right Hand', 3: 'Mug'}, 13: {1: 'Knife', 2: 'Right Hand', 3: 'Apple'}, 14: {1: 'Body of the Trash Can', 2: 'Right Hand', 3: 'Lid', 4: 'Bottle'}, 17: {1: 'Base', 2: 'Right Hand', 3: 'Tube', 4: 'Top Link', 5: 'Bottom Link', 6: 'Left Hand'}, 18: {1: 'Base', 2: 'Right Hand', 3: 'Slot', 4: 'Lid', 5: 'Paper'}, 20: {1: 'Chair', 2: 'Right Hand', 3: 'Left Hand'}}
+
+
+class SplitFiles(io.RawIOBase):
+    """Read concatenated source gzip parts without concatenating 119 GB on disk."""
+    def __init__(self, paths):
+        self.paths = iter(paths)
+        self.current = None
+
+    def readable(self):
+        return True
+
+    def readinto(self, buffer):
+        while True:
+            if self.current is None:
+                path = next(self.paths, None)
+                if path is None:
+                    return 0
+                self.current = path.open('rb')
+            n = self.current.readinto(buffer)
+            if n:
+                return n
+            self.current.close()
+            self.current = None
+
+    def close(self):
+        if self.current is not None:
+            self.current.close()
+        super().close()
+
+
+def extract_depths(root, sequences, work):
+    directory = find_file(root, 'DEPTH')
+    if directory is None:
+        return {}, []
+    parts = sorted(directory.glob('HOI4D_depth_video.tar.gz*'), key=lambda p: int(p.name.split('.gz')[1] or 0))
+    if not parts:
+        return {}, []
+    wanted = {f'HOI4D_depth_video/{seq}/align_depth/depth_video.avi': seq for seq in sequences}
+    found = {}
+    with SplitFiles(parts) as joined, io.BufferedReader(joined) as inp, gzip.GzipFile(fileobj=inp) as gz:
+        with tarfile.open(fileobj=gz, mode='r|') as tf:
+            for member in tf:
+                if member.isfile() and member.name in wanted:
+                    seq = wanted[member.name]
+                    dest = work / (seq.replace('/', '_') + '.avi')
+                    with tf.extractfile(member) as src, dest.open('wb') as dst:
+                        shutil.copyfileobj(src, dst)
+                    found[seq] = (dest, {'files': [str(p.resolve()) for p in parts], 'member': member.name})
+                    if len(found) == len(wanted):
+                        break
+    return found, parts
+
+
+def palette_mask(data):
+    with Image.open(io.BytesIO(data)) as im:
+        arr = np.asarray(im)
+    if arr.ndim == 2:
+        return arr.astype(np.uint16)
+    if arr.shape[-1] != 3:
+        raise ValueError('unsupported HOI4D mask representation')
+    # VOC's 256-color palette interleaves eight label bits into the high RGB bits.
+    # Invert it directly instead of sorting every pixel with np.unique.
+    if np.any(arr & np.array([31, 31, 63], dtype=np.uint8)):
+        raise ValueError('source mask has colors outside official VOC palette')
+    mask = np.zeros(arr.shape[:2], dtype=np.uint16)
+    for bit in range(8):
+        mask |= ((arr[..., bit % 3] >> (7 - bit // 3)) & 1).astype(np.uint16) << bit
+    return mask
+
+
+# Four workers each reuse annotation and hand indices instead of rebuilding the ZIP directory.
+@lru_cache(maxsize=8)
+def index_frames(z):
+    index = {}
+    for name in z.namelist():
+        parent, _, filename = name.rpartition('/')
+        stem, dot, suffix = filename.rpartition('.')
+        if not dot or not stem.isdigit() or suffix not in ('png', 'json', 'pickle'):
+            continue
+        entries = index.setdefault((parent + '/', '.' + suffix), {})
+        f = int(stem)
+        if f in entries:
+            raise ValueError(f'duplicate frame number under {parent}')
+        entries[f] = name
+    return index
+
+
+def frame_members(z, prefix, suffix):
+    return index_frames(z).get((prefix, suffix), {})
+
+
+def convert_sequence(seq, args, target, rgbz, annz, hpz, calz, cadz, paths, models, depth):
+    # ponytail: cache only one sequence's encoded annotation bytes; no decoded masks or CAD cache.
+    @lru_cache(maxsize=None)
+    def read_annotation(archive, name):
+        return archive.read(name)
+
+    category = int(re.search(r'/C(\d+)/', seq)[1])
+    definition = CATEGORIES.get(category)
+    if definition is None:
+        raise ValueError(f'category mapping unavailable: {category}')
+    prefix = f'HOI4D_annotations/{seq}'
+    mask_members = frame_members(annz, prefix + '/2Dseg/mask/', '.png')
+    shifted = frame_members(annz, prefix + '/2Dseg/shift_mask/', '.png')
+    if not mask_members:
+        mask_members = shifted
+    pose_members = frame_members(annz, prefix + '/objpose/', '.json')
+    hand_members = {side: frame_members(hpz, f'Hand_pose/handpose_{side}_hand/{seq}/', '.pickle')
+                    for side in ('left', 'right')} if hpz else {'left': {}, 'right': {}}
+    with tempfile.TemporaryDirectory(prefix='hoi4d-source-') as wd:
+        work = Path(wd)
+        member = f'HOI4D_release/{seq}/align_rgb/image.mp4'
+        video = work / 'rgb.mp4'
+        with rgbz.open(member) as src, video.open('wb') as dst:
+            shutil.copyfileobj(src, dst)
+        info = probe(video)
+        if 'nb_frames' not in info:
+            info['nb_frames'] = json.loads(run(['ffprobe', '-v', 'error', '-count_frames', '-select_streams', 'v:0',
+                '-show_entries', 'stream=nb_read_frames', '-of', 'json', str(video)], stdout=subprocess.PIPE).stdout)['streams'][0]['nb_read_frames']
+        n = int(info['nb_frames'])
+        if args.max_frames:
+            n = min(n, args.max_frames)
+        if shifted and mask_members != shifted and any(f not in mask_members and f in shifted for f in range(n)):
+            raise ValueError('shift_mask has frames absent from mask; correspondence unverified')
+        frames = []
+        for f in range(n):
+            sources = {}
+            if f in mask_members:
+                sources['segment'] = locator(paths['annotations'], mask_members[f], read_annotation(annz, mask_members[f]))
+            if f in pose_members:
+                sources['object'] = locator(paths['annotations'], pose_members[f], read_annotation(annz, pose_members[f]))
+            for side in ('left', 'right'):
+                if f in hand_members[side]:
+                    sources['hand_' + side] = locator(paths['hand'], hand_members[side][f], read_annotation(hpz, hand_members[side][f]))
+            frames.append(dict(source_frame_id=f, source_locator=locator(paths['release'], member, frame=f),
+                component_sources=sources, annotation_matches={k: dict(source_frame_id=f, source_timestamp_ns=None,
+                    clock_domain=None, match_method='index_map', delta_ns=None) for k in sources}))
+        view = View(target, 'HOI4D', args.release_id, seq.replace('/', '_'), 'ego', frames)
+        view.valid_rules = {
+            'mano': 'Matched side pickle contains poseCoeff, beta, trans and corresponding model available; zero angles are not missing',
+            'keypoints': 'Matched side pickle contains kps2D; independently provided source points, no visibility gate',
+            'pose': 'Matched source object record contains rotation and center; record identity retained',
+            'dimensions': 'Matched source object record contains dimensions; size/mesh relation remains source_unspecified',
+            'segment': 'Source mask or shift_mask present with verified palette codes and entity/part mapping',
+        }
+        view.full_jpeg = getattr(args, 'jpeg_full_check', False)
+        view.meta['view_type'] = 'ego'
+        view.meta['sampling']['selection'] = {'max_frames': args.max_frames, 'source_sequence': seq}
+        view.rgb([], locator(paths['release'], member), video=video, count=n)
+        intr_name = f'camera_params/{seq.split("/")[0]}/intrin.npy'
+        if calz and intr_name in calz.namelist():
+            view.camera(np.load(io.BytesIO(calz.read(intr_name))), locator(paths['camera'], intr_name))
+        extr_name = prefix + '/3Dseg/output.log'
+        if extr_name in annz.NameToInfo:
+            lines = [line for line in annz.read(extr_name).decode().splitlines() if line.strip()]
+            if len(lines) % 5:
+                raise ValueError('invalid Open3D camera trajectory')
+            extr, valid = np.full((n, 4, 4), np.nan, np.float32), np.zeros(n, bool)
+            for i in range(0, len(lines), 5):
+                idx = int(lines[i].split()[0])
+                if idx < n:
+                    if valid[idx]:
+                        raise ValueError('duplicate trajectory index')
+                    extr[idx] = np.linalg.inv(np.loadtxt(io.StringIO('\n'.join(lines[i + 1:i + 5]))))
+                    valid[idx] = True
+            view.tensor('extrinsic', 'transform/se3/v1', 'camera/extrinsic.pt', extr, valid,
+                ['T', 'row', 'column'], locator(paths['annotations'], extr_name),
+                convention='T_camera_from_world', source_convention='Open3D LOG camera-to-world; inverted')
+        loc = {'frames_ref': 'camera/frames.jsonl', 'source_field': 'component_sources'}
+        motion, mv = np.full((n, 2, 61), np.nan, np.float32), np.zeros((n, 2), bool)
+        points, pv = np.full((n, 2, 21, 2), np.nan, np.float32), np.zeros((n, 2, 21), bool)
+        for slot, side in enumerate(('left', 'right')):
+            for f, name in hand_members[side].items():
+                if f >= n:
+                    continue
+                data = pickle.loads(read_annotation(hpz, name), encoding='latin1')
+                if all(k in data for k in ('poseCoeff', 'beta', 'trans')):
+                    motion[f, slot] = np.r_[np.asarray(data['poseCoeff']).reshape(48),
+                        np.asarray(data['beta']).reshape(10), np.asarray(data['trans']).reshape(3)]
+                    mv[f, slot] = True  # Zero angles are a valid pose, not a missing sentinel.
+                if 'kps2D' in data:
+                    points[f, slot] = np.asarray(data['kps2D']).reshape(21, 2)
+                    pv[f, slot] = True
+        if mv.any():
+            if all(side in models for slot, side in enumerate(('left', 'right')) if mv[:, slot].any()):
+                view.tensor('mano', 'mano/axis_angle61/v1', 'annotation/mano/motion.pt', motion, mv,
+                    ['T', 'hand', 'parameter'], loc, model_profile=model_profile(models))
+            else:
+                view.absent('mano', 'Source MANO exists; provide --mano-dir to fix model identity', 'dependency_missing')
+        if pv.any():
+            view.tensor('keypoints_uv', 'keypoints/uv/v1', 'annotation/keypoints/uv.pt', points, pv,
+                ['T', 'hand', 'keypoint', 'coordinate'], loc, keypoint_schema='mano21/v1', unit='pixel',
+                source_method='Upstream fitted MANO projection already present in kps2D')
+        objects = {}
+        groups = {key: (vals, name, cls) for key, vals, name, cls in definition['groups']}
+        def ensure_object(key, source_loc, category_id=None, frame=None, source_id=None):
+            if key not in objects:
+                view.object(key, source_loc, scope='frame' if frame is not None else 'sequence',
+                    scope_id=f'{seq}:{frame}' if frame is not None else None,
+                    category_id=category_id, source_object_id=source_id)
+                objects[key] = {}
+        mappings = []
+        mask_valid = [f in mask_members for f in range(n)]
+        if any(mask_valid):
+            width, height = view.meta['camera']['image_size']
+            def masks():
+                for f in range(n):
+                    labels = []
+                    if f in mask_members:
+                        mask = palette_mask(read_annotation(annz, mask_members[f]))
+                        for value in np.unique(mask):
+                            value = int(value)
+                            if value in (0, 255):
+                                if value == 255:
+                                    raise ValueError('HOI4D 255 region semantics unverified')
+                                continue
+                            if REGION_NAMES[category].get(value) in ('Left Hand', 'Right Hand'):
+                                entity, part, cls = 'hand_' + REGION_NAMES[category][value].split()[0].lower(), None, 'hand'
+                            else:
+                                group = next((k for k, (vals, _, _) in groups.items() if value in vals), None)
+                                entity = group or f'unknown_region_{value}'
+                                cls = groups[group][2] if group else None
+                                part = f'region_{value}' if group and len(groups[group][0]) > 1 else None
+                                ensure_object(entity, frames[f]['component_sources']['segment'], cls)
+                                if part:
+                                    objects[entity].setdefault('parts', {})[part] = {'part_id': part, 'parent_id': None,
+                                        'has_model': False, 'mesh_ref': None, 'source_region_code': value, 'name': REGION_NAMES[category].get(value)}
+                            labels.append(dict(mask_value=value, category_id=cls, entity_id=entity,
+                                               part_id=part, source_instance_id=value,
+                                               source_attributes={'region_name': REGION_NAMES[category].get(value)}))
+                    else:
+                        mask = np.zeros((height, width), np.uint16)
+                    mappings.append(dict(frame_index=f, labels=labels))
+                    yield mask
+            view.raster('segment', masks(), mask_valid, loc, mappings)
+            view.meta['components']['segment']['coverage'] = {'kind': 'unknown'}
+            if any(label['part_id'] for row in mappings for label in row['labels']):
+                view.meta['components']['segment']['segmentation_type'] = 'part'
+        else:
+            view.absent('segment', 'No native segmentation for this source sequence')
+        records = {}
+        for f, name in pose_members.items():
+            if f >= n:
+                continue
+            data = json.loads(read_annotation(annz, name))
+            # isEffective is not used to erase existing numerical annotations.
+            source_records = data.get('dataList', data.get('objects'))
+            if source_records is None:
+                raise ValueError('unrecognized source object annotation structure')
+            if category == 2 and source_records and (len(source_records) != 1 or
+                    source_records[0].get('id') != 1 or source_records[0].get('label') != 'Watercup'):
+                raise ValueError('C2 native pose differs from the verified one-object target; cannot map to mask region 1')
+            label_counts = Counter(record['label'] for record in source_records)
+            per_frame = set()
+            for record_index, record in enumerate(source_records):
+                label = record['label']
+                group = definition['objmap'].get(label)
+                # Pinned official ObjTracking reads dataList[0] with segmentation label 1.
+                source_primary = category == 2 and record_index == 0 and record.get('id') == 1 and group == 'a'
+                ambiguous = group is not None and sum(g[2] == groups[group][2] for g in groups.values()) > 1 and not source_primary
+                ambiguous = ambiguous or (label_counts[label] > 1 and not source_primary) or group is None
+                entity = f'pose_frame_{f}_{record_index}' if ambiguous else group
+                ensure_object(entity, locator(paths['annotations'], name, record=record_index),
+                    groups[group][2] if group else None, frame=f if ambiguous else None, source_id=record.get('id'))
+                if ambiguous:
+                    view.build['unresolved'].append({'component': entity, 'status': 'source_semantics_unverified',
+                        'reason': 'Source pose retained with frame identity; no guessed association to same-class mask instances'})
+                label_key = (entity, label)
+                if label_key in per_frame:
+                    raise ValueError(f'ambiguous repeated object label {label}; source identity must be resolved')
+                per_frame.add(label_key)
+                records.setdefault(label_key, {})[f] = (record, locator(paths['annotations'], name, record=record.get('id')))
+        if category == 2 and any(entity == 'a' for entity, _ in records):
+            view.build['validation'].append({'component': 'C2_source_pose_mask_identity',
+                'rule': 'Official ObjTracking process_data.py reads dataList[0] with semantic mask label 1',
+                'source_revision': 'afa7bf37b02b7e33cf245acd08b98c207ccc175c',
+                'source_pose_id': 1, 'mask_value': 1, 'matched_frames': sum(len(by_frame) for (entity, _), by_frame in records.items() if entity == 'a')})
+        for (entity, label), by_frame in records.items():
+            part_labels = [k[1] for k in records if k[0] == entity]
+            articulated = definition['cad'].startswith('articulated/') and entity == definition['groups'][0][0]
+            part = 'pose_' + re.sub(r'[^a-zA-Z0-9_]', '_', label) if articulated or len(part_labels) > 1 else None
+            base = f'annotation/object/{entity}'
+            if part:
+                base += '/parts/' + part
+                objects[entity].setdefault('parts', {})[part] = dict(part_id=part, parent_id=None,
+                    has_model=False, mesh_ref=None, source_label=label)
+            poses, valid = np.full((n, 4, 4), np.nan, np.float32), np.zeros(n, bool)
+            dimensions = np.full((n, 3), np.nan, np.float32)
+            dv = np.zeros(n, bool)
+            for f, (record, _) in by_frame.items():
+                if all(k in record for k in ('rotation', 'center')):
+                    poses[f] = np.eye(4)
+                    poses[f, :3, :3] = Rotation.from_euler('XYZ', [record['rotation'][k] for k in ('x', 'y', 'z')]).as_matrix()
+                    poses[f, :3, 3] = [record['center'][k] for k in ('x', 'y', 'z')]
+                    valid[f] = True
+                if 'dimensions' in record:
+                    dimensions[f] = [record['dimensions'][k] for k in ('length', 'width', 'height')]
+                    dv[f] = True
+            key = entity + ('_' + part if part else '')
+            view.tensor('pose_' + key, 'transform/se3/v1' if part else 'object/rigid/v1', base + '/pose.pt', poses, valid,
+                ['T', 'row', 'column'], loc, convention='T_camera_from_part' if part else 'T_camera_from_object',
+                unit='m', source_label=label, source_euler_order='XYZ',
+                definition='https://github.com/hoi4d/ObjTracking/blob/main/process_data.py')
+            if dv.any():
+                view.tensor('dimensions_' + key, 'object/dimensions/v1', base + '/dimensions.pt', dimensions, dv,
+                    ['T', 'dimension'], loc, dimensions_semantics='source_unspecified', unit='m', scale_applied_to_mesh=False)
+        if cadz:
+            primary = definition['groups'][0][0]
+            if category != 2 and sum(g[2] == category for g in groups.values()) > 1:
+                primary = 'source_cad_instance'
+            instance = int(re.search(r'/N(\d+)/', seq)[1])
+            mesh_prefix = f'HOI4D_CAD_Model_for_release/{definition["cad"]}/{instance:03d}'
+            candidates = [name for name in cadz.namelist() if not name.endswith('/') and
+                          (name == mesh_prefix + '.obj' or name.startswith(mesh_prefix + '/'))]
+            if candidates:
+                # Retain OBJ material/texture dependencies beside their original names.
+                queue = list(candidates)
+                for name in queue:
+                    if Path(name).suffix.lower() not in ('.obj', '.mtl'):
+                        continue
+                    for line in cadz.read(name).decode().splitlines():
+                        fields = shlex.split(line, comments=True)
+                        if not fields:
+                            continue
+                        refs = fields[1:] if fields[0] == 'mtllib' else ([fields[-1]] if
+                            fields[0].startswith('map_') or fields[0] in ('bump', 'norm', 'disp', 'decal') else [])
+                        for ref in refs:
+                            dependency = posixpath.normpath(posixpath.join(posixpath.dirname(name), ref))
+                            if dependency not in cadz.NameToInfo:
+                                raise ValueError(f'missing CAD dependency: {dependency}')
+                            if dependency not in candidates:
+                                candidates.append(dependency)
+                                queue.append(dependency)
+                ensure_object(primary, locator(paths['cad'], mesh_prefix), category)
+                asset_dir = view.path / f'annotation/object/{primary}/assets'
+                asset_refs = []
+                for name in candidates:
+                    relative = posixpath.relpath(name, mesh_prefix if definition['cad'].startswith('articulated/') else posixpath.dirname(mesh_prefix))
+                    dest = safe_member(asset_dir, relative)
+                    dest.parent.mkdir(parents=True, exist_ok=True)
+                    dest.write_bytes(cadz.read(name))
+                    ref = dest.relative_to(view.path).as_posix()
+                    if dest.suffix.lower() == '.obj':
+                        asset_key = re.sub(r'[^a-zA-Z0-9_]', '_', relative)
+                        view.component('mesh_' + primary + '_' + asset_key, 'mesh/obj/v1', ref, ['vertex'],
+                            locator(paths['cad'], name), unit='m', valid=True,
+                            scale_applied_to_mesh=False, dimensions_semantics='source_unspecified')
+                        asset_refs.append(ref)
+                        if definition['cad'].startswith('articulated/'):
+                            part_id = 'asset_' + asset_key
+                            objects[primary].setdefault('parts', {})[part_id] = dict(part_id=part_id,
+                                parent_id=None, has_model=True, mesh_ref=ref, source_asset=relative,
+                                note='Source asset; correspondence to motion part labels remains unverified')
+                for asset_ref in asset_refs:
+                    mesh_check = check_mesh(view.path / asset_ref, asset_dir)
+                    view.build['validation'].append({'component': asset_ref, **mesh_check,
+                        'unit_basis': 'official ObjTracking loads source canonical CAD without scaling in metre camera coordinates',
+                        'geometric_acceptance': 'physical dimensions-to-mesh relation remains source_unspecified'})
+                metadata_path = view.path / f'annotation/object/{primary}/metainfo.json'
+                metadata = json.loads(metadata_path.read_text())
+                metadata.update(has_model=bool(asset_refs), asset_refs=asset_refs)
+                if definition['cad'].startswith('rigid/') and len(asset_refs) == 1:
+                    metadata['mesh_ref'] = asset_refs[0]
+                json_write(metadata_path, metadata)
+                if definition['cad'].startswith('articulated/'):
+                    view.build['unresolved'].append({'component': 'object_part_asset_association',
+                        'status': 'source_semantics_unverified',
+                        'reason': 'All source assets and motion labels retained, but no guessed correspondence or joint_state'})
+        for entity, info in objects.items():
+            if 'parts' in info:
+                ref = f'annotation/object/{entity}/articulation.json'
+                json_write(view.path / ref, {'parts': list(info['parts'].values()), 'joints': [],
+                    'hierarchy_semantics': 'semantic_containment',
+                    'note': 'Unassociated source region, motion and asset nodes; original CAD trees retained in assets'})
+                view.component('articulation_' + entity, 'object/articulated/v1', ref, [], loc)
+                meta_path = view.path / f'annotation/object/{entity}/metainfo.json'
+                metadata = json.loads(meta_path.read_text())
+                metadata['articulation_ref'] = ref
+                json_write(meta_path, metadata)
+        if depth is not None:
+            dpath, dloc = depth
+            dmeta = probe(dpath)
+            if [dmeta['width'], dmeta['height']] != view.meta['camera']['image_size']:
+                raise ValueError('HOI4D depth grid differs from RGB')
+            if dmeta['pix_fmt'] != 'gray16le':
+                raise ValueError('unexpected HOI4D depth pixel format')
+            width, height = view.meta['camera']['image_size']
+            raw = work / 'depth.raw'
+            limits = ['-frames:v', str(n)] if args.max_frames else []
+            run(decode_args(dpath, 'gray16le') + limits + ['-y', str(raw)])
+            if raw.stat().st_size != n * width * height * 2:
+                raise ValueError('depth/RGB frame count or grid mismatch')
+            def depths():
+                with raw.open('rb') as inp:
+                    for _ in range(n):
+                        yield np.frombuffer(inp.read(width * height * 2), '<u2').reshape(height, width).astype(np.float32) / 1000
+            view.raster('depth', depths(), None, dloc | {'conversion': 'registered millimetre camera-Z to metres'})
+        else:
+            if find_file(args.root, 'DEPTH'):
+                raise ValueError('DEPTH source exists, but no matching depth video was found')
+            view.absent('depth', 'Depth source not downloaded', 'not_downloaded')
+        return view.finish()
+
+
+def main(argv=None):
+    p = parser('HOI4D', __doc__)
+    p.set_defaults(release_id='shared-v1.0')
+    # Four workers reached 21 observed OS threads including timing; stay below the 24-thread budget.
+    p.add_argument('--workers', type=int, choices=range(1, 5), default=1,
+                   help='Independent views in parallel, 1..4; codec/BLAS threads fixed at 1')
+    p.add_argument('--tasks', help='Comma-separated substrings of source sequence paths')
+    p.add_argument('--sessions', help='Comma-separated sessions')
+    p.add_argument('--categories', help='Comma-separated category numbers')
+    args = p.parse_args(argv)
+    if args.views and args.views != 'ego':
+        p.error('HOI4D source camera ID is ego')
+    paths = {key: find_file(args.root, name) for key, name in {
+        'release': 'HOI4D_release.zip', 'annotations': 'HOI4D_annotations.zip', 'hand': 'HOI4D_Hand_pose.zip',
+        'camera': 'camera_params.zip', 'cad': 'HOI4D_CAD_Model_for_release.zip'}.items()}
+    if not paths['release'] or not paths['annotations']:
+        p.error('HOI4D release and annotation ZIPs are required')
+    out = release(args, 'HOI4D')
+    mapping = Path(__file__).resolve().parent.parent / 'docs/hoi4d-RAW_TO_SHARED.md'
+    published_mapping = out / 'doc/RAW_TO_SHARED.md'
+    if published_mapping.exists() and published_mapping.read_bytes() != mapping.read_bytes():
+        raise ValueError('RAW_TO_SHARED.md changed; choose a new release-id')
+    if not published_mapping.exists():
+        shutil.copyfile(mapping, published_mapping)
+    torch.set_num_threads(1)
+    models = mano_models(args.mano_dir)
+    with ExitStack() as stack, tempfile.TemporaryDirectory(prefix='hoi4d-depth-') as wd:
+        archives = {key: stack.enter_context(zipfile.ZipFile(path)) if path else None for key, path in paths.items()}
+        sequences = sorted(name.removeprefix('HOI4D_release/').removesuffix('/align_rgb/image.mp4')
+                           for name in archives['release'].namelist() if name.endswith('/align_rgb/image.mp4'))
+        if args.tasks:
+            sequences = [s for s in sequences if any(token in s for token in args.tasks.split(','))]
+        if args.sessions:
+            sequences = [s for s in sequences if s.split('/')[0] in args.sessions.split(',')]
+        if args.categories:
+            sequences = [s for s in sequences if re.search(r'/C(\d+)/', s)[1] in args.categories.split(',')]
+        if args.limit:
+            sequences = sequences[:args.limit]
+        if not sequences:
+            p.error('no matching source sequences')
+        depths, _ = extract_depths(args.root, sequences, Path(wd))
+        worker = threading.local()
+        worker_stacks = []
+        def init_worker():
+            owned = ExitStack()
+            try:
+                # ZipFile.read mutates a shared seek position: never share a handle across workers.
+                worker.archives = {key: owned.enter_context(zipfile.ZipFile(path)) if path else None
+                                   for key, path in paths.items()}
+                index_frames(worker.archives['annotations'])
+                if worker.archives['hand']:
+                    index_frames(worker.archives['hand'])
+            except BaseException:
+                owned.close()
+                raise
+            worker_stacks.append(owned)
+        def convert_one(seq):
+            source = archives if args.workers == 1 else worker.archives
+            name = seq.replace('/', '_')
+            with sequence_dir(out, name) as target:
+                report = convert_sequence(seq, args, target / 'ego', source['release'], source['annotations'],
+                    source['hand'], source['camera'], source['cad'], paths, models, depths.get(seq))
+            build_record(out, name, {'ego': report})
+            return seq
+        if args.workers == 1:
+            for seq in map(convert_one, sequences):
+                print(f'Converted {seq}', flush=True)
+        else:
+            try:
+                with ThreadPoolExecutor(max_workers=args.workers, initializer=init_worker) as pool:
+                    for seq in pool.map(convert_one, sequences):
+                        print(f'Converted {seq}', flush=True)
+            finally:
+                for owned in worker_stacks:
+                    owned.close()
+    return 0
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())
+
+# TODO(v2.0): 旧 bbox IoU 实例重分配属于推断，注释保留，不在 v1.0 执行。
+#                         # 同类多实例（C8 双桶 / C2 双杯）：2dBox 可用时按 IoU 复核归属
+#                         cls = next(x[3] for x in cat["groups"] if x[0] == gk)
+#                         sibs = [g for g in cat["groups"]
+#                                 if g[0] != gk and cls is not None and g[3] == cls]
+#                         box = (d.get("2dBox") or {}).get("camera1")
+#                         box = (box[0][0], box[0][1], box[1][0], box[1][1]) \
+#                             if box and len(box) == 2 else None
+#                         if box is not None and sibs:
+#                             gvals = next(x[1] for x in cat["groups"] if x[0] == gk)
+#                             best, best_iou = gk, box_iou(box, bbox_of(mask, gvals))
+#                             for s in sibs:
+#                                 if ("obj", s[0]) not in tid_of_group:
+#                                     continue
+#                                 i = box_iou(box, bbox_of(mask, s[1]))
+#                                 if i > best_iou + 0.1 and i > 0.15:
+#                                     best, best_iou = s[0], i
+#                             if best != gk:
+#                                 gk, iou_fixups = best, iou_fixups + 1

H2O/shared-v1.0/script/provenance/previous-run/script/convert_visor_v1.0.py

@@ -0,0 +1,271 @@
+#!/usr/bin/env python3
+"""VISOR source polygons and contact → CONTRACT.md v1.0, without rasterization or side inference."""
+from collections import Counter
+import hashlib
+import json
+import re
+import shutil
+import tempfile
+from pathlib import Path
+import zipfile
+
+import numpy as np
+import torch
+
+from warehouse_v1_0 import (View, build_record, find_file, jsonl_write, locator, parser,
+                            release, roots, sequence_dir, source)
+
+FRAME = re.compile(r'frame_(\d+)')
+HAND = {300: 'left', 301: 'right'}
+GLOVE = {303: 'left', 304: 'right'}
+STATES = {'hand-not-in-contact': 'negative', 'glove-not-in-contact': 'negative',
+          'none-of-the-above': 'target_unannotated', 'inconclusive': 'unknown'}
+
+
+def hand_side(ann):
+    if ann['class_id'] in HAND:
+        return HAND[ann['class_id']]
+    if ann['class_id'] not in (60, 303, 304):
+        return None
+    raw = ann.get('on_which_hand')
+    raw = raw if isinstance(raw, list) else [raw]
+    sides = set()
+    for item in raw:
+        if item == 'left hand':
+            sides.add('left')
+        if item in ('right hand', 'rigth hand'):
+            sides.add('right')
+    if len(sides) == 1:
+        return next(iter(sides))
+    if not sides and ann['class_id'] in GLOVE:
+        return GLOVE[ann['class_id']]
+    return None
+
+
+def convert_video(video, split, annotation, ann_loc, rgb_zip, rgb_loc, mapping, args, target):
+    indexed = sorted(enumerate(annotation['video_annotations']), key=lambda row: int(FRAME.search(row[1]['image']['name'])[1]))
+    if args.max_frames:
+        indexed = indexed[:args.max_frames]
+    if len({row['image']['name'] for _, row in indexed}) != len(indexed):
+        raise ValueError('duplicate VISOR frame')
+    keys = lambda row, ann: (row['image']['subsequence'], ann['class_id'], ann['name'])
+    collisions = set()
+    for _, row in indexed:
+        collisions.update(key for key, number in Counter(keys(row, a) for a in row['annotations']).items() if number > 1)
+    frames, paths = [], []
+    with tempfile.TemporaryDirectory(prefix='visor-rgb-') as wd, zipfile.ZipFile(rgb_zip) as z:
+        work = Path(wd)
+        for original_index, row in indexed:
+            name = row['image']['name']
+            data = z.read(name)
+            path = work / Path(name).name
+            path.write_bytes(data)
+            paths.append(path)
+            loc = rgb_loc | {'nested_member': name, 'sha256': hashlib.sha256(data).hexdigest()}
+            frames.append(dict(source_frame_id=int(FRAME.search(name)[1]), source_locator=loc,
+                source_image=row['image'], original_frame_name=mapping.get(name),
+                annotation_matches={'segment': dict(source_frame_id=name, source_timestamp_ns=None,
+                    clock_domain=None, match_method='index_map', delta_ns=None,
+                    source_locator=ann_loc | {'record': f'video_annotations[{original_index}]'})}))
+        view = View(target, 'EPIC-KITCHENS-VISOR', args.release_id, video, 'ego', frames, sampling='sparse', split=split)
+        view.valid_rules = {
+            'segment': 'Selected sparse annotation frame present; all source polygons finite with at least three vertices and entity mapping valid',
+            'relations': 'At least one source in_contact_object value present; positive targets resolve in the same frame; unknown is an explicit usable status',
+        }
+        view.full_jpeg = getattr(args, 'jpeg_full_check', False)
+        view.meta.update(view_type='ego', split_scheme='VISOR_GroundTruth_SparseAnnotations', split_source_locator=ann_loc)
+        view.meta['sampling']['selection'] = {'representation': 'official sparse annotations', 'max_frames': args.max_frames}
+        view.rgb(paths, rgb_loc)
+    entities = {}
+    polygons, mappings, relations, relation_valid = [], [], [], []
+    for f, (original_index, row) in enumerate(indexed):
+        frame_loc = ann_loc | {'record': f'video_annotations[{original_index}]'}
+        id_map, labels, polys = {}, [], []
+        for code, ann in enumerate(row['annotations'], 1):
+            if code >= 65535:
+                raise ValueError('too many source instances')
+            sid = ann['id']
+            if sid in id_map:
+                raise ValueError('duplicate source instance ID within a frame')
+            side = hand_side(ann)
+            if side is not None:
+                entity = 'hand_' + side
+            else:
+                key = keys(row, ann)
+                identity = key + (sid, row['image']['name']) if key in collisions else key
+                entity = 'object_' + hashlib.sha256(json.dumps(identity).encode()).hexdigest()[:20]
+                if entity not in entities:
+                    scope = 'frame' if key in collisions else 'subsequence'
+                    scope_id = f'{video}:{row["image"]["name"]}' if scope == 'frame' else f'{video}:{key[0]}'
+                    view.object(entity, frame_loc | {'source_instance_id': sid}, scope, scope_id,
+                        category_id=ann['class_id'], name=ann['name'],
+                        association_basis='source_frame_instance' if scope == 'frame' else 'official_subsequence_entity_name')
+                    entities[entity] = True
+            id_map[sid] = entity
+            for polygon in ann.get('segments', []):
+                p = np.asarray(polygon)
+                if p.ndim != 2 or p.shape[1] != 2 or len(p) < 3 or not np.isfinite(p).all():
+                    raise ValueError(f'invalid source polygon {sid}')
+            polys.append(dict(mask_value=code, polygons=ann.get('segments', [])))
+            labels.append(dict(mask_value=code, category_id=ann['class_id'], entity_id=entity,
+                               part_id=None, source_instance_id=sid,
+                               source_attributes={k: v for k, v in ann.items() if k != 'segments'}))
+        rr = []
+        provided = False
+        for ann in row['annotations']:
+            raw = ann.get('in_contact_object')
+            values = raw if isinstance(raw, list) else [raw]
+            for value in values:
+                if value is None:
+                    continue
+                provided = True
+                status = STATES.get(value)
+                obj = None
+                if status is None:
+                    if value not in id_map:
+                        raise ValueError(f'unresolved source contact target {value} in {video} frame {f}')
+                    status, obj = 'positive', id_map[value]
+                rr.append(dict(relation_type='contact', subject_id=id_map[ann['id']], object_id=obj,
+                    status=status, source_value=value, source_locator=frame_loc | {'source_instance_id': ann['id']},
+                    provenance=source('Resolve source contact ID against instances in this frame')))
+        polygons.append(dict(frame_index=f, instances=polys))
+        mappings.append(dict(frame_index=f, labels=labels))
+        relations.append(dict(frame_index=f, relations=rr))
+        relation_valid.append(provided)
+    jsonl_write(view.path / 'annotation/segment/polygons.jsonl', polygons)
+    jsonl_write(view.path / 'annotation/segment/instances.jsonl', mappings)
+    torch.save(torch.ones(len(frames), dtype=torch.bool, device='cpu'), view.path / 'annotation/segment/mask_valid.pt')
+    view.component('segment', 'segment/instance_polygons/v1', 'annotation/segment/polygons.jsonl', ['T'], ann_loc,
+        valid_ref='annotation/segment/mask_valid.pt', valid_axes=['T'], segmentation_type='instance',
+        mapping_ref='annotation/segment/instances.jsonl', coverage={'kind': 'unknown', 'exhaustive_ref': 'annotation/segment/instances.jsonl'})
+    view.relations(relations, relation_valid, ann_loc)
+    for component in ('depth', 'mano', 'keypoints_uv', 'keypoints_xyz', 'intrinsic', 'extrinsic'):
+        view.absent(component, 'Not provided by VISOR sparse source annotations')
+    return view.finish()
+
+
+def main(argv=None):
+    p = parser('EPIC-KITCHENS-VISOR', __doc__)
+    p.add_argument('--videos', help='Comma-separated video names')
+    p.add_argument('--splits', default='train,val')
+    args = p.parse_args(argv)
+    wanted = set(args.videos.split(',')) if args.videos else None
+    splits = set(args.splits.split(','))
+    if not splits <= {'train', 'val'}:
+        p.error('only annotated train/val splits are available')
+    out = release(args, 'EPIC-KITCHENS-VISOR')
+    archive = find_file(args.root, 'EPIC-KITCHENS VISOR.zip')
+    with tempfile.TemporaryDirectory(prefix='visor-source-') as wd:
+        work = Path(wd)
+        if archive:
+            with zipfile.ZipFile(archive) as z:
+                suffix = 'GroundTruth-SparseAnnotations/annotations/'
+                members = [n for n in z.namelist() if suffix in n and n.endswith('.json')]
+                prefix = members[0].split(suffix)[0]
+                fm = json.loads(z.read(prefix + 'frame_mapping.json'))
+                done = 0
+                for member in sorted(members):
+                    split, filename = member.split(suffix)[1].split('/')
+                    video = Path(filename).stem
+                    if split not in splits or wanted and video not in wanted:
+                        continue
+                    data = z.read(member)
+                    rgb_member = prefix + f'GroundTruth-SparseAnnotations/rgb_frames/{split}/{video.split("_")[0]}/{video}.zip'
+                    rgb_zip = work / (video + '.zip')
+                    with z.open(rgb_member) as inp, rgb_zip.open('wb') as dest:
+                        shutil.copyfileobj(inp, dest)
+                    with sequence_dir(out, video) as dest:
+                        report = convert_video(video, split, json.loads(data), locator(archive, member, data),
+                            rgb_zip, locator(archive, rgb_member), fm.get(video, {}), args, dest / 'ego')
+                    build_record(out, video, {'ego': report})
+                    rgb_zip.unlink()
+                    done += 1
+                    print(f'Converted {video}', flush=True)
+                    if args.limit and done >= args.limit:
+                        break
+        else:
+            root = next((r for r in roots(args.root) if (r / 'GroundTruth-SparseAnnotations').is_dir()), None)
+            if root is None:
+                p.error('VISOR outer ZIP or extracted source directory not found')
+            fm = json.loads((root / 'frame_mapping.json').read_text()) if (root / 'frame_mapping.json').exists() else {}
+            done = 0
+            for split in sorted(splits):
+                for ann in sorted((root / 'GroundTruth-SparseAnnotations/annotations' / split).glob('*.json')):
+                    video = ann.stem
+                    if wanted and video not in wanted:
+                        continue
+                    rgb = root / f'GroundTruth-SparseAnnotations/rgb_frames/{split}/{video.split("_")[0]}/{video}.zip'
+                    with sequence_dir(out, video) as dest:
+                        report = convert_video(video, split, json.loads(ann.read_text()), locator(ann), rgb,
+                            locator(rgb), fm.get(video, {}), args, dest / 'ego')
+                    build_record(out, video, {'ego': report})
+                    done += 1
+                    if args.limit and done >= args.limit:
+                        return 0
+    if not done:
+        p.error('no matching annotated videos found')
+    return 0
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())
+
+# TODO(v2.0): 旧距离猜侧和单层掩码生成仅注释保留；v1.0 不执行。
+#                     if side is None and len(frame_hands) == 1:
+#                         # 空间门槛：脱下放在远处的手套不并入（实测双峰 0..25px / ≥44px）
+#                         hb = frame_hands[0][1]
+#                         gb = _bbox_of(polys)
+#                         if hb and gb and _bbox_dist(gb, hb) <= GLOVE60_MERGE_BBOX_DIST:
+#                             side = frame_hands[0][0]
+#                             stats["glove60_merged_into_sole_hand"] += 1
+#                     if side is None:
+#                         stats["dropped_glove60_unresolved"] += 1
+#                         id_disp[mid] = ("dropped", "glove60")
+#                         continue
+# def rasterize_and_encode(vdir, frames, frame_draw, zip_path, n, stats):
+#     """流式读 zip 内 jpg + 光栅化多边形 → rgb.mkv / mask.mkv；返回 (w, h)。"""
+#     from PIL import Image, ImageDraw
+#
+#     with zipfile.ZipFile(zip_path) as zf:
+#         member_of = {}
+#         for nm in zf.namelist():
+#             member_of[os.path.basename(nm)] = nm
+#         names = [im["name"] for im in frames]
+#         missing = [nm for nm in names if nm not in member_of]
+#         if missing:
+#             raise RuntimeError(f"{len(missing)} annotated frames missing in zip, "
+#                                f"e.g. {missing[:3]}")
+#         stats["zip_extra_members"] = len(member_of) - len(names)
+#
+#         # 首帧定分辨率（官方 1920×1080，实测每帧校验）
+#         img0 = Image.open(io.BytesIO(zf.read(member_of[names[0]]))).convert("RGB")
+#         w, h = img0.size
+#
+#         bin_img = Image.new("L", (w, h), 0)
+#         bin_draw = ImageDraw.Draw(bin_img)
+#         canvas = np.zeros((h, w), np.uint16)
+#         with VideoPipe(vdir / "mask.mkv", w, h, FPS, "mask") as mw, \
+#                 VideoPipe(vdir / "rgb.mkv", w, h, FPS, "rgb") as rw:
+#             for f in range(n):
+#                 canvas[:] = 0
+#                 for pid, polys in frame_draw[f]:
+#                     bin_draw.rectangle((0, 0, w, h), fill=0)
+#                     for poly in polys:
+#                         bin_draw.polygon([(float(x), float(y)) for x, y in poly], fill=255)
+#                     arr = np.asarray(bin_img)
+#                     canvas[arr > 0] = pid          # 后画覆盖先画：物体→手（modal）
+#                 mw.write(canvas)
+#                 img = img0 if f == 0 else Image.open(
+#                     io.BytesIO(zf.read(member_of[names[f]]))).convert("RGB")
+#                 if img.size != (w, h):
+#                     raise RuntimeError(f"frame {names[f]}: size {img.size} != {(w, h)}")
+#                 rw.write(np.asarray(img))
+#             n_mask, n_rgb = mw.n, rw.n
+#     if (n_rgb, n_mask) != (n, n):
+#         raise RuntimeError(f"wrote rgb={n_rgb} mask={n_mask} != {n}")
+#     got = (count_frames(vdir / "rgb.mkv"), count_frames(vdir / "mask.mkv"))
+#     if got != (n, n):
+#         raise RuntimeError(f"encoded rgb={got[0]} mask={got[1]} != {n} (decode verify)")
+#     return w, h
+#
+#

H2O/shared-v1.0/script/provenance/previous-run/script/provenance/operations/h2o-benchmark-decode.py

@@ -0,0 +1,74 @@
+"""Benchmark full FFV1 decodes, PNG RGB decodes, and source depth PNG loading; preserve every byte."""
+import hashlib
+import json
+import os
+from pathlib import Path
+import statistics
+import subprocess
+import time
+from concurrent.futures import ThreadPoolExecutor
+import numpy as np
+from PIL import Image
+
+old = Path('/data/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-w4-c10')
+bench = Path('/data/xuzhefeng/Datasets/uni-hoi-h2o-decode-bench-20260930')
+bench.mkdir(exist_ok=True)
+media = next((old / 'H2O/shared-v1.0/data').glob('subject4_h1_0*/cam0/observation/rgb.mkv')).parent
+inputs = bench / 'inputs'
+inputs.mkdir(exist_ok=True)
+for name in ('rgb.mkv', 'depth.mkv', 'depth_valid.mkv'):
+    dest = inputs / name
+    if not dest.exists():
+        os.link(media / name, dest)
+source = next((old / 'scratch').glob('h2o-source-*/subject4/h1/0/cam0/rgb')).parent
+for kind in ('rgb', 'depth'):
+    dest = inputs / kind
+    dest.mkdir(exist_ok=True)
+    for frame in sorted((source / kind).glob('*.png'))[:100]:
+        if not (dest / frame.name).exists():
+            os.link(frame, dest / frame.name)
+records = []
+expected = {}
+def record(stage, threads, digest, size, seconds):
+    assert expected.setdefault(stage, (digest, size)) == (digest, size), stage
+    row = dict(stage=stage, threads=threads, seconds=seconds, sha256=digest, bytes=size)
+    records.append(row)
+    print(json.dumps(row), flush=True)
+    (bench / 'report.json').write_text(json.dumps(records, indent=2) + '\n')
+def decode(stage, arguments, threads, pix_fmt):
+    cmd = ['ffmpeg','-v','error','-nostdin','-filter_threads',str(threads),'-threads',str(threads),
+           '-noautorotate',*arguments,'-map','0:v:0','-vsync','0','-threads',str(threads),
+           '-f','rawvideo','-pix_fmt',pix_fmt,'pipe:1']
+    digest = hashlib.sha256()
+    size = 0
+    start = time.monotonic()
+    with subprocess.Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE) as proc:
+        while chunk := proc.stdout.read(1024*1024):
+            digest.update(chunk)
+            size += len(chunk)
+        stderr = proc.stderr.read()
+        assert proc.wait() == 0, stderr.decode()
+    record(stage, threads, digest.hexdigest(), size, time.monotonic()-start)
+for threads in (1,10,10,1):
+    for name, pix in (('rgb','rgb24'),('depth','gray16le'),('depth_valid','gray')):
+        decode('ffv1_'+name, ['-i',str(inputs/(name+'.mkv'))], threads, pix)
+    decode('source_rgb_png_100', ['-framerate','30','-i',str(inputs/'rgb/%06d.png')], threads, 'rgb24')
+    files = sorted((inputs/'depth').glob('*.png'))
+    def read(path):
+        with Image.open(path) as im:
+            return np.asarray(im).astype(np.float32)/1000.0
+    for depth_threads in (1,2,4):
+        digest = hashlib.sha256()
+        start = time.monotonic()
+        size = 0
+        with ThreadPoolExecutor(max_workers=depth_threads) as pool:
+            # Bound decoded arrays to one small batch, as the production candidate would.
+            for begin in range(0,len(files),depth_threads):
+                for arr in pool.map(read,files[begin:begin+depth_threads]):
+                    digest.update(arr.tobytes())
+                    size += arr.nbytes
+        record('source_depth_png_100', depth_threads, digest.hexdigest(), size, time.monotonic()-start)
+print('MEDIANS',flush=True)
+for stage in expected:
+    for threads in sorted({r['threads'] for r in records if r['stage']==stage}):
+        print(stage,threads,statistics.median(r['seconds'] for r in records if r['stage']==stage and r['threads']==threads),flush=True)

H2O/shared-v1.0/script/provenance/previous-run/script/provenance/operations/h2o-prepare-d10.py

@@ -0,0 +1,55 @@
+"""Reuse only atomically published sequences; preserve previous runtime provenance."""
+import datetime
+import json
+import os
+from pathlib import Path
+import shutil
+
+old = Path('/data/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-w4-c10')
+new = Path('/data/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-w4-c10-d10')
+snapshot = Path('/tmp/h2o-w4-c10-d10-recovery')
+new.mkdir()
+shutil.copytree(snapshot, new / 'frozen-code')
+release = new / 'H2O/shared-v1.0'
+for directory in ('data', 'doc/validation', 'script/provenance/previous-run', 'script/provenance/operations'):
+    (release / directory).mkdir(parents=True, exist_ok=True)
+history = release / 'script/provenance/previous-run'
+shutil.copytree(old / 'H2O/shared-v1.0/script', history / 'script')
+shutil.copytree(old / 'frozen-code/docs', history / 'docs')
+for name in ('full.log', 'full.status', 'run-full.sh'):
+    shutil.copy2(old / name, history / name)
+records = []
+for record in sorted((old / 'H2O/shared-v1.0/doc').glob('*.build.json')):
+    build = json.loads(record.read_text())
+    seq = build['sequence_id']
+    assert Path(seq).name == seq and seq not in ('.', '..')
+    assert set(build['views']) == {f'cam{i}' for i in range(5)}, seq
+    data = old / 'H2O/shared-v1.0/data' / seq
+    frames = {}
+    for cam in build['views']:
+        json.loads((data / cam / 'metainfo.json').read_text())
+        frames[cam] = sum(1 for _ in (data / cam / 'camera/frames.jsonl').open())
+        assert frames[cam] > 0
+    shutil.copytree(data, release / 'data' / seq, copy_function=os.link)
+    shutil.copy2(record, release / 'doc' / record.name)
+    records.append({'sequence_id': seq, 'frames_by_camera': frames,
+                    'original_script_sha256': build['scripts']})
+manifest = {'time': datetime.datetime.now().astimezone().isoformat(),
+            'source_run': str(old), 'destination_run': str(new),
+            'configuration': {'workers': 4, 'codec_threads': 10, 'decode_threads': 10,
+                              'depth_threads': 4, 'cpu_limit': None},
+            'reuse': 'Hard-linked immutable completed views; source timeline/SHA256 still rechecked without staging published source frames',
+            'excluded': 'Unpublished .tmp directories; historical 62-sequence output',
+            'original_runtime_provenance': 'script/provenance/previous-run', 'sequences': records}
+(release / 'doc/RESTART_PROVENANCE.json').write_text(json.dumps(manifest, indent=2) + '\n')
+operations = release / 'script/provenance/operations'
+for name in ('h2o-prepare-d10.py', 'h2o-switch-d10.sh', 'h2o-benchmark-decode.py'):
+    shutil.copy2(Path('/tmp') / name, operations / name)
+for src, dest in (('/tmp/h2o-run-d10.sh', new / 'run-full.sh'),
+                  ('/tmp/h2o-run-d10.sh', release / 'script/run-full.sh'),
+                  ('/tmp/h2o-decode-bench.log', release / 'doc/validation/DECODE_THREAD_BENCHMARK.log'),
+                  ('/data/xuzhefeng/Datasets/uni-hoi-h2o-decode-bench-20260930/report.json', release / 'doc/validation/DECODE_THREAD_BENCHMARK.json'),
+                  ('/tmp/h2o-d10-final-tests.log', operations / 'worker4-encode10-decode10-depth4-regression.log'),
+                  ('/tmp/h2o-io-sample.log', release / 'doc/validation/IO_SAMPLE.log')):
+    shutil.copy2(src, dest)
+print(f'REUSED_SEQUENCES={len(records)} REUSED_VIEWS={sum(len(r["frames_by_camera"]) for r in records)}', flush=True)

H2O/shared-v1.0/script/provenance/previous-run/script/provenance/operations/h2o-switch-d10.sh

@@ -0,0 +1,16 @@
+#!/usr/bin/env bash
+set -euo pipefail
+cd /home/xuzhefeng/Projects/uni-hoi-dataset
+test "$(cat /tmp/h2o-d10-final-tests.exit)" = 0
+test "$(cat /tmp/h2o-decode-bench.exit)" = 0
+test ! -e /data/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-w4-c10-d10
+old=/data/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-w4-c10
+# Let the already nearly complete fifth camera publish, instead of discarding four completed cameras.
+while test ! -s "$old/H2O/shared-v1.0/doc/subject4_h1_0.build.json"; do
+    test "$(systemctl --user show --no-pager h2o-full-w4c10-r2-20260930.scope -p ActiveState --value)" = active
+    sleep 20
+done
+printf '\nRECONFIGURE=%s REASON=decoder10_depth4_no_resume_png_writes\n' "$(date -Is)" >> "$old/full.status"
+systemctl --user stop h2o-full-w4c10-r2-20260930.scope
+.venv/bin/python /tmp/h2o-prepare-d10.py
+systemd-run --user --scope --unit=h2o-full-d10-20260930 -p CPUAccounting=yes /bin/bash /data/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-w4-c10-d10/run-full.sh

H2O/shared-v1.0/script/provenance/previous-run/script/provenance/operations/worker4-encode10-decode10-depth4-regression.log

@@ -0,0 +1,37 @@
+Converted subject1/h1/0
+Converted subject2/h1/0
+Converted subject3/h1/0
+Converted subject4/h1/0
+Converted subject2/h1/0
+Converted subject1/h1/0
+Archive subject1_v1_1.tar.gz: 1 sequences
+Archive subject2_v1_1.tar.gz: 1 sequences
+Converted subject4/h1/0
+Converted subject3/h1/0
+Archive subject3_v1_1.tar.gz: 1 sequences
+Archive subject4_v1_1.tar.gz: 1 sequences
+Converted subject1/h1/0
+Archive subject1_v1_1.tar.gz: 1 sequences
+Converted subject2/h1/0
+Archive subject2_v1_1.tar.gz: 1 sequences
+Converted subject4/h1/0
+Converted subject3/h1/0
+Archive subject3_v1_1.tar.gz: 1 sequences
+Archive subject4_v1_1.tar.gz: 1 sequences
+Verified existing subject2/h1/0
+Verified existing subject1/h1/0
+Verified existing subject4/h1/0
+Archive subject1_v1_1.tar.gz: 1 sequences
+Archive subject2_v1_1.tar.gz: 1 sequences
+Verified existing subject3/h1/0
+Archive subject3_v1_1.tar.gz: 1 sequences
+Archive subject4_v1_1.tar.gz: 1 sequences
+Converted subject2/h1/0
+Converted subject1/h1/0
+Archive subject1_v1_1.tar.gz: 1 sequences
+Archive subject2_v1_1.tar.gz: 1 sequences
+Converted subject3/h1/0
+Archive subject3_v1_1.tar.gz: 1 sequences
+Converted subject4/h1/0
+Archive subject4_v1_1.tar.gz: 1 sequences
+PASS H2O four archive workers, codec/decode 1/4/10, ordered depth 1/2/4 and missing-depth preservation, no resume PNG writes, RGB FFV1-only, depth and persistent QC

H2O/shared-v1.0/script/provenance/previous-run/script/provenance/operations/workspace-regression.log

@@ -0,0 +1,37 @@
+Converted subject1/h1/0
+Converted subject2/h1/0
+Converted subject3/h1/0
+Converted subject4/h1/0
+Converted subject1/h1/0
+Archive subject1_v1_1.tar.gz: 1 sequences
+Converted subject2/h1/0
+Archive subject2_v1_1.tar.gz: 1 sequences
+Converted subject3/h1/0
+Archive subject3_v1_1.tar.gz: 1 sequences
+Converted subject4/h1/0
+Archive subject4_v1_1.tar.gz: 1 sequences
+Converted subject1/h1/0
+Archive subject1_v1_1.tar.gz: 1 sequences
+Converted subject3/h1/0
+Converted subject2/h1/0
+Archive subject2_v1_1.tar.gz: 1 sequences
+Archive subject3_v1_1.tar.gz: 1 sequences
+Converted subject4/h1/0
+Archive subject4_v1_1.tar.gz: 1 sequences
+Verified existing subject4/h1/0
+Verified existing subject3/h1/0
+Verified existing subject1/h1/0
+Verified existing subject2/h1/0
+Archive subject1_v1_1.tar.gz: 1 sequences
+Archive subject2_v1_1.tar.gz: 1 sequences
+Archive subject3_v1_1.tar.gz: 1 sequences
+Archive subject4_v1_1.tar.gz: 1 sequences
+Converted subject1/h1/0
+Archive subject1_v1_1.tar.gz: 1 sequences
+Converted subject2/h1/0
+Archive subject2_v1_1.tar.gz: 1 sequences
+Converted subject3/h1/0
+Archive subject3_v1_1.tar.gz: 1 sequences
+Converted subject4/h1/0
+Archive subject4_v1_1.tar.gz: 1 sequences
+PASS H2O four archive workers, codec/decode 1/4/10, ordered depth 1/2/4 and missing-depth preservation, no resume PNG writes, RGB FFV1-only, depth and persistent QC

H2O/shared-v1.0/script/provenance/previous-run/script/provenance/previous-run/docs/CONTRACT.md

@@ -0,0 +1,457 @@
+# 手物数据仓库统一约定 v1.0
+
+格式版本：`hand-warehouse/1.0`。存储约定：`lossless-video-2026-09-18`。
+
+范围限于手物交互重建。按本任务纳入的源模态转换，不要求每个数据集填满所有组件。缺失模态、尚未下载、语义未核实分别报告，不能互相替代。
+
+本文定义各字段的含义与使用规则，机器可读注册表见 [schema.json](schema.json)。两者互相补充；如有冲突，应先核实并统一定义。
+
+## 1. 目录
+
+```text
+hand-datasets/
+  <dataset_name>/
+    raw/
+      doc/                         # 官方说明、来源与校验记录
+      script/                      # 官方下载脚本、必要包装脚本
+      data/                        # 保留原始包，不原地解压或改写
+    <dataset_version>/
+      doc/                         # 发布说明、schema、来源清单
+      script/                      # 转换、校验脚本及运行配置
+      data/
+        <sequence_id>/
+          <camera_id>/
+            metainfo.json
+            observation/
+              rgb.mkv
+              rgb_valid.pt
+              depth.mkv           # float32 深度的可逆视频编码
+              depth_valid.mkv     # 逐像素 bool 的无损视频编码
+            camera/
+              frames.jsonl
+              intrinsic.pt
+              intrinsic_valid.pt
+              extrinsic.pt
+              extrinsic_valid.pt
+            annotation/
+              mano/
+                motion.pt
+                motion_valid.pt
+                visibility.pt
+              keypoints/
+                uv.pt
+                uv_valid.pt
+                xyz.pt
+                xyz_valid.pt
+                visibility.pt    # 源数据有时提供
+              object/
+                <object_id>/
+                  assets/          # 该物体的网格、材质、纹理及部件资产
+                    <mesh_id>.obj
+                  pose.pt
+                  pose_valid.pt
+                  visibility.pt
+                  metainfo.json    # 物体身份、类别、mesh、尺寸与部件引用
+                  articulation.json
+                  keypoints/
+                  vertex_part_ids.pt
+                  joint_state.pt
+                  joint_state_valid.pt
+                  parts/<part_id>/pose.pt
+                  parts/<part_id>/pose_valid.pt
+              segment/
+                mask.mkv
+                mask_valid.pt
+                instances.jsonl
+              contact/
+                relations.jsonl
+                <subject_id>__<object_id>/  # 对应实体间的距离/接触区域
+```
+
+目录展示可能存在的组件，不要求填满。具体文件、字段和缺省规则分别见后续对应小节。原始包的解压发生在临时工作区，raw 内原始包保持不变。
+
+## 2. 相机视图的 metainfo.json
+
+### 2.1 顶层字段与发布信息
+
+每个相机视图必须有 `metainfo.json`。以下字段除另行说明外均为必填：
+
+| 字段 | 填写内容 |
+|---|---|
+| schema_version | 结构版本，本版为 `hand-warehouse/1.0` |
+| dataset_id | 数据集的逻辑标识 |
+| release_id | 本次数据发布版本，对应目录中的 dataset_version |
+| sequence_id | 所属序列或独立图片样本的标识 |
+| camera_id | 当前相机视图标识 |
+| sample_type | `image` 或 `video`，规则见 2.2 |
+| num_frames | 当前视图帧数 T，所有随帧组件按同一时间轴组织 |
+| hand_slots | 固定为 `["left", "right"]`；解剖学左右的使用规则见 5.1 |
+| sampling | 采样类型、源 FPS、帧索引引用及选择策略，见 2.2 |
+| camera | 当前相机模型与图像尺寸，见 2.3 |
+| components | 实际提供的组件及其文件、schema、有效性和来源，见 2.5 |
+| split | `train/val/test/unknown`，划分协议见 2.4 |
+| view_type | 可选，`ego/exo/unknown` |
+
+`dataset_version` 是数据发布版本；`schema_version` 是结构版本，两者独立。原始包校验摘要、官方来源和转换配置应记录在发布清单中。解压发生在临时工作区，原始包保持不变。
+
+### 2.2 图片、视频与 sampling
+
+`sample_type=image`：允许单图 T=1、`sampling.type=single`，也允许将独立图片批量装入一个视频容器，T>1 时 `sampling.type=independent`、`source_fps=null`。分片大小按数据集选择；FreiHAND 默认每片最多 4096 张，训练与评估分开，尾片不补帧。该数量不是其他数据集的强制要求；同批分辨率、相机模型类型及划分协议一致。每行 frames 必须有稳定且唯一的 `sample_id`、原始文件定位；内参、手形状等仍逐帧保留，不能因打包而共享错误的标定或身份。`sequence_id` 在这里仅是批次标识，不代表真实序列；不得按相邻帧推断跟踪、同步或接触延续。源 camera_id、image_variant、split 等随样本变化时放在 frames 的对应字段中。真实采集时间未知则 timestamp_ns=null。合成背景变体不视为真实多相机。
+
+`sample_type=video`：`sampling.type=dense/sparse`；`source_fps` 是真实采集 FPS，未知为 null。非连续取样必须保留源帧号和已知源时间，不可用采样后索引除以 FPS 冒充源时间。可另记采样策略与步长。
+
+shared 可以仅存所选 RGB 帧对应的数据。OakInk2 图像/动捕不同采样率时用 annotation_matches 明确映射；不要求将未采样动捕点硬塞进 T。多视角身份和源同步关系要保留；选择了哪些帧/相机及原因写在 sampling 中。
+
+`sampling.frames_ref` 指向 `camera/frames.jsonl`。帧号、逐帧时间、同步与标注匹配记录的结构见 4.1；MKV 封装时间的约定见 3.1。
+
+### 2.3 camera 模型声明
+
+`camera` 至少包含 `model`、`schema_id`、`image_size:[W,H]`。模型未知使用 `unknown` 和 `null`，不猜成针孔。实际内参、畸变及外参数组放在 `camera/`，各模型的固定结构见第 4 节。
+
+### 2.4 split 与划分协议
+
+保留上游划分，不将同一采集序列的相邻帧随意拆到不同集合。
+
+有官方划分时必填 split_scheme 和 split_source_locator，split_scheme 区分如 Dex-YCB s0/s1、ARCTIC P1/P2 等选定协议；不知道时 split=unknown 并给理由，不凭空分配。每发布可选择一种协议；存在逐帧差异时 frames 中的 split 覆盖视图默认值，并登记其来源。
+
+### 2.5 components 组件登记
+
+组件表的键是本发布内唯一名称，例如 `rgb`、`depth`、`mano`、`keypoints_uv`。每个已提供组件声明以下信息；模态特有参数在第 3–5 节对应小节中定义。
+
+| 字段 | 约定 |
+|---|---|
+| available | 本发布是否提供该组件；缺省也表示未提供 |
+| schema_id | 已登记的具体存储与语义类型，不得临时创造未登记名称 |
+| data_ref | 数据文件路径，相对当前相机视图根目录，使用 POSIX 路径 |
+| axes | 数组各维度的含义 |
+| valid_ref / valid_axes | 有效性文件及其维度；语义由所属组件规定 |
+| storage / valid_storage | 视频数据／视频有效性的存储配置；包含已登记 profile、实际 codec、pixel_format、encoded_size，见 3.1；非视频组件按原数组规则 |
+| provenance | source / format_conversion / derived，以及转换或生成方法 |
+| source_locator | 原文件、压缩包成员和记录／帧的定位信息 |
+
+不存在的组件允许缺省或写作 `{ "available": false, "reason": "..." }`，不用引用不存在的文件。静态网格可声明标量有效性；具体见 5.4。观测的 valid 见第 3 节，标注的 valid 与 visibility 区分见 5.1。
+
+`provenance.kind` 使用 source/format_conversion/derived，`method` 说明转换或生成方法；发布清单补充脚本版本、参数、输入摘要。`source_locator` 定位到原始文件/压缩包成员和记录、帧号；允许引用有完整溯源清单的前一发布，但链条最终应能回到 raw。不得仅写一个无法定位记录的数据集名字。
+
+`.pt` 仅保存 CPU Tensor，不保存 Python 自定义对象；浮点默认 float32，有效性为 bool，浮点无效位置填 NaN。所有随帧数组保留首维 T，单帧也不 squeeze；静态资产按其 schema 单独声明。单位、坐标及固定形状见各组件小节。
+
+机器注册表 `schema.json` 中，`components` 固定解码后的数据类型、形状、单位和语义；`storage_profiles` 固定视频内的编码／还原方式。不得临时创造未登记的 schema_id 或 storage.profile。同一语义允许选择其登记的不同无损存储 profile。
+
+## 3. observation 观测数据
+
+### 3.1 rgb.mkv 与 rgb_valid.pt
+
+RGB、深度和栅格 mask 在 shared 中统一使用 MKV 视频容器，不展开为逐帧图片或图像 tar。MANO、关键点、位姿等数值标注仍用数组；分割的重叠多边形仍用 JSONL，不为了视频化而有损栅格化。单图可为一帧视频，独立图批量封装规则见 2.2。
+
+**无损指共享层不增加信息损失。** 源 JPEG／H.264 等本身可能已是有损编码，原样保留并不恢复已丢失信息；不得重新以有损 JPEG、H.264、H.265、AV1 参数编码。源视频／图片的色彩、位深、分辨率和几何关系不得被静默改变。RGB schema `rgb/uint8/v1` 的统一读取结果是 `[T,H,W,3] uint8`、RGB 通道，有效性仍为 `rgb_valid.pt [T] bool`。超出该 schema 的高位深或 HDR 源必须报告并扩展定义，不能降到 8 位冒充无损。
+
+#### 3.1.1 允许的 RGB 存储 profile
+
+| storage.profile | 编码方式 | 适用规则 |
+|---|---|---|
+| mkv/mjpeg_copy/v1 | 源 JPEG 压缩帧直接封装成 MJPEG/MKV | 不解码重编码；保持每张 JPEG 码流字节；普通 8 位 JPEG 且 MKV/解码器支持时使用 |
+| mkv/source_copy/v1 | 已有视频的 RGB 对应码流直接封装为 MKV | 仅源编码能被 MKV 和选定解码器支持时使用，不伪造帧；需维持正确的呈现顺序与依赖帧 |
+| mkv/h264_rgb_lossless/v1 | libx264rgb 的真正无损模式 | 必须 RGB 路径、CRF=0，并通过全帧 RGB 字节比较；不能替换成默认 libx264 YUV420 路径 |
+| mkv/ffv1_rgb/v1 | FFV1 的 RGB 可逆编码路径 | 作为通用候选；解码回 RGB 后必须逐字节相同，不做色度降采样 |
+
+优先尝试源码流直接封装；不能封装或需要选择稀疏帧时，比较 FFV1 与无损 RGB H.264。对完整待发布片段，只在通过无损验收的候选中选择文件较小者，不假设某个编码器总最小；编码时间和随机读取代价记入构建记录。中间候选不作为 shared 多份常驻。H.264 跨帧压缩允许较长 GOP，必须记录 GOP／关键帧间隔，随机取帧需从前一关键帧正确解码，不能把 seek 返回的近似帧当目标帧。
+
+普通源视频可以保留它已有的有损码流，但不允许新增有损重编码。源是 VRS 等记录容器时，先查明图像流实际编码：已有兼容 JPEG／视频包可直接提取封装，否则解码后使用真正无损编码；不把整个 VRS 改后缀当 MKV。PNG 不要求直接塞进 MKV，可选上表两种无损重编码方式。
+
+#### 3.1.2 视频配置与校验
+
+每个视频组件在 metainfo 中登记 `storage`，至少包含：
+
+| 字段 | 内容 |
+|---|---|
+| profile | 上述登记的存储 profile；语义仍由组件 schema_id 指定 |
+| codec | 文件中实际 codec，例如 mjpeg、h264、ffv1；source_copy 必须写实际 codec |
+| pixel_format | 实际编码像素格式；不得仅凭编码器参数猜测 |
+| encoded_size | 视频物理像素尺寸 `[width,height]`；深度可能不同于逻辑尺寸，见 3.2 |
+| decoder | 解码器名称、版本，以及颜色矩阵／范围／旋转处理等实际像素解释配置；未知的源色彩声明 unknown，不擅自称为 sRGB |
+| encoder_parameters | 重编码时的编码器、版本、无损模式和 GOP 等参数；直接封装时记录 stream_copy |
+| source_lossy | 源是否已有损：true / false / unknown；不是共享层再有损编码的许可 |
+
+无损重新编码的 RGB 必须使用同一固定解码配置，比较**全部帧**的统一 RGB 字节、数量及顺序。MJPEG-copy 不重新编码，生产校验采用 `jpeg-packets-all-rgb-sample16/v1`：从 MKV 逐包取回，对全部源 JPEG 的 SHA256、数量、顺序和图像尺寸进行精确核对；每视图另用同一固定解码配置比较最多 16 帧的源与目标 RGB 字节（含首尾与四分位位置，其余由固定种子抽取）。记录策略、实际索引、种子与结果；不足 16 帧时全查。试点与最终完整序列随机抽查，对选中视图的全部帧复核解码 RGB。历史已做全帧像素检查的证明继续保留。
+
+逐包一致与像素抽查的范围必须分别声明，不能将抽查写成全量像素检查。不能用 PSNR 高、视觉相似、编码器参数称为 lossless 等替代精确检查。同时保留源编码的色彩标记和必要元数据；原图有 EXIF/色彩/方向等元数据时保留 raw 和必要来源信息；不能静默自动旋转图像而不更新像素对应与标定。以上加速策略仅适用于原 JPEG 码流直接封装，不适用于深度、标签或重新编码的 RGB。
+
+所有共享视频的呈现帧顺序与统一 frame_index 一一对应，每个栅格组件恰好 T 帧。MKV 的 `storage_fps` 只服务封装，声明 `timebase_role=index_only`；source_copy 可以保留源 PTS 以避免改坏码流，但真实时间与标注对齐仍以 frames 为准。未知真实时间保留 null。缺失某帧组件时保留索引位置并使用该组件的无效占位／valid 标识，不删帧；不能以重复相邻图像冒充已知观测。
+
+### 3.2 depth.mkv 与 depth_valid.mkv
+
+深度语义仍是 `depth/z_float32/v1`：统一读取结果 `[T,H,W] float32`，米制相机 Z 分量，`depth_type=z`。有效深度有限且大于 0，无效位置为 NaN。`depth_origin=measured/rendered/unknown` 必填，来源不明时 unknown。默认与 RGB 共用像素网格及相机参考系，不处理未注册 RGB-D。射线距离必须依据源定义转换成 z，不能只改声明。
+
+#### 3.2.1 float32 的可逆视频编码
+
+采用 `mkv/depth_float32_bits/v1`：**不是把深度量化成 uint16**，而是将规范化后每个 IEEE-754 float32 的完整 32 位拆成两个 uint16，使用 FFV1 gray16le 保存。视频物理尺寸为 `[2W,H]`：每行前 W 列存原 float32 位串的低 16 位，后 W 列存高 16 位。读取时先合并为 uint32 位串，再按 float32 重新解释，不能数值 cast 成 float32。
+
+```text
+u = reinterpret_uint32(depth_float32)
+encoded[:, :, :W] = uint16(u & 65535)
+encoded[:, :, W:] = uint16(u >> 16)
+u_restored = uint32(low) | (uint32(high) << 16)
+depth_float32 = reinterpret_float32(u_restored)
+```
+
+统一采用 little-endian 原始字节接口。`camera.image_size` 始终是逻辑 RGB 尺寸 `[W,H]`，不能改成编���视频的 2W；`storage.encoded_size=[2W,H]`。包括 NaN 在内的 float32 位串必须全帧完全恢复，不允许改成 float16、整数毫米、归一化灰度或 RGB 伪彩色。这里的无损验收基准是完成既定米制／坐标转换后的标准 float32 数组，不声称消除 float64→float32 等原约定转换本身的舍入。
+
+#### 3.2.2 逐像素有效性
+
+`depth_valid.mkv` 使用 `mkv/bool_ffv1/v1`，FFV1 gray8（FFmpeg pixel_format=`gray`），物理尺寸 `[W,H]`。像素只能为 0 或 1，统一解码为 `[T,H,W] bool`，逐像素含义不变。这样不再把庞大的 bool 数组保存为未压缩 `.pt`；标注的帧级／手级 valid 仍保留原有小数组。
+
+登记示例（W、H 以下取 640、480；decoder/encoder_parameters 的实际版本由转换器填写）：
+
+```json
+{
+  "available": true,
+  "schema_id": "depth/z_float32/v1",
+  "data_ref": "observation/depth.mkv",
+  "axes": ["T", "height", "width"],
+  "valid_ref": "observation/depth_valid.mkv",
+  "valid_axes": ["T", "height", "width"],
+  "unit": "m",
+  "depth_type": "z",
+  "depth_origin": "measured",
+  "invalid_value": "NaN",
+  "storage": {
+    "profile": "mkv/depth_float32_bits/v1",
+    "codec": "ffv1",
+    "pixel_format": "gray16le",
+    "encoded_size": [1280, 480],
+    "source_lossy": false,
+    "decoder": {"name": "ffmpeg", "version": "填写实际版本"},
+    "encoder_parameters": {"name": "ffv1", "level": 3, "version": "填写实际版本"}
+  },
+  "valid_storage": {
+    "profile": "mkv/bool_ffv1/v1",
+    "codec": "ffv1",
+    "pixel_format": "gray",
+    "encoded_size": [640, 480],
+    "source_lossy": false,
+    "decoder": {"name": "ffmpeg", "version": "填写实际版本"},
+    "encoder_parameters": {"name": "ffv1", "version": "填写实际版本"}
+  },
+  "provenance": {"kind": "format_conversion", "method": "源深度转米制 z float32；逐位打包为 FFV1"},
+  "source_locator": {"file": "raw/data/depth.zip", "member": "depth/0001.png"}
+}
+```
+
+深度和有效性视频都必须逐帧解码验收：float32 位串相同、bool 相同、T 相同，并检查 valid=true 的正有限值和 valid=false 的 NaN。读取器必须按 profile 还原，不能把 depth.mkv 的编码像素直接当深度值。
+
+## 4. camera 相机、帧索引与标定
+
+### 4.1 frames.jsonl：索引、时间与对应关系
+
+`camera/frames.jsonl` 每行对应统一帧号 `frame_index=0..T-1`，是所有组件的唯一行索引。至少包含 `frame_index`、`timestamp_ns`、`clock_domain`、`source_locator`；原始帧号/文件名保存在 `source_frame_id`，不能被新帧号替代。每行还可有 `annotation_matches`，按组件键保存 `source_frame_id`、`source_timestamp_ns`、`clock_domain`、`match_method` 和 `delta_ns`；方法为 exact/index_map/nearest/interpolated，索引映射无时间依据时 delta_ns=null。nearest 必须记录允许误差；interpolated 必须记录两端源标识和权重。delta_ns 仅在同一时间域中计算为源标注时间减图像时间。多相机共享 `sequence_id`；已知同步关系必须用 `sync_id` 或明确的源帧/时间映射表达，不能仅凭相同数组索引认定同步。
+
+每个视图必须提供此文件。源时间使用整数纳秒及 `clock_domain`，未知写 JSON `null`；不同时间域不能直接比较。逐帧 split 可以覆盖 metainfo 的视图默认划分，并保留来源，见 2.4。
+
+### 4.2 坐标系与标定通则
+
+| 项目 | 统一约定 |
+|---|---|
+| 相机坐标 | 右手系，X 向右、Y 向下、Z 向前，长度单位米 |
+| 齐次变换 | 列向量，`p_target = T_target_from_source @ p_source` |
+| 图像坐标 | `(u,v)`；左上像素中心为 `(0,0)`，u 向右、v 向下，单位 pixel |
+| 固定标定 | 本版展开为 T 份，同一 schema 不混用静态与动态形状 |
+| 分辨率或模型变化 | 分开存储为视图／片段，不在同一组件中混装 |
+| 未知标定 | 相机内外参允许缺省，不填单位阵冒充已知标定 |
+
+### 4.3 intrinsic.pt 与畸变参数
+
+| model / schema_id | intrinsic.pt 的固定结构 |
+|---|---|
+| pinhole / camera/pinhole/v1 | `[T,3,3]`，标准 K，最后一行为 `[0,0,1]`；限无畸变或已校正图像 |
+| fisheye624 / camera/fisheye624/v1 | `[T,15]`，`[f,cx,cy,k0,k1,k2,k3,k4,k5,p0,p1,s0,s1,s2,s3]`，采用 Project Aria Fisheye624 投影定义 |
+| opencv_fisheye4 / camera/opencv_fisheye4/v1 | `intrinsic.pt [T,3,3]` 保存 K；同组件 `distortion_ref` 指向 `[T,4]`，顺序 `[k1,k2,k3,k4]`，使用 OpenCV fisheye 等距角度多项式；已由 ReInterHand 实际原始相机参数验证 |
+| rational8 / camera/rational8/v1 | `intrinsic.pt [T,3,3]` 保存 K；同组件必需 `distortion_ref` 指向 `[T,8]`，顺序 `[k1,k2,p1,p2,k3,k4,k5,k6]` |
+| unknown / null | 不提供可用 intrinsic，不声称可以投影 |
+
+有畸变的其他模型必须注册新 schema，固定参数名、顺序和投影公式；不能塞进针孔 K 后遗失畸变项。`metainfo.camera.schema_id` 必须与 intrinsic 组件一致。同一模型的别名不可隐式改变参数布局。
+
+`opencv_fisheye4`：对前方点令 `x=X/Z, y=Y/Z, r=sqrt(x²+y²), theta=atan(r)`，`theta_d=theta*(1+k1*theta²+k2*theta⁴+k3*theta⁶+k4*theta⁸)`；`u=fx*x*theta_d/r+cx, v=fy*y*theta_d/r+cy`。光轴 r=0 时使用连续极限 `theta_d/r=1`。四系数不能按普通径向畸变解释，也不能补零冒充 Fisheye624。
+
+rational8 对归一化坐标 x=X/Z、y=Y/Z，令 r²=x²+y²，径向项为 `(1+k1*r²+k2*r⁴+k3*r⁶)/(1+k4*r²+k5*r⁴+k6*r⁶)`；再加 p1/p2 切向项并经 K 得到像素。`distortion_ref` 与 intrinsic 共用 `[T]` valid，保存到 camera/distortion.pt。不得把这种畸变三维坐标当作真实相机系 xyz 保存。
+
+内参有效性为 `intrinsic_valid.pt [T] bool`；有畸变时使用对应 `distortion_ref`。同一模型的参数名和顺序固定。
+
+### 4.4 extrinsic.pt 与 extrinsic_valid.pt
+
+`extrinsic.pt` 为 `[T,4,4]`，schema `transform/se3/v1`，方向固定 `T_camera_from_world`，有效性 `[T]`。世界坐标含义及源坐标转换记录在 provenance 中；只有相机坐标标注不要求虚构世界坐标。旋转块须正交且 determinant≈1。
+
+## 5. annotation 标注数据
+
+### 5.1 标注通则：手槽、valid 与 visibility
+
+手槽固定为 0=left、1=right，指解剖学左右，不按画面位置或可见性重排。本阶段只处理左右手两槽，不处理更多手。实体引用使用 `hand_left`、`hand_right`；手槽不证明跨采集片段的个人身份。
+
+#### 5.1.1 valid：标注是否可用于监督
+
+标注的 `valid=true` 表示：有对应实际标注、通过本 schema 数值检查，且通过来源明确提供并适用于该组件的有效性/QA 条件。`valid=false` 可由缺标注、无法建立帧对应、源无效占位或官方 QA 不通过导致；不等于画面没有手或手不可见。各组件独立判断，MANO 按 `[T,2]` 标记左右手，关键点按 `[T,2,K]` 标记。
+
+**2026-09-20 将标注 valid 从“仅表示存在”修订为“可用于监督”，字段形状不变。** 公式为 `source_annotation_present AND numerical_checks_pass AND applicable_official_QA_pass`。没有官方 QA 时省略最后一项，不伪造 QA；不能把未知当成已通过已存在的 QA。适配器在组件 `valid_policy` 中记录具体规则、来源、聚合/广播作用域。缺失官方必需 QA 的记录须明确处理为未知或暂停验收，不能静默通过。无效浮点项仍填 NaN，原始数据保留在 raw；被排除原值可用于 doc 中明确标识的诊断渲染，不能冒充 shared 有效监督。
+
+本地筛选须依据明确的规则，并在 `valid_policy.local_excluded_frame_indices`、排除原因和正式报告中记录。单个可疑投影作为问题记录，不直接硬编码排除该帧。已确认的曝光异常须明确作用范围，不能把一个序列的“前两帧有问题”套到所有序列。RGB 编码完好与几何监督可用是两件事，允许保留原图及帧索引而屏蔽对应标注；未发现问题不等于完成全部人工审核，试点待确认时不能作为已验收发布使用。
+
+源字段必须按官方语义映射，不能直接按名字复制：
+
+| 来源 | 官方含义 | 映射原则 |
+|---|---|---|
+| [H2O](https://github.com/taeinkwon/h2odataset) 的手标注首位 | 0 未标注、1 已标注 | 可作为对应手标注存在性的依据，不推断未标注原因 |
+| [HOT3D](https://github.com/facebookresearch/hot3d#using-mask-files) 的 pose_available、visible、mask_qa_pass | 分别表达姿态存在、可见性、对画面中可见手物位姿的人工 QA | 逐手/物体记录决定存在性，再 AND 当前相机当前帧 QA；QA 是整帧聚合判定，保守排除该帧全部手物几何监督，不声称每个实体均有错误；源可见性基于位姿推导时，其自身 valid 同样受 QA 约束 |
+| [Dex-YCB](https://github.com/NVlabs/dex-ycb-toolkit#loading-dataset-and-visualizing-samples) 的全零 pose_m、全 -1 joints | 不可见的手或缺标注共用占位表示 | 对应组件 valid=false；不能从该值判定 visibility=false，也不能声称原因是动捕缺失 |
+| [ARCTIC](https://github.com/zc-alexfan/arctic/blob/master/docs/data/data_doc.md) 的 is_valid/left_valid/right_valid | 是否有效用于官方评估 | 核实每项筛选依据；不把可见性/评估选择直接等同 QA。本次原始 mano.npy 不含这些派生评估标记，按原记录与数值检查生成 valid，并声明没有额外官方 QA 输入 |
+
+源 valid=false 若代表评估筛选或其他条件，应检查实际标注和源约定后映射；不能一律清空坐标。反之，占位零或 -1 只有在源官方明确规定时才判为缺失，不泛化成所有数据的判断规则。
+
+`valid=true` 的标注应通过其 schema 的数值与形状检查；检查失败必须报告或修复，不能静默隐藏。QA 不通过应报告排除数量与原因，并保留可追溯信息；不能把这类无效称为缺标注。RGB、深度等观测有效性仍按各自 schema；手物 QA 不自动否定 RGB、静态相机内参或物体资产。相机外参是否过滤须依据其自身来源规则；HOT3D 本次保留源外参用于溯源和诊断，不能用其存在性证明 QA 失败帧的几何监督正确。
+
+#### 5.1.2 visibility：可见性属性
+
+附加质量属性只保留 `visibility`，存放在对应模态或标注集内，不设 quality 目录。同一份可见性仅存一次；uv 与 xyz 共用时引用同一个文件。
+
+visibility 通过所属组件 attributes 登记 data_ref、axes、dtype、visibility_type、meaning、provenance、source_locator；未知值通过属性 valid_ref 标识，全已知可声明 valid=true。比例与布尔值不相互强转，也不按任意阈值二值化。没有 MANO 时，整手可见性挂在 keypoints 组件，使用独立的 whole_hand_visibility.pt，明确 meaning=whole_hand，不与逐关键点可见性混用。
+
+`valid` 回答“这项标注能否用于监督”，`visibility` 回答“在当前图像中是否可见（或可见多少）”。不可见的手或关键点仍可有有效标注；不能因 visibility=0 就清空坐标。源没有 visibility 时缺省或标为未知，不由 valid 推断。
+
+| 情况 | 手标注 valid | visibility |
+|---|---|---|
+| RGB 已开始录制，动捕尚未开始，没有其他对应手标注 | false | 以独立源可见性标注为准；没有则未知 |
+| 手被遮挡，但源有标注且通过适用有效性检查 | true | 若源声明不可见，则 false 或源可见比例 |
+| 手可见且有对应标注，通过适用 QA | true | 若源有此信息，则 true 或源可见比例 |
+| 手可见，但该帧漏标或动捕中断 | false | 可以为 true，但必须有独立来源依据 |
+| 有手部参数，但官方 QA 明确不通过 | false | 不据此断言不可见；源可见比例若基于不可靠位姿，其属性 valid 也为 false |
+
+visibility 自己的 valid_ref 表示该可见性标注是否可用，不复用 MANO 或关键点坐标的 valid_ref；明确记录源存在性及适用的 QA 条件。上述例子不要求凭图像自动生成可见性。
+
+不设 quality 组件或目录，不另存拟合误差、曝光、截断比例等通用质量数组。适用的官方 QA 合并进对应 valid，在 `valid_policy` 记录来源与规则；正式校验报告保留原始 QA 时间区间及排除统计。原始 QA 文件仍在 raw，官方 split 保留。
+
+### 5.2 mano：参数与整手可见性
+
+MANO schema `mano/axis_angle61/v1`：`motion.pt [T,2,61]`，顺序为完整 16 关节 axis-angle 48、shape betas 10、全局平移 3。姿态使用弧度，平移米，位于当前相机系。有效性 `[T,2]`；缺少一侧整槽 NaN。PCA 必须先按对应模型展开，不能将 PCA 系数直接填入 axis-angle。
+
+`model_profile` 必须固定 MANO 版本、`flat_hand_mean=true`、左手 shapedirs 修正状态。平移表示模型输出顶点所加的全局平移，不是某个腕关节的坐标；源语义不同须转换。shape 无量纲。模型权重不作为共享数据附带上传。
+
+`mano/visibility.pt [T,2]` 保存源整手可见比例时使用 float32 [0,1]，`visibility_type=ratio`；只有实际采用 0–100 的百分比才除以100，不能只凭表头判断。HOT3D 的 `visibility_ratio[%]` 实值已经是0–1，官方读取器原样读取，必须保留原值。通过 MANO 组件 attributes 登记，使用独立有效性，不复用 motion_valid。
+
+手部共享表示仅保留 MANO 参数与选定关键点，不发布独立手表面几何。需要渲染或接触计算时由 MANO 参数生成，几何引用使用固定 MANO 模型配置与顶点顺序；不影响 object 下的物体模型和材质资产。
+
+### 5.3 keypoints：二维点、三维点与逐点可见性
+
+关键点独立于 MANO 参数存在；没有 motion 也可有关键点。`uv.pt [T,2,K,2]` 与 `xyz.pt [T,2,K,3]`，有效性 `[T,2,K]`，分别采用 `keypoints/uv/v1` 和 `keypoints/xyz/v1`。K 由 `keypoint_schema` 固定，不能在同一数组中逐帧变化；同体系统一顺序。数值使用像素坐标或米制相机系。
+
+| keypoint_schema | K 与固定索引 |
+|---|---|
+| mano21/v1 | 21；wrist 后依次 thumb/index/middle/ring/little，每指由近端至指尖 |
+| rhd21/v1 | 21；保持 wrist，然后五指由近端至指尖；原始每指 tip→palm 反序。只统一顺序，不宣称解剖学位置等价于 MANO |
+
+具体名称与源索引映射见 `schema.json`。当前只登记 MANO21 与 RHD21；不登记 MediaPipe、ReInterHand capture21 或 fit24 中间关键点。
+
+每个发布中的每个视图只提供一套标准手关键点，直接放在 keypoints 下，不设 annotation_set_id 字段或目录。uv 与 xyz 是这套关键点的二维与三维表示，不是两套标注。provenance 分别记录二者来自原始标注还是计算结果；不要求已有 2D 必须由当前 xyz 重新投影覆盖。
+
+ReInterHand 的 Original keypoints���Original fits keypoints 及带前臂关键点不纳入 shared；如需手关键点，从选定 MANO 参数生成 MANO21 并记录模型/指尖回归约定。其他数据集默认使用选定的官方关键点，不额外生成平行的 MANO 派生集合。后续主动更换生成方法或结果时发布新的 dataset_version，不在同一版本内增设方法分支。
+
+manual/projection/fitting 只记录在 provenance，不作为目录或类型分支。不得通过坐标为零、是否出画或 visibility 判断标注是否存在。派生标注仅在实际生成并通过数值检查后声明 valid=true；没有生成的项为 false，不伪造坐标。
+
+`keypoints/visibility.pt [T,2,K]` 保存源逐点可见性，布尔表示使用 `visibility_type=boolean`。uv 与 xyz 共用时引用同一份文件。整手比例属性没有 MANO 可挂载时使用 `whole_hand_visibility.pt [T,2]`，`meaning=whole_hand`；它与逐点可见性分开，通用登记规则见 5.1。
+
+不使用 annotation_set_id；ReInterHand 中间关键点集合不进入 shared，raw 保留来源文件。
+
+### 5.4 object：身份、资产、位姿与几何
+
+#### 5.4.1 metainfo.json：物体身份与跟踪作用域
+
+手用 `hand_left`、`hand_right` 及槽位 0、1 表示，不再另存 entities.json。物体身份登记在 `annotation/object/<object_id>/metainfo.json`，必填 `object_id`、`identity_scope`、`scope_id`、`source_locator`；可含 `category_id`、`source_object_id`、`mesh_ref`、`has_model`、部件引用。即使只有分割、没有 pose/mesh，也可只有这份物体元信息。
+
+`identity_scope` 为 frame/subsequence/sequence，`scope_id` 指定具体作用域；跨相机共享身份只在已有源对应证据时使用。VISOR 按官方 subsequence 与实体名关联时记录 `association_basis=official_subsequence_entity_name`，不能声称这是全视频物理实例跟踪。无需与 object_id 一一对应的标准 track_id；官方独立轨迹标识存在时保留为 source_track_id，并声明作用域。手槽不证明跨采集片段的个人身份。
+
+#### 5.4.2 assets：网格、材质与纹理
+
+资产属于物体，统一存放在 `annotation/object/<object_id>/` 内，可用 `assets/` 子目录组织网格、材质、纹理和部件资产。序列层不设共享 assets。所有引用采用视图根目录为基准；其中物体资产组件的 `data_ref` 与部件的 `mesh_ref` 采用相对于当前视图目录的 POSIX 路径，例如 `annotation/object/cup_01/assets/cup.obj`，且必须指向所属物体目录内的资产。OBJ 内部的材质、纹理引用遵循文件格式自身的相对路径规则，依赖文件一并保留。不同相机视图允许各自保存物体资产，不要求跨视图去重；下载物体时包含其资产依赖。
+
+网格 schema 为 `mesh/obj/v1`，顶点米制，静态有效性可直接声明标量 `valid`，或通过 `valid_ref` 指向标量 bool Tensor；不按 T 重复静态资产。由物体／部件通过引用关联，不仅凭文件名猜测。
+
+#### 5.4.3 pose.pt、pose_valid.pt 与 visibility.pt
+
+刚体 schema 为 `object/rigid/v1`，`pose.pt [T,4,4]` 固定为 `T_camera_from_object`，平移米制，有效性 `[T]`。物体坐标原点、轴向必须与 mesh 一致；旋转块为正交矩阵且 determinant≈1。
+
+`object/<id>/visibility.pt [T]` 保存源物体可见比例，float32 [0,1]，`visibility_type=ratio`。通过所属物体组件 attributes 登记，独立表达其有效性。
+
+#### 5.4.4 articulation.json、部件位姿与 joint_state.pt
+
+关节物体：schema `object/articulated/v1`，`articulation.json` 登记 `parts` 与 `joints`。每个 part 有 `part_id`、`parent_id`（根为 null）、`has_model`、`mesh_ref`；has_model=false 时 mesh_ref=null，结构节点允许没有网格。父链不得成环。每个 joint 有唯一 `joint_id`、`parent_part`、`child_part`、`type`；fixed 单位 null、revolute 弧度、prismatic 米。有源限位时保存 `limits:[lower,upper]`；没有时 null，不伪造上下限。部件姿态统一为 `T_camera_from_part [T,4,4]`，不与父局部变换混装。
+
+若提供 `joint_state.pt [T,J]`，必须登记固定 joint_order、轴在父部件系的表达、零位变换与正方向，valid 为 `[T,J]`；缺少这些运动学定义时只发布部件位姿与已有来源信息，不伪造可驱动的关节模型。关节值与部件位姿同时存在时需验证一致性。
+
+部件层级应声明语义：纯语义包含树允许 `joints=[]`，不将包含关系推断为机械关节。关节状态 schema 为 `object/joint_state/v1`；转动用弧度、移动用米，混合类型按 joint_order 对应的关节定义解释。
+
+#### 5.4.5 dimensions、几何关键点与顶点部件标签
+
+`dimensions.pt [T,3]` 采用 `object/dimensions/v1`，有效性为 `[T]`。
+
+物体/部件 `dimensions` 为 `[length,width,height]` 米，轴顺序与该 mesh 的局部坐标关联，登记 `dimensions_semantics` 和来源。HOI4D dimensions 与 mesh 的具体尺度关���未验证前标为 source_unspecified，保留源值并暂停依赖该关系的几何验收，不能默认为待施加缩放。需要施加缩放时明确 scale_applied_to_mesh，禁止重复应用。
+
+源物体关键点放 `object/<id>/keypoints/xyz.pt [T,K,3]`（相机系）或 `template_xyz.pt [K,3]`（物体系），schema 分别为 object_keypoints/camera/v1 和 object_keypoints/template/v1。明确 keypoint_schema、keypoint_names、单位及有效性；模板点配合 pose 可投影，无须重复生成全部帧。H2O 中心/角点/棱中点，以及 ARCTIC 部件点分别登记具体源顺序，不套用手的 21 点 schema。每个发布选择一套源几何点并记录选择，不另建方法分支。
+
+`vertex_part_ids.pt [V] int64` 绑定 `mesh_ref` 与 topology_id，登记整数到 part_id 的映射；未知用 -1。ARCTIC 不同拓扑的资产分别绑定，不能共用未经转换的顶点标签。部件层级中 is_instance/has_model 等源属性保留，几何文件仍全部归 object。
+
+### 5.5 segment：分割类别、编码与覆盖范围
+
+每个发布选择一套任务使用的分割表示，不建立方法／标注集分支。没有分割就不创建 segment 目录，不为其他模态创建空的实体表。
+
+| schema_id | 文件与数值语义 |
+|---|---|
+| segment/uint16_labels/v1 | mask.mkv；整数区域编号，通过 instances.jsonl 解释 |
+| segment/instance_polygons/v1 | polygons.jsonl；实例多边形，可重叠；当前不支持孔洞拓扑 |
+| segment/foreground_intensity/v1 | mask.mkv；0–255 前景强度，不是区域编号或经过校准的概率 |
+
+#### 5.5.1 uint16 标签图与 instances.jsonl
+
+组件声明 `segmentation_type=semantic/instance/part`；同一组件固定一种类型。`mask.mkv` 解码为 `[T,H,W] uint16`，schema `segment/uint16_labels/v1`，使用 `mkv/uint16_ffv1/v1`，FFV1 gray16le 无损；只有所有区域编码均 ≤255 时允许 gray8（pixel_format=gray），读取后转 uint16，数值不变。禁止调色板和有损编码。有效性 `[T]`，表示该帧有可用分割标注。数值 0 固定为 unassigned，65535 固定为 ignore，1–65534 为区域编码，不能直接等同 entity_id。
+
+`instances.jsonl` 每帧一行，包含 `frame_index` 与 `labels`。每条标签包含 `mask_value`、`category_id`、`entity_id`、`part_id`；无意义或未知字段为 null。semantic 允许 entity_id=null；instance 必须引用 hand_left/hand_right 或登记的 object_id；part 还须关联登记部件。源实例 ID 和其他属性分别保存在 `source_instance_id`、`source_attributes`。mask 中每个非保留编码必须有且仅有一个映射。
+
+FreiHAND 的试点按本任务需要，统一到手／背景类别，保留每个原生区域编码；不要求给原生细分区域猜测手指名称。需要细部件监督时，尚未核实的部件名称必须保持未知，不能因粗粒度手类别已核实便声称部件语义也已统一。
+
+#### 5.5.2 重叠实例多边形
+
+每发布选择一套任务使用的分割表示，不建立方法/标注集分支。VISOR 的实际试点存在重叠实例多边形，单层标签图不能无损表达。对此补充 `segment/instance_polygons/v1`：`segment/polygons.jsonl` 每帧包含 `frame_index` 和 `instances`，每实例包含 `mask_value` 与 `polygons`（列表，每个多边形为源图像像素坐标顶点序列，保留原始精度）。身份和类别仍由同一 `instances.jsonl` 映射，有效性仍为 `[T]`，通过 schema 区分物理编码。允许多实例重叠，不采用后画覆盖前画。消费者按需栅格化为独立实例二值 mask；不能声称合并的单层 mask 无损。
+
+#### 5.5.3 灰度前景强度
+
+ReInterHand 的官方渲染 mask 为双手合并的灰度前景掩码，实测存在 0–255 的中间值，不是左右手实例编号。登记 `segment/foreground_intensity/v1`：仍存 `segment/mask.mkv`，解码为 `[T,H,W] uint16`，使用 `mkv/uint16_ffv1/v1`，此类数据优先编码为 FFV1 gray8（pixel_format=gray），读取后转 uint16。有效值范围 0–255，归一化值为 `float32(value)/255`；0 为背景端点、255 为前景端点。组件声明 `foreground_category=hand`、`entity_id=null` 和双手合并语义，不使用区域编号映射。这里保留原始渲染 mask 强度，不把它称为经过校准的概率。若任务需要二值 mask，阈值化属于任务层派生操作，记录阈值；共享层不丢弃中间灰度。不同 schema 的数值含义不同，不能将该文件按整数类别图解码。
+
+FreiHAND 训练集的 JPEG mask 使用 `segment/foreground_intensity/v1`，逐像素保留 JPEG 解码后的灰度，`entity_id=hand_right`。该 schema 不限定双手合并：单手或合并范围由 `entity_id` 与 `coverage` 明确声明。评估集整数区域图仍用 `segment/uint16_labels/v1`，两者不能仅凭文件名使用同一种解码语义。
+
+#### 5.5.4 coverage 与有效性
+
+`coverage.kind=full/selected_entities/partial/unknown`。full 指声明���标注范围完整；selected_entities 还应列出目标范围。无论哪种，0 本身不自动成为背景；背景要有明确类别映射，未标注区域不能用于构造负样本。保留源逐类/逐实例 exhaustive 信息，未知不可改写为完整。
+
+三种表示的帧级有效性均为 `[T]`，表示该帧是否提供此分割标注；对应 valid_ref 在组件中登记，仍用小型 bool `.pt` 数组。栅格 mask 必须登记 storage，格式见 3.1.2，所有帧的整数编码或强度逐值完全相同。编码语义由 schema_id 决定，读取器不能仅依据 mask.mkv 文件名选择解码方式。
+
+### 5.6 contact：接触、抓取、区域与距离
+
+#### 5.6.1 relations.jsonl 与 relations_valid.pt
+
+`annotation/contact/relations.jsonl`，schema contact/relations/v1，每帧一行含 frame_index、relations。每条记录含 relation_type（contact/grasp）、subject_id（手槽或登记实体）、object_id、status、source_locator、provenance。status 固定为 positive/negative/unknown/target_unannotated：positive 必须有可解析目标，negative 表示源明确否定该关系，unknown 为不确定，target_unannotated 为确有接触但目标未标注。非 positive 的 object_id 可 null，不能用 null 冒充“不接触”。稀疏源无记录的帧不自动补 negative。
+
+VISOR 的对象 mask ID 必须先按同帧 source_instance_id 解析成 object_id；hand-not-in-contact、none-of-the-above、inconclusive 分别映射 negative、target_unannotated、unknown，保留原始字符串。Dex ycb_grasp_ind 通过 ycb_ids 映射到登记物体，以 relation_type=grasp 保存；不得当作逐顶点接触。
+
+relations 组件的 valid_ref 指向 `relations_valid.pt [T] bool`，表示该帧是否有可用源关系标注。有效帧中可以存在 status=unknown 的明确不确定标注；无源标注的帧为 valid=false、relations=[]。组件仍声明 available/provenance/source_locator。
+
+#### 5.6.2 实体对的顶点区域与距离
+
+已发布的接触区域/距离按参与关系的 subject_id 与 object_id 组织，每对实体保留选定的一套结果，以精确 geometry_ref/topology_id 关联到手或物体/部件；不要求每套数据都有。顶点区域 schema contact/vertex_region/v1：bool `[T,V]` 加同形 valid；顶点距离 schema contact/vertex_distance/v1：float32 `[T,V]`、米、同形 valid，声明 source/target geometry_ref、distance_definition（signed/unsigned、测量到点集或表面等明确源定义）。geometry_ref 必须指向已登记 mesh 或由固定 MANO 模型配置确定的拓扑；拓扑/顶点数变化时按拓扑拆分片段。其他源区域表示只有在实际接入并登记 schema 后才可转换，不能随意栅格化宣称等价。
+
+顶点区域／距离已有存储定义，但尚未完成真实输入试点；接入时必须提供源语义与几何验收证据，不因已经登记 schema 就视为通过。
+
+## 附录：任务范围
+
+不强制收录 2D bbox、完整高频动捕、动作程序、身体 SMPL-X、点云或网络特征。范围内的接触、visibility、选定手关键点、物体几何不能因不便处理而静默丢弃。分割采用的物理编码由第 5.5 节决定，存在重叠时不能以单层标签图声称无损。
+
+不设置 entities_ref/segment/entities.json，不要求与 object_id 重复的 track_id。MediaPipe 不在当前关键点定义中；旧版本迁移必须保留可追溯关系，不能直接删表。手部共享表示不含独立 hand_mesh。
+
+schema.json 是普通注册表，不是 JSON Schema 文档或可执行验收器；定义存在不代表真实数据已验证。验收须保留实测范围与未决项。
+
+本存储约定只更新标准，不代表既有 FFV1／depth.pt 试点已迁移。新发布必须显式写 storage 配置；旧产物按其旧定义读取，不能只改文件后缀或版本标识伪装为合规。

H2O/shared-v1.0/script/provenance/previous-run/script/provenance/previous-run/docs/h2o-RAW_TO_SHARED.md

@@ -0,0 +1,30 @@
+# H2O raw → shared-v1.0
+
+## 范围与来源
+
+本版本仅转换 `raw/data/subject{1,2,3,4}_v1_1.tar.gz` 的真实五视角 (`cam0`–`cam4`) 序列；`h2o_CASA.tar` 是单独的扩展，不与标准序列混合。每个原始序列存于 `data/subjectN_<scene>_<sequence>/cam{0..4}/`。源压缩包保持不变，不合成源不存在的标注。输出仅在本地制作；完整序列试点及人工检查通过、更新全量预估之前，不启动全量。
+
+| 源包/字段 | shared 文件/处理 |
+| --- | --- |
+| `subjectN_v1_1.tar.gz`: `camX/rgb/<frame>.png` | `observation/rgb.mkv`：逐帧源 RGB 解码为 RGB24，FFV1 直接编码；输出**每帧**解码与源 RGB 字节摘要完全一致。`rgb_valid.pt` 记录存在性，源帧号和每组件的包成员及 SHA256 写入 `camera/frames.jsonl`。不再额外编码未采用的 H.264 RGB 候选；FFV1 编码器进程失败时才尝试 H.264。 |
+| 同包 `camX/depth/<frame>.png` | `observation/depth.mkv` 和 `depth_valid.mkv`；按源毫米单位除以 1000 转米，保留无效深度与全帧时间轴，深度位串以 FFV1 保真封装并逐帧校验。既不做 3 米裁剪也不造出缺失的深度。 |
+| `camX/cam_intrinsics.txt`、`camX/cam_pose/<frame>.txt` | `camera/intrinsic.pt`、`extrinsic.pt`：保留源内参；逐帧源 `T_world_from_camera` 求逆后存 `T_camera_from_world`；声明变换约定与有效性。 |
+| `camX/hand_pose_mano/<frame>.txt` | `annotation/mano/motion.pt`、`motion_valid.pt`：源左/右手标志各自判断有效性；姿态与平移保持源数值，声明 MANO 模型配置，缺模型则明确标依赖缺失，不伪造网格或掩码。 |
+| `camX/hand_pose/<frame>.txt` | `annotation/keypoints/xyz.pt`、`xyz_valid.pt`：源双手 21 点，相机坐标系米制，与 MANO 有效性彼此独立。HTML 中的投影仅用于人工预览，不写回训练数据。 |
+| `camX/obj_pose/<frame>.txt`、`obj_pose_rt/<frame>.txt` | `annotation/object/<name>/keypoints/xyz.pt`、`pose.pt`：按源类号关联 21 个物体点及物体→相机位姿；仅源存在时登记组件及有效性。 |
+| `object.zip` | `annotation/object/<name>/assets/` 与源 OBJ、MTL、纹理；静态形状预览不等于所有帧网格位姿的投影证明。 |
+| `label_split.zip` 的 `action_{train,val,test}.txt` | 序列元数据的官方划分及来源；未知映射明确记 `unknown`，不自行编排。 |
+
+`cam4` 是 ego，`cam0`–`cam3` 是 exo。真实时间/FPS 缺乏可靠来源时使用 null，不把 MKV 封装帧率冒充采集时钟。源 `action_label`、`verb_label`、pose 任务帧列表和 `manolabel_v1.1.tar.gz` 保留在 raw；当前契约没有相应动作组件，本发布不宣称它们已转换。源没有像素分割，本版本不投影网格生成替代分割。
+
+## 源资产文件名修复
+
+源 `object/cocoa/cocoa.obj` 引用 `ovomaltine.mtl`，材质内写的是 `map_Kd cocoa.png`，但原 ZIP 只提供 `ovomaltine.png`。仅在临时解压目录为该已确认的纹理增加**字节相同**的 `cocoa.png` 别名，保留原 PNG、OBJ、MTL 不变；在网格 `source_locator.texture_aliases` 中记录别名、原始包成员和 SHA256。别名与已保留的旧 cocoa 资产一致，不替换图像、标注或网格。全部八类物体的材质依赖在读取大 subject 包之前预检，其他缺失/越界依赖仍报错。
+
+## 加速与验收
+
+四个互不依赖的 subject 压缩包可由 `--workers 1..4` 并行顺序读取，各 worker 一次只处理一个序列。数值库保持单线程；FFV1 编码支持 `--codec-threads 1|2|4|8|10`，当前全量按用户要求使用 10，逐帧解码校验仍保留。并发数由相同输入实测吞吐量决定，记录 OS 线程峰值和磁盘占用，不以 CPU 线程数本身作为提速证据。RGB 源为 PNG，不适用 ARCTIC/HOT3D 的 JPEG 包直封和 16 帧抽查：RGB、深度、有效性与数值模态仍逐帧精确校验。临时文件在 `/data` 上隔离放置，历史 62 条目录不复用、不覆盖；调整并发时，在新的运行目录复用本次全量已经原子发布的完整序列，并从源重新核对帧号及逐组件 SHA256。未完成的临时序列不复用，原运行的脚本、配置及 QC 保留用于溯源。
+
+先运行不截帧的代表性完整序列试点，保存自动 QC、真实耗时和 `doc/validation/index.html`（RGB、深度、源 3D 关键点投影、原生物体网格和全组件数值/有效性）。试点由人工核查之后再决定全量；全量后再次抽样人工核查。报告须区分数据完整性、源未提供的模态与仅供人工判断的几何对齐。
+
+2026-09-29 完整序列试点已获人工确认，用户已授权本地全量。2026-09-30 用户取消 CPU 使用量上限，指定 4 个独立压缩包 worker、每个 FFV1 编码器 10 线程；Python worker、读取及校验子进程不计入这最多 40 个编码线程，因此总线程数不固定为 40。该配置不代表已证明加速倍数。源包和旧产物保留；代码变化使用独立运行目录，已完��序列保留原始构建记录及生产脚本快照。

H2O/shared-v1.0/script/provenance/previous-run/script/provenance/previous-run/docs/schema.json

@@ -0,0 +1,869 @@
+{
+  "schema_version": "hand-warehouse/1.0",
+  "validity_revision": "usable-annotation-2026-09-20",
+  "annotation_validity": {
+    "meaning": "annotation usable for supervision, not merely record presence",
+    "formula": "source_annotation_present AND numerical_checks_pass AND applicable_official_QA_pass AND NOT explicitly_reviewed_local_exclusion",
+    "without_official_qa": "omit QA term; declare source rules; do not fabricate QA",
+    "policy_field": "components.<name>.valid_policy",
+    "visibility_is_independent": true,
+    "invalid_float_value": "NaN",
+    "hot3d": "broadcast per-camera frame manual QA conservatively to hand/object labels and pose-derived visibility; retain RGB and source calibration; keep native QA in raw and review report",
+    "hot3d_visibility": "native visibility_ratio[%] already ranges from 0 to 1; do not divide by 100"
+  },
+  "hand_slots": [
+    "left",
+    "right"
+  ],
+  "keypoints": {
+    "mano21/v1": {
+      "count": 21,
+      "names": [
+        "wrist",
+        "thumb_1",
+        "thumb_2",
+        "thumb_3",
+        "thumb_4",
+        "index_1",
+        "index_2",
+        "index_3",
+        "index_4",
+        "middle_1",
+        "middle_2",
+        "middle_3",
+        "middle_4",
+        "ring_1",
+        "ring_2",
+        "ring_3",
+        "ring_4",
+        "little_1",
+        "little_2",
+        "little_3",
+        "little_4"
+      ],
+      "source_index_map": null,
+      "note": "adapter maps official MANO joint and fingertip convention"
+    },
+    "rhd21/v1": {
+      "count": 21,
+      "names": [
+        "wrist",
+        "thumb_1",
+        "thumb_2",
+        "thumb_3",
+        "thumb_4",
+        "index_1",
+        "index_2",
+        "index_3",
+        "index_4",
+        "middle_1",
+        "middle_2",
+        "middle_3",
+        "middle_4",
+        "ring_1",
+        "ring_2",
+        "ring_3",
+        "ring_4",
+        "little_1",
+        "little_2",
+        "little_3",
+        "little_4"
+      ],
+      "source_index_map_per_hand": [
+        0,
+        4,
+        3,
+        2,
+        1,
+        8,
+        7,
+        6,
+        5,
+        12,
+        11,
+        10,
+        9,
+        16,
+        15,
+        14,
+        13,
+        20,
+        19,
+        18,
+        17
+      ]
+    }
+  },
+  "camera": {
+    "opencv_fisheye4": {
+      "K": [
+        3,
+        3
+      ],
+      "distortion": [
+        "k1",
+        "k2",
+        "k3",
+        "k4"
+      ],
+      "projection": "theta=atan(r); theta_d=theta*(1+k1*theta^2+k2*theta^4+k3*theta^6+k4*theta^8); source ReInterHand"
+    },
+    "pinhole": {
+      "K": [
+        3,
+        3
+      ]
+    },
+    "fisheye624": {
+      "parameters": [
+        "f",
+        "cx",
+        "cy",
+        "k0",
+        "k1",
+        "k2",
+        "k3",
+        "k4",
+        "k5",
+        "p0",
+        "p1",
+        "s0",
+        "s1",
+        "s2",
+        "s3"
+      ]
+    },
+    "rational8": {
+      "K": [
+        3,
+        3
+      ],
+      "distortion": [
+        "k1",
+        "k2",
+        "p1",
+        "p2",
+        "k3",
+        "k4",
+        "k5",
+        "k6"
+      ]
+    }
+  },
+  "segmentation": {
+    "segment/uint16_labels/v1": {
+      "storage": "mask.mkv",
+      "dtype": "uint16",
+      "mapping": "instances.jsonl",
+      "valid_shape": [
+        "T"
+      ],
+      "required_metadata": [
+        "segmentation_type",
+        "coverage",
+        "mapping_ref",
+        "storage"
+      ],
+      "shape": [
+        "T",
+        "H",
+        "W"
+      ],
+      "reserved_codes": {
+        "0": "unassigned",
+        "65535": "ignore"
+      },
+      "region_codes": [
+        1,
+        65534
+      ],
+      "allowed_storage_profiles": [
+        "mkv/uint16_ffv1/v1"
+      ]
+    },
+    "segment/instance_polygons/v1": {
+      "storage": "polygons.jsonl",
+      "mapping": "instances.jsonl",
+      "frame_keys": [
+        "frame_index",
+        "instances"
+      ],
+      "instance_keys": [
+        "mask_value",
+        "polygons"
+      ],
+      "coordinate_unit": "pixel",
+      "overlap_allowed": true,
+      "valid_shape": [
+        "T"
+      ],
+      "required_metadata": [
+        "segmentation_type",
+        "coverage",
+        "mapping_ref"
+      ],
+      "polygon_shape": [
+        "N",
+        2
+      ],
+      "holes_supported": false
+    },
+    "segment/foreground_intensity/v1": {
+      "storage": "mask.mkv",
+      "dtype": "uint16",
+      "value_range": [
+        0,
+        255
+      ],
+      "normalization_scale": 0.00392156862745098,
+      "axes": [
+        "T",
+        "height",
+        "width"
+      ],
+      "meaning": "native hand foreground intensity; individual hand or union declared in metadata; not instance codes or calibrated probability",
+      "zero": "background",
+      "full_scale": "foreground",
+      "mapping": "foreground_category in component metadata",
+      "valid_shape": [
+        "T"
+      ],
+      "required_metadata": [
+        "segmentation_type",
+        "coverage",
+        "value_range",
+        "normalization_scale",
+        "foreground_category",
+        "entity_id",
+        "storage"
+      ],
+      "shape": [
+        "T",
+        "H",
+        "W"
+      ],
+      "allowed_storage_profiles": [
+        "mkv/uint16_ffv1/v1"
+      ]
+    }
+  },
+  "note": "Registry for hand-warehouse/1.0. Read CONTRACT.md for field semantics. This is not a JSON Schema document or executable validator.",
+  "components": {
+    "mano/axis_angle61/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        2,
+        61
+      ],
+      "valid_shape": [
+        "T",
+        2
+      ],
+      "unit": "m",
+      "angle_unit": "rad",
+      "coordinate_frame": "camera",
+      "layout": {
+        "pose_axis_angle": [
+          0,
+          48
+        ],
+        "shape": [
+          48,
+          58
+        ],
+        "translation": [
+          58,
+          61
+        ]
+      }
+    },
+    "keypoints/uv/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        2,
+        21,
+        2
+      ],
+      "valid_shape": [
+        "T",
+        2,
+        21
+      ],
+      "unit": "pixel",
+      "required_metadata": [
+        "keypoint_schema"
+      ]
+    },
+    "keypoints/xyz/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        2,
+        21,
+        3
+      ],
+      "valid_shape": [
+        "T",
+        2,
+        21
+      ],
+      "unit": "m",
+      "required_metadata": [
+        "keypoint_schema"
+      ]
+    },
+    "depth/z_float32/v1": {
+      "storage": "mkv",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        "H",
+        "W"
+      ],
+      "valid_shape": [
+        "T",
+        "H",
+        "W"
+      ],
+      "unit": "m",
+      "depth_type": "z",
+      "pixel_grid": "current_rgb",
+      "coordinate_frame": "current_rgb_camera",
+      "valid_storage": "mkv",
+      "allowed_storage_profiles": [
+        "mkv/depth_float32_bits/v1"
+      ],
+      "allowed_valid_storage_profiles": [
+        "mkv/bool_ffv1/v1"
+      ],
+      "required_metadata": [
+        "storage",
+        "valid_storage",
+        "depth_origin"
+      ],
+      "depth_origin_values": [
+        "measured",
+        "rendered",
+        "unknown"
+      ]
+    },
+    "transform/se3/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        4,
+        4
+      ],
+      "valid_shape": [
+        "T"
+      ],
+      "unit": "m",
+      "required_metadata": [
+        "convention"
+      ]
+    },
+    "object/rigid/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        4,
+        4
+      ],
+      "valid_shape": [
+        "T"
+      ],
+      "unit": "m",
+      "required_metadata": [
+        "convention"
+      ]
+    },
+    "camera/pinhole/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        3,
+        3
+      ],
+      "valid_shape": [
+        "T"
+      ]
+    },
+    "camera/rational8/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        3,
+        3
+      ],
+      "valid_shape": [
+        "T"
+      ]
+    },
+    "camera/opencv_fisheye4/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        3,
+        3
+      ],
+      "valid_shape": [
+        "T"
+      ]
+    },
+    "camera/fisheye624/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        15
+      ],
+      "valid_shape": [
+        "T"
+      ]
+    },
+    "camera/distortion8/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        8
+      ],
+      "valid_shape": [
+        "T"
+      ]
+    },
+    "camera/fisheye4_distortion/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        4
+      ],
+      "valid_shape": [
+        "T"
+      ]
+    },
+    "object/joint_state/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        "J"
+      ],
+      "valid_shape": [
+        "T",
+        "J"
+      ],
+      "required_metadata": [
+        "joint_order"
+      ],
+      "unit_rule": "revolute radians; prismatic metres; mixed types declared per joint"
+    },
+    "object_keypoints/camera/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        "K",
+        3
+      ],
+      "valid_shape": [
+        "T",
+        "K"
+      ],
+      "unit": "m",
+      "coordinate_frame": "camera",
+      "required_metadata": [
+        "keypoint_schema"
+      ]
+    },
+    "object/dimensions/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        3
+      ],
+      "valid_shape": [
+        "T"
+      ],
+      "unit": "m",
+      "required_metadata": [
+        "dimensions_semantics"
+      ],
+      "axis_order": [
+        "length",
+        "width",
+        "height"
+      ],
+      "scale_rule": "do not apply to mesh without source-supported size/scale relation"
+    },
+    "object_keypoints/template/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "K",
+        3
+      ],
+      "valid_shape": [
+        "K"
+      ],
+      "unit": "m",
+      "coordinate_frame": "object",
+      "static": true,
+      "required_metadata": [
+        "keypoint_schema"
+      ]
+    },
+    "mesh/obj/v1": {
+      "storage": "obj",
+      "unit": "m",
+      "static": true,
+      "valid_shape": [],
+      "dependencies": "MTL and textures retained under owning object"
+    },
+    "rgb/uint8/v1": {
+      "storage": "mkv",
+      "dtype": "uint8",
+      "shape": [
+        "T",
+        "H",
+        "W",
+        3
+      ],
+      "valid_shape": [
+        "T"
+      ],
+      "color_order": "RGB",
+      "allowed_storage_profiles": [
+        "mkv/mjpeg_copy/v1",
+        "mkv/source_copy/v1",
+        "mkv/h264_rgb_lossless/v1",
+        "mkv/ffv1_rgb/v1"
+      ],
+      "required_metadata": [
+        "storage"
+      ]
+    },
+    "contact/relations/v1": {
+      "storage": "jsonl",
+      "valid_shape": [
+        "T"
+      ],
+      "frame_keys": [
+        "frame_index",
+        "relations"
+      ],
+      "absence_semantics": "no record is not a negative contact annotation"
+    },
+    "contact/vertex_region/v1": {
+      "storage": "pt",
+      "dtype": "bool",
+      "shape": [
+        "T",
+        "V"
+      ],
+      "valid_shape": [
+        "T",
+        "V"
+      ],
+      "required_metadata": [
+        "source_geometry_ref",
+        "target_geometry_ref",
+        "topology_id"
+      ],
+      "pilot_status": "defined_not_empirically_validated"
+    },
+    "contact/vertex_distance/v1": {
+      "storage": "pt",
+      "dtype": "float32",
+      "shape": [
+        "T",
+        "V"
+      ],
+      "valid_shape": [
+        "T",
+        "V"
+      ],
+      "unit": "m",
+      "required_metadata": [
+        "source_geometry_ref",
+        "target_geometry_ref",
+        "topology_id",
+        "distance_definition"
+      ],
+      "pilot_status": "defined_not_empirically_validated"
+    },
+    "object/articulated/v1": {
+      "storage": "json",
+      "required_keys": [
+        "parts",
+        "joints"
+      ],
+      "part_keys": [
+        "part_id",
+        "parent_id",
+        "has_model",
+        "mesh_ref"
+      ],
+      "hierarchy_semantics": "declare mechanical articulation or semantic containment; containment alone does not imply joints",
+      "acyclic": true
+    },
+    "segment/uint16_labels/v1": {
+      "storage": "mkv",
+      "dtype": "uint16",
+      "mapping": "instances.jsonl",
+      "valid_shape": [
+        "T"
+      ],
+      "required_metadata": [
+        "segmentation_type",
+        "coverage",
+        "mapping_ref",
+        "storage"
+      ],
+      "shape": [
+        "T",
+        "H",
+        "W"
+      ],
+      "reserved_codes": {
+        "0": "unassigned",
+        "65535": "ignore"
+      },
+      "region_codes": [
+        1,
+        65534
+      ],
+      "allowed_storage_profiles": [
+        "mkv/uint16_ffv1/v1"
+      ]
+    },
+    "segment/instance_polygons/v1": {
+      "storage": "jsonl",
+      "mapping": "instances.jsonl",
+      "frame_keys": [
+        "frame_index",
+        "instances"
+      ],
+      "instance_keys": [
+        "mask_value",
+        "polygons"
+      ],
+      "coordinate_unit": "pixel",
+      "overlap_allowed": true,
+      "valid_shape": [
+        "T"
+      ],
+      "required_metadata": [
+        "segmentation_type",
+        "coverage",
+        "mapping_ref"
+      ],
+      "polygon_shape": [
+        "N",
+        2
+      ],
+      "holes_supported": false
+    },
+    "segment/foreground_intensity/v1": {
+      "storage": "mkv",
+      "dtype": "uint16",
+      "value_range": [
+        0,
+        255
+      ],
+      "normalization_scale": 0.00392156862745098,
+      "axes": [
+        "T",
+        "height",
+        "width"
+      ],
+      "meaning": "native hand foreground intensity; individual hand or union declared in metadata; not instance codes or calibrated probability",
+      "zero": "background",
+      "full_scale": "foreground",
+      "mapping": "foreground_category in component metadata",
+      "valid_shape": [
+        "T"
+      ],
+      "required_metadata": [
+        "segmentation_type",
+        "coverage",
+        "value_range",
+        "normalization_scale",
+        "foreground_category",
+        "entity_id",
+        "storage"
+      ],
+      "shape": [
+        "T",
+        "H",
+        "W"
+      ],
+      "allowed_storage_profiles": [
+        "mkv/uint16_ffv1/v1"
+      ]
+    }
+  },
+  "storage_revision": "lossless-video-2026-09-18",
+  "sampling": {
+    "image": {
+      "single": {
+        "T": 1
+      },
+      "independent": {
+        "T": ">=2",
+        "required_frame_fields": [
+          "sample_id",
+          "source_locator"
+        ],
+        "temporal_relationship": false
+      },
+      "source_fps": null
+    },
+    "video": {
+      "sampling_types": [
+        "dense",
+        "sparse"
+      ]
+    }
+  },
+  "storage_profiles": {
+    "mkv/mjpeg_copy/v1": {
+      "container": "matroska",
+      "codec": "mjpeg",
+      "method": "stream_copy_from_source_jpeg",
+      "logical_dtype": "uint8",
+      "decoded_pixel_format": "rgb24",
+      "validation": [
+        "source_jpeg_packet_sha256_exact",
+        "all_packet_count_order_and_dimensions_exact",
+        "deterministic_decoded_rgb_sample_exact",
+        "pilot_and_final_selected_whole_sequences_decoded_rgb_exact"
+      ],
+      "validation_policy": {
+        "id": "jpeg-packets-all-rgb-sample16/v1",
+        "sample_count_per_view": 16,
+        "short_views": "all_frames_if_T_less_than_16",
+        "fixed_indices": "0,T-1,T//4,T//2,3*T//4, deduplicated",
+        "remaining_indices": "sample without replacement to min(T,16) using Python random.Random(seed)",
+        "seed": "first 8 bytes of SHA256(sequence_id/camera_id/20260920), big-endian integer",
+        "report": "policy, exact selected frame indices, seed, separate packet and pixel coverage",
+        "legacy_all_frame_pixel_proofs": "still accepted; preserve historical evidence"
+      }
+    },
+    "mkv/source_copy/v1": {
+      "container": "matroska",
+      "codec": "source_codec_declared_in_metadata",
+      "method": "stream_copy_only",
+      "logical_dtype": "uint8",
+      "decoded_pixel_format": "rgb24",
+      "validation": [
+        "no_reencoding",
+        "all_decoded_rgb_frames_exact",
+        "presentation_frame_order_exact"
+      ]
+    },
+    "mkv/h264_rgb_lossless/v1": {
+      "container": "matroska",
+      "codec": "h264",
+      "encoder": "libx264rgb",
+      "crf": 0,
+      "decoded_pixel_format": "rgb24",
+      "logical_dtype": "uint8",
+      "gop": "declared_in_encoder_parameters",
+      "validation": [
+        "all_decoded_rgb_frames_exact"
+      ],
+      "forbidden": [
+        "lossy_rate_control",
+        "rgb_to_yuv420_subsampling"
+      ]
+    },
+    "mkv/ffv1_rgb/v1": {
+      "container": "matroska",
+      "codec": "ffv1",
+      "decoded_pixel_format": "rgb24",
+      "logical_dtype": "uint8",
+      "validation": [
+        "all_decoded_rgb_frames_exact"
+      ]
+    },
+    "mkv/depth_float32_bits/v1": {
+      "container": "matroska",
+      "codec": "ffv1",
+      "pixel_format": "gray16le",
+      "logical_dtype": "float32",
+      "logical_shape": [
+        "T",
+        "H",
+        "W"
+      ],
+      "encoded_shape": [
+        "T",
+        "H",
+        "2*W"
+      ],
+      "packing": "IEEE754 little-endian float32 bits; low uint16 in first W columns, high uint16 in last W columns",
+      "decode": "reinterpret_float32(uint32(low) | (uint32(high) << 16))",
+      "validation": [
+        "all_float32_bits_exact_including_nan"
+      ]
+    },
+    "mkv/bool_ffv1/v1": {
+      "container": "matroska",
+      "codec": "ffv1",
+      "pixel_format": "gray",
+      "logical_dtype": "bool",
+      "stored_values": [
+        0,
+        1
+      ],
+      "validation": [
+        "all_bool_values_exact"
+      ]
+    },
+    "mkv/uint16_ffv1/v1": {
+      "container": "matroska",
+      "codec": "ffv1",
+      "pixel_formats": [
+        "gray",
+        "gray16le"
+      ],
+      "logical_dtype": "uint16",
+      "gray_condition": "all values <=255; decode by exact uint8-to-uint16 promotion",
+      "validation": [
+        "all_integer_values_exact"
+      ]
+    }
+  },
+  "video_metadata": {
+    "required": [
+      "profile",
+      "codec",
+      "pixel_format",
+      "encoded_size",
+      "decoder",
+      "encoder_parameters",
+      "source_lossy"
+    ],
+    "source_lossy_values": [
+      true,
+      false,
+      "unknown"
+    ],
+    "timebase_role": "index_only",
+    "lossless_reference": "source compressed packets or fixed-decoder source pixels; depth uses normalized standard float32 bits",
+    "selection": "choose smaller complete candidate only after exact roundtrip validation"
+  }
+}

H2O/shared-v1.0/script/provenance/previous-run/script/provenance/previous-run/full.log

@@ -0,0 +1,8 @@
+Verified existing subject1/h1/0
+Verified existing subject2/h1/0
+Verified existing subject1/h1/6
+Verified existing subject1/h1/3
+Verified existing subject2/h1/6
+Converted subject3/h1/0
+Verified existing subject1/h1/7
+Converted subject4/h1/0

H2O/shared-v1.0/script/provenance/previous-run/script/provenance/previous-run/full.status

@@ -0,0 +1,7 @@
+START=2026-09-30T10:22:30+08:00
+WORKERS=4 CODEC_THREADS=10 SUBJECTS=1,2,3,4 CASA=false
+CPU_LIMIT=none
+CGROUP=/user.slice/user-25044.slice/user@25044.service/app.slice/h2o-full-w4c10-r2-20260930.scope
+CPU_MAX=max controller_disabled
+
+RECONFIGURE=2026-09-30T11:03:03+08:00 REASON=decoder10_depth4_no_resume_png_writes

H2O/shared-v1.0/script/provenance/previous-run/script/provenance/previous-run/run-full.sh

@@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+set -u
+base="$HOME/Datasets/uni-hoi-h2o-full-20260930-w4-c10"
+raw="$HOME/Datasets/Hand-Datasets-Raw/H2O"
+cd "$HOME/Projects/uni-hoi-dataset" || exit 1
+mkdir -p "$base/scratch"
+export TMPDIR="$base/scratch" PYTHONUNBUFFERED=1
+for n in 1 2 3 4; do test -s "$raw/raw/data/subject${n}_v1_1.tar.gz" || exit 2; done
+test ! -e "$base/full.status" || exit 3
+cg=$(awk -F: '$1=="0" {print $3}' /proc/self/cgroup)
+if [[ -f "/sys/fs/cgroup$cg/cpu.max" ]]; then
+    read -r quota period < "/sys/fs/cgroup$cg/cpu.max"
+else
+    quota=max
+    period=controller_disabled
+fi
+printf 'START=%s\nWORKERS=4 CODEC_THREADS=10 SUBJECTS=1,2,3,4 CASA=false\nCPU_LIMIT=none\nCGROUP=%s\nCPU_MAX=%s %s\n' "$(date -Is)" "$cg" "$quota" "$period" > "$base/full.status"
+if [[ "$quota" != max ]]; then
+    printf 'EXIT=4 ERROR=UNEXPECTED_CPU_QUOTA\n' >> "$base/full.status"
+    exit 4
+fi
+/usr/bin/time -v .venv/bin/python "$base/frozen-code/scripts/convert_h2o_v1.0.py" --root "$raw" --out "$base" --release-id shared-v1.0 --workers 4 --codec-threads 10 --mano-dir "$HOME/Projects/WiLoR/mano_data" > "$base/full.log" 2>&1
+code=$?
+printf 'EXIT=%s END=%s\n' "$code" "$(date -Is)" >> "$base/full.status"
+if [[ -f "/sys/fs/cgroup$cg/cpu.stat" ]]; then cat "/sys/fs/cgroup$cg/cpu.stat" >> "$base/full.status"; fi
+exit "$code"

H2O/shared-v1.0/script/provenance/previous-run/script/provenance/previous-run/script/check_dexycb_projection.py

@@ -0,0 +1,31 @@
+#!/usr/bin/env python3
+"""Check saved Dex-YCB source 3D joints against saved source 2D joints in sampled views."""
+import json
+from pathlib import Path
+import sys
+
+import numpy as np
+import torch
+
+root = Path(sys.argv[1])
+report = json.loads((root / 'doc/validation/STRUCTURAL_AUDIT.json').read_text())
+checks = []
+for name in report['sampled_views']:
+    view = root / name
+    load = lambda path: torch.load(view / path, weights_only=True, map_location='cpu').numpy()
+    k = load('camera/intrinsic.pt')
+    xyz = load('annotation/keypoints/xyz.pt')
+    uv = load('annotation/keypoints/uv.pt')
+    valid = load('annotation/keypoints/xyz_valid.pt') & load('annotation/keypoints/uv_valid.pt')
+    z = xyz[..., 2]
+    valid &= np.isfinite(xyz).all(axis=-1) & np.isfinite(uv).all(axis=-1) & (z > 0)
+    z = np.where(valid, z, 1)
+    pred = np.stack((k[:, 0, 0, None, None] * xyz[..., 0] / z + k[:, 0, 2, None, None],
+                     k[:, 1, 1, None, None] * xyz[..., 1] / z + k[:, 1, 2, None, None]), axis=-1)
+    errors = np.linalg.norm(pred[valid] - uv[valid], axis=-1)
+    checks.append(dict(view=name, points=int(errors.size), median_px=float(np.median(errors)) if errors.size else None,
+                       p95_px=float(np.percentile(errors, 95)) if errors.size else None))
+(root / 'doc/validation/KEYPOINT_PROJECTION_QC.json').write_text(json.dumps({
+    'scope': 'Saved source 3D joints projected with saved color intrinsics vs saved source 2D joints; independent of hand mesh rendering.',
+    'views': checks}, ensure_ascii=False, indent=2) + '\n')
+print(json.dumps(checks, ensure_ascii=False), flush=True)

H2O/shared-v1.0/script/provenance/previous-run/script/provenance/previous-run/script/config.json

@@ -0,0 +1,15 @@
+{
+  "root": "/home/xuzhefeng/Datasets/Hand-Datasets-Raw/H2O",
+  "out": "/home/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-w4-c10",
+  "release_id": "shared-v1.0",
+  "limit": null,
+  "max_frames": null,
+  "mano_dir": "/home/xuzhefeng/Projects/WiLoR/mano_data",
+  "views": null,
+  "jpeg_full_check": false,
+  "workers": 4,
+  "codec_threads": 10,
+  "subjects": null,
+  "include_casa": false,
+  "archive": null
+}

H2O/shared-v1.0/script/provenance/previous-run/script/provenance/previous-run/script/convert_dexycb_v1.0.py

@@ -0,0 +1,349 @@
+#!/usr/bin/env python3
+"""Dex-YCB source annotations → CONTRACT.md v1.0. No derived annotations."""
+import io
+import re
+import shutil
+import tarfile
+import tempfile
+from pathlib import Path
+
+import numpy as np
+from PIL import Image
+import yaml
+
+from warehouse_v1_0 import (View, build_record, find_file, locator, mano_models, mesh_asset,
+    model_profile, parser, pca_to_axis_angle, release, roots, safe_member, sequence_dir)
+
+IMG_W, IMG_H = 640, 480
+
+class _Loader(yaml.SafeLoader):
+    pass
+
+
+_Loader.add_constructor("tag:yaml.org,2002:python/tuple",
+                        lambda l, n: tuple(l.construct_sequence(n)))
+
+
+def yaml_load_bytes(data: bytes):
+    return yaml.load(io.BytesIO(data), Loader=_Loader)
+
+
+class Calibration:
+    def __init__(self, dexycb_root: Path):
+        self.intrinsics = {}   # serial -> {fx,fy,cx,cy,w,h}
+        self.extrinsics = {}   # group  -> {serial: [12] 行优先 3×4 cam→world}
+        self.betas = {}        # mano_calib_name -> [10]
+        calib_dir = dexycb_root / "calibration"
+        if calib_dir.is_dir():
+            self._load((p.as_posix(), p.read_bytes())
+                       for p in sorted(calib_dir.rglob("*")) if p.is_file())
+        else:
+            tar_path = dexycb_root / "calibration.tar.gz"
+            if not tar_path.exists():
+                raise FileNotFoundError(f"no calibration/ nor calibration.tar.gz in {dexycb_root}")
+            with tarfile.open(tar_path, "r:gz") as tf:
+                self._load((m.name, tf.extractfile(m).read()) for m in tf if m.isfile())
+
+    _RE_INTR = re.compile(r"(?:^|/)intrinsics/(\d+)_\d+x\d+\.yml$")
+    _RE_EXTRA = re.compile(r"(?:^|/)extrinsics_(\d{8}_\d{6})/extrinsics\.yml$")
+    _RE_MANO = re.compile(r"(?:^|/)mano_(\d{8}_\d{6}_subject-\d{2}_right)/mano\.yml$")
+
+    def _load(self, files):
+        for rel, data in files:
+            m = self._RE_INTR.search(rel)
+            if m:
+                c = yaml_load_bytes(data)["color"]
+                self.intrinsics[m.group(1)] = {"fx": float(c["fx"]), "fy": float(c["fy"]),
+                                               "cx": float(c["ppx"]), "cy": float(c["ppy"]),
+                                               "w": IMG_W, "h": IMG_H}
+                continue
+            m = self._RE_EXTRA.search(rel)
+            if m:
+                raw = yaml_load_bytes(data)["extrinsics"]
+                self.extrinsics[m.group(1)] = {
+                    str(s): [float(x) for x in np.asarray(v).ravel()] for s, v in raw.items()}
+                continue
+            m = self._RE_MANO.search(rel)
+            if m:
+                self.betas[m.group(1)] = [float(b) for b in yaml_load_bytes(data)["betas"]]
+
+
+
+def calibration_locator(path, member, **extra):
+    if path.is_dir():
+        return locator(path / member.removeprefix('calibration/'), **extra)
+    return locator(path, member, **extra)
+
+
+def prepare_models(root, destination):
+    archive = find_file(root, 'models.tar.gz')
+    if archive is None:
+        return {}, None
+    pattern = re.compile(r'^models/(\d{3}_[a-z0-9_]+)/(textured_simple\.obj(?:\.mtl)?|texture_map\.png)$')
+    names = set()
+    with tarfile.open(archive, 'r|gz') as tf:
+        for member in tf:
+            match = pattern.fullmatch(member.name)
+            if member.isfile() and match:
+                names.add(match[1])
+                path = safe_member(destination, '/'.join(match.groups()))
+                path.parent.mkdir(parents=True, exist_ok=True)
+                path.write_bytes(tf.extractfile(member).read())
+    if len(names) != 21:
+        raise ValueError('expected all 21 source YCB model directories')
+    return {i: destination / name for i, name in enumerate(sorted(names), 1)}, archive
+
+
+def convert_view(path, args, seq, serial, frames, meta, meta_loc, calib, calib_path, models, assets, asset_archive):
+    total = int(meta.get('num_frames', max(frames) + 1))
+    ids = list(range(min(total, args.max_frames) if args.max_frames else total))
+    for f in ids:
+        frames.setdefault(f, {})
+    if not ids:
+        raise ValueError('no RGB frames')
+    source_rows = []
+    for f in ids:
+        row = frames[f]
+        source_rows.append(dict(source_frame_id=f, sync_id=f'{seq}:{f}',
+            source_locator=row['color'][1] if 'color' in row else (meta_loc | {'source_frame_id': f, 'missing_rgb': True}), annotation_matches={
+                'labels': {'source_frame_id': f, 'source_timestamp_ns': None, 'clock_domain': None,
+                           'match_method': 'index_map', 'delta_ns': None, 'source_locator': row['labels'][1]}
+                for _ in [0] if 'labels' in row},
+            component_sources={k: v[1] for k, v in row.items()}))
+    view = View(path, 'dex-ycb', args.release_id, seq, serial, source_rows,
+                sampling='dense' if ids == list(range(ids[0], ids[-1] + 1)) else 'sparse')
+    view.valid_rules = {
+        'mano': 'Source pose_m present and not all zero; side model/calibration available',
+        'keypoints': 'Corresponding joint_2d/joint_3d present; whole-hand all -1 is missing, broadcast to its 21 points',
+        'pose': 'Source pose_y for this ycb_ids slot present and not an all-zero matrix',
+        'segment': 'Source seg array provided for this frame; region codes and entity references validated',
+        'relations': 'Source sequence ycb_grasp_ind resolves through ycb_ids; broadcast sequence grasp to selected frames',
+    }
+    view.full_jpeg = getattr(args, 'jpeg_full_check', False)
+    view.meta['view_type'] = 'exo'
+    view.meta['sampling']['selection'] = {'max_frames': args.max_frames, 'views': args.views}
+    subject = re.search(r'subject-(\d+)', seq)
+    if subject and 1 <= int(subject[1]) <= 10:
+        index = int(subject[1])
+        view.meta.update(split='val' if index == 7 else 'test' if index in (8, 9) else 'train',
+            split_scheme='Dex-YCB/s1', split_source_locator={
+                'url': 'https://github.com/NVlabs/dex-ycb-toolkit/blob/master/dex_ycb_toolkit/dex_ycb.py',
+                'rule': 's1 unseen-subject: 01-06/10 train, 07 val, 08-09 test'})
+        view.meta.pop('split_reason', None)
+    view.rgb([frames[f]['color'][0] if 'color' in frames[f] else None for f in ids], {'frames_ref': 'camera/frames.jsonl'})
+    n = len(ids)
+    intr = calib.intrinsics.get(serial)
+    K = None if intr is None else np.array([[intr['fx'], 0, intr['cx']], [0, intr['fy'], intr['cy']], [0, 0, 1]])
+    ext = calib.extrinsics.get(str(meta['extrinsics']), {}).get(serial)
+    E = None
+    if ext is not None:
+        E = np.eye(4)
+        E[:3] = np.asarray(ext).reshape(3, 4)
+        E = np.linalg.inv(E)
+    view.camera(K, calibration_locator(calib_path, f'calibration/intrinsics/{serial}_640x480.yml'), E,
+        calibration_locator(calib_path, f'calibration/extrinsics_{meta["extrinsics"]}/extrinsics.yml',
+                conversion='inverse of source T_world_from_camera'))
+    sides = meta.get('mano_sides')
+    if sides not in (['left'], ['right']):
+        raise ValueError(f'expected exactly one anatomical source hand: {sides}')
+    side = sides[0]
+    slot = ['left', 'right'].index(side)
+    cal_name = meta['mano_calib'][0] if isinstance(meta['mano_calib'], list) else meta['mano_calib']
+    labels = [dict(np.load(frames[f]['labels'][0])) if 'labels' in frames[f] else {} for f in ids]
+    label_loc = {'frames_ref': 'camera/frames.jsonl', 'source_field': 'component_sources.labels'}
+    motion = np.full((n, 2, 61), np.nan, np.float32)
+    mv = np.zeros((n, 2), bool)
+    if side in models and any('pose_m' in lab for lab in labels):
+        for i, lab in enumerate(labels):
+            if 'pose_m' not in lab:
+                continue
+            pose = np.asarray(lab['pose_m']).reshape(51)
+            if not np.any(pose):  # Official Dex-YCB missing-pose sentinel, independent of joints/seg.
+                continue
+            motion[i, slot] = np.r_[pose[:3], pca_to_axis_angle(pose[3:48], models[side]),
+                                     calib.betas[cal_name], pose[48:51]]
+            mv[i, slot] = True
+        view.tensor('mano', 'mano/axis_angle61/v1', 'annotation/mano/motion.pt', motion, mv,
+                    ['T', 'hand', 'parameter'], label_loc, model_profile=model_profile(models),
+                    calibration_source=calibration_locator(calib_path, f'calibration/mano_{cal_name}/mano.yml'),
+                    provenance={'kind': 'format_conversion', 'method': 'source PCA @ hands_components + hands_mean; explicit axis-angle consumed with flat_hand_mean=true'})
+    elif any('pose_m' in lab for lab in labels):
+        view.absent('mano', 'Source PCA exists; provide --mano-dir to expand it', 'dependency_missing')
+    else:
+        view.absent('mano', 'No source pose_m field')
+    for source_key, key, dim, unit in [('joint_2d', 'uv', 2, 'pixel'), ('joint_3d', 'xyz', 3, 'm')]:
+        if not any(source_key in x for x in labels):
+            view.absent('keypoints_' + key, 'No source ' + source_key)
+            continue
+        points = np.full((n, 2, 21, dim), np.nan, np.float32)
+        valid = np.zeros((n, 2, 21), bool)
+        for i, lab in enumerate(labels):
+            if source_key not in lab:
+                continue
+            p = np.asarray(lab[source_key]).reshape(21, dim)
+            points[i, slot] = p
+            valid[i, slot] = not np.all(p == -1)  # Official sentinel covers the whole source hand.
+            # An individual off-image point (-1, -1) is not evidence of missing annotation.
+        view.tensor('keypoints_' + key, f'keypoints/{key}/v1', f'annotation/keypoints/{key}.pt',
+                    points, valid, ['T', 'hand', 'keypoint', 'coordinate'], label_loc,
+                    keypoint_schema='mano21/v1', unit=unit)
+    ycb_ids = [int(x) for x in meta['ycb_ids']]
+    for j, oid in enumerate(ycb_ids):
+        entity = f'ycb_{oid:02d}'
+        view.object(entity, meta_loc, source_object_id=oid, category_id=oid)
+        poses = np.full((n, 4, 4), np.nan, np.float32)
+        valid = np.zeros(n, bool)
+        for i, lab in enumerate(labels):
+            if 'pose_y' in lab:
+                pose = np.asarray(lab['pose_y'])
+                if pose.shape != (len(ycb_ids), 3, 4):
+                    raise ValueError('pose_y shape differs from ycb_ids')
+                if np.any(pose[j]):  # Official all-zero missing object pose sentinel.
+                    poses[i] = np.eye(4)
+                    poses[i, :3] = pose[j]
+                    valid[i] = True
+        if any('pose_y' in lab for lab in labels):
+            view.tensor('pose_' + entity, 'object/rigid/v1', f'annotation/object/{entity}/pose.pt',
+                    poses, valid, ['T', 'row', 'column'], label_loc,
+                    convention='T_camera_from_object', unit='m')
+        else:
+            view.absent('pose_' + entity, 'No source pose_y field')
+        if oid in assets:
+            mesh_asset(view, entity, assets[oid], assets[oid] / 'textured_simple.obj',
+                locator(asset_archive, f'models/{assets[oid].name}/textured_simple.obj'))
+        else:
+            view.absent('mesh_' + entity, 'models.tar.gz unavailable', 'not_downloaded')
+    if any('seg' in lab for lab in labels):
+        mask_rows = []
+        width, height = view.meta['camera']['image_size']
+        for i, lab in enumerate(labels):
+            values = np.unique(lab.get('seg', np.zeros((height, width), np.uint8)))
+            mappings = []
+            for value in values:
+                value = int(value)
+                if value == 0:
+                    continue
+                if value != 255 and value not in ycb_ids:
+                    raise ValueError(f'unknown Dex-YCB region value {value}')
+                mappings.append(dict(mask_value=value, category_id='hand' if value == 255 else value,
+                    entity_id='hand_' + side if value == 255 else f'ycb_{value:02d}', part_id=None,
+                    source_instance_id=value))
+            mask_rows.append(dict(frame_index=i, labels=mappings))
+        view.raster('segment', (lab.get('seg', np.zeros((height, width), np.uint16)) for lab in labels),
+                    ['seg' in lab for lab in labels], label_loc, mask_rows)
+        view.meta['components']['segment'].update(coverage={'kind': 'selected_entities',
+            'entities': ['hand_' + side] + [f'ycb_{o:02d}' for o in ycb_ids]}, source_zero_meaning='background')
+    else:
+        view.absent('segment', 'No source seg field')
+    if any('aligned_depth_to_color' in frames[f] for f in ids):
+        def depths():
+            width, height = view.meta['camera']['image_size']
+            for f in ids:
+                item = frames[f].get('aligned_depth_to_color')
+                if item is None:
+                    yield np.full((height, width), np.nan, np.float32)
+                else:
+                    with Image.open(item[0]) as im:
+                        yield np.asarray(im).astype(np.float32) / 1000.0
+        view.raster('depth', depths(), None, {'frames_ref': 'camera/frames.jsonl',
+            'source_field': 'component_sources.aligned_depth_to_color', 'conversion': 'millimetres to metres'})
+    else:
+        view.absent('depth', 'No source aligned_depth_to_color frames')
+    grasp = meta.get('ycb_grasp_ind')
+    if grasp is not None:
+        if not 0 <= int(grasp) < len(ycb_ids):
+            raise ValueError('invalid ycb_grasp_ind')
+        relation = dict(relation_type='grasp', subject_id='hand_' + side,
+            object_id=f'ycb_{ycb_ids[int(grasp)]:02d}', status='positive', source_locator=meta_loc,
+            provenance={'kind': 'source', 'method': 'sequence ycb_grasp_ind resolved through ycb_ids; not per-frame contact'})
+        view.relations([{'frame_index': i, 'relations': [relation]} for i in range(n)], np.ones(n, bool), meta_loc)
+    else:
+        view.absent('relations', 'No source grasp/contact relation')
+    return view.finish()
+
+
+MEMBER = re.compile(r'^([^/]+)/(\d{8}_\d{6})/(meta\.yml|(\d{12})/(color|labels|aligned_depth_to_color)_(\d+)\.(jpg|npz|png))$')
+
+
+def convert_archive(archive, args, out, calib, calib_path, models, assets, asset_archive):
+    selected_views = set(args.views.split(',')) if args.views else None
+    done = 0
+    current, meta, meta_loc, rows = None, None, None, {}
+    with tempfile.TemporaryDirectory(prefix='dex-source-') as wd:
+        work = Path(wd)
+        def flush():
+            nonlocal done
+            if not rows:
+                return
+            seq = archive.name.split('.tar')[0] + '_' + current
+            reports = {}
+            with sequence_dir(out, seq) as target:
+                for serial, frames in sorted(rows.items()):
+                    reports[serial] = convert_view(target / serial, args, seq, serial, frames, meta,
+                                                   meta_loc, calib, calib_path, models, assets, asset_archive)
+            build_record(out, seq, reports)
+            for staged in list(work.iterdir()):
+                if staged.is_dir():
+                    shutil.rmtree(staged)
+                else:
+                    staged.unlink()
+            done += 1
+            print(f'Converted {seq}', flush=True)
+        with tarfile.open(archive, 'r|gz') as tf:
+            for member in tf:
+                match = MEMBER.fullmatch(member.name)
+                if not member.isfile() or not match:
+                    continue
+                seq = match[2]
+                if seq != current:
+                    flush()
+                    if args.limit and done >= args.limit:
+                        break
+                    current, meta, meta_loc, rows = seq, None, None, {}
+                data_kind = match[3]
+                if data_kind == 'meta.yml':
+                    data = tf.extractfile(member).read()
+                    meta = yaml_load_bytes(data)
+                    meta_loc = locator(archive, member.name, data)
+                    continue
+                serial, kind, f = match[4], match[5], int(match[6])
+                if selected_views and serial not in selected_views or args.max_frames and f >= args.max_frames:
+                    continue
+                data = tf.extractfile(member).read()
+                dest = safe_member(work, member.name)
+                dest.parent.mkdir(parents=True, exist_ok=True)
+                dest.write_bytes(data)
+                row = rows.setdefault(serial, {}).setdefault(f, {})
+                if kind in row:
+                    raise ValueError('duplicate archive frame')
+                row[kind] = (dest, locator(archive, member.name, data))
+            else:
+                flush()
+    return done
+
+
+def main(argv=None):
+    p = parser('dex-ycb', __doc__)
+    p.add_argument('--subjects')
+    args = p.parse_args(argv)
+    archives = sorted({a for r in roots(args.root) for a in r.glob('*-subject-*.tar.gz')})
+    if args.subjects:
+        subjects = {s.strip().zfill(2) for s in args.subjects.split(',')}
+        archives = [a for a in archives if re.search(r'subject-(\d+)', a.name)[1] in subjects]
+    if not archives:
+        p.error('no Dex-YCB subject archives found')
+    cp = find_file(args.root, 'calibration.tar.gz') or find_file(args.root, 'calibration')
+    if cp is None:
+        p.error('calibration source missing')
+    calib = Calibration(cp.parent)
+    models = mano_models(args.mano_dir)
+    out = release(args, 'dex-ycb')
+    with tempfile.TemporaryDirectory(prefix='dex-models-') as wd:
+        assets, asset_archive = prepare_models(args.root, Path(wd))
+        for archive in archives:
+            convert_archive(archive, args, out, calib, cp, models, assets, asset_archive)
+    return 0
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())

H2O/shared-v1.0/script/provenance/previous-run/script/provenance/previous-run/script/convert_h2o_v1.0.py

@@ -0,0 +1,527 @@
+#!/usr/bin/env python3
+"""H2O native RGB-D, MANO, joints and objects → CONTRACT.md v1.0 (no rendered masks)."""
+from concurrent.futures import ThreadPoolExecutor
+import json
+import os
+import re
+import shutil
+import tarfile
+import tempfile
+from pathlib import Path
+import zipfile
+
+# Bound native libraries before loading NumPy/Torch; each worker still uses independent FFmpeg processes.
+for key in ('OMP_NUM_THREADS', 'OPENBLAS_NUM_THREADS', 'MKL_NUM_THREADS', 'NUMEXPR_NUM_THREADS'):
+    os.environ[key] = '1'
+
+import numpy as np
+from PIL import Image
+import torch
+
+from warehouse_v1_0 import (View, build_record, check_mesh, find_file, locator, mano_models, mesh_asset,
+    model_profile, parser, release, roots, safe_member, sequence_dir, sha256)
+
+OBJECTS = {1: 'book', 2: 'espresso', 3: 'lotion', 4: 'spray',
+           5: 'milk', 6: 'cocoa', 7: 'chips', 8: 'cappuccino'}
+MEMBER = re.compile(r'^(?:h2o_CASA/)?(subject\d+)/([^/]+)/(\d+)/(cam\d)/(.+)$')
+FRAME = re.compile(r'(rgb|depth|cam_pose|hand_pose|hand_pose_mano|obj_pose|obj_pose_rt)/(\d+)\.(png|jpg|txt)$')
+
+
+def load_splits(root):
+    archive = find_file(root, 'label_split.zip')
+    result = {}
+    if archive is not None:
+        with zipfile.ZipFile(archive) as z:
+            for split in ('train', 'val', 'test'):
+                member = f'label_split/action_{split}.txt'
+                for line_no, line in enumerate(z.read(member).decode().splitlines(), 1):
+                    fields = line.split()
+                    if len(fields) < 2 or fields[1] == 'path':
+                        continue
+                    key = fields[1]
+                    if key in result and result[key][0] != split:
+                        raise ValueError('conflicting official splits')
+                    result[key] = (split, locator(archive, member, record=line_no))
+    return result
+
+
+def load_assets(root, work):
+    archive = find_file(root, 'object.zip')
+    if archive is None:
+        return {}, None
+    with zipfile.ZipFile(archive) as z:
+        for member in z.infolist():
+            if member.is_dir():
+                continue
+            path = safe_member(work, member.filename)
+            path.parent.mkdir(parents=True, exist_ok=True)
+            path.write_bytes(z.read(member))
+    result = {}
+    for oid, name in OBJECTS.items():
+        directory = work / 'object' / name
+        aliases = []
+        # Source cocoa.obj -> ovomaltine.mtl -> cocoa.png, but the ZIP names the
+        # supplied texture ovomaltine.png. Keep all originals and add a byte-identical alias.
+        if name == 'cocoa' and not (directory / 'cocoa.png').exists():
+            mtl = directory / 'ovomaltine.mtl'
+            if mtl.is_file() and 'map_Kd cocoa.png' in mtl.read_text().splitlines():
+                texture = directory / 'ovomaltine.png'
+                shutil.copyfile(texture, directory / 'cocoa.png')
+                aliases.append({'file': 'cocoa.png', 'sha256': sha256(texture),
+                                'source': locator(archive, 'object/cocoa/ovomaltine.png')})
+        candidates = list(directory.glob('*.obj'))
+        if len(candidates) != 1:
+            raise ValueError(f'expected one source OBJ for {name}')
+        check_mesh(candidates[0], directory)  # Fail before decoding any large subject archive.
+        result[oid] = (directory, candidates[0], aliases)
+    return result, archive
+
+
+def convert_view(path, args, seq, cam, rows, intr, splits, models, assets, asset_archive):
+    codec_threads = getattr(args, 'codec_threads', 1)
+    ids = list(range(min(rows), max(rows) + 1))
+    for f in ids:
+        rows.setdefault(f, {})
+    n = len(ids)
+    if not n:
+        raise ValueError('no RGB frames')
+    split, split_loc = splits.get(seq, ('unknown', None))
+    frame_rows = [dict(source_frame_id=f, sync_id=f'{seq}:{f}', source_locator=rows[f]['rgb'][1] if 'rgb' in rows[f] else {'sequence': seq, 'camera': cam, 'source_frame_id': f, 'missing_rgb': True},
+                       component_sources={k: item[1] for k, item in rows[f].items()},
+                       annotation_matches={k: dict(source_frame_id=f, source_timestamp_ns=None,
+                           clock_domain=None, match_method='index_map', delta_ns=None)
+                           for k in rows[f] if k != 'rgb'}) for f in ids]
+    view = View(path, 'H2O', args.release_id, seq.replace('/', '_'), cam, frame_rows, split=split)
+    view.valid_rules = {
+        'mano': 'Per-hand hand_pose_mano flag equals 1; corresponding side model available',
+        'keypoints': "Per-hand hand_pose flag equals 1, broadcast to that hand's 21 source points independently of MANO",
+        'pose': 'Source obj_pose_rt class equals this nonzero object class',
+        'points': 'Source obj_pose class equals this nonzero object class, broadcast to its 21 points',
+    }
+    view.full_jpeg = getattr(args, 'jpeg_full_check', False)
+    view.meta['view_type'] = 'ego' if cam == 'cam4' else 'exo' if cam in ('cam0', 'cam1', 'cam2', 'cam3') else 'unknown'
+    view.meta['sampling']['selection'] = {'max_frames': args.max_frames, 'views': args.views}
+    if split_loc:
+        view.meta.update(split_scheme='official_action_split', split_source_locator=split_loc)
+    # In the measured 275 H2O views, exact FFV1 was always smaller than exact H.264.
+    view.rgb([rows[f]['rgb'][0] if 'rgb' in rows[f] else None for f in ids],
+             {'frames_ref': 'camera/frames.jsonl'}, ffv1_first=True, codec_threads=codec_threads)
+    loc = {'frames_ref': 'camera/frames.jsonl', 'source_field': 'component_sources'}
+    if intr is not None:
+        fx, fy, cx, cy, width, height = np.loadtxt(intr[0]).ravel()
+        if [int(width), int(height)] != view.meta['camera']['image_size']:
+            raise ValueError('source intrinsics resolution mismatch')
+        K = np.array([[fx, 0, cx], [0, fy, cy], [0, 0, 1]])
+        view.camera(K, intr[1])
+    camps = np.full((n, 4, 4), np.nan, np.float32)
+    cv = np.zeros(n, bool)
+    data = []
+    for i, f in enumerate(ids):
+        lab = {k: np.loadtxt(item[0]).ravel() for k, item in rows[f].items() if k not in ('rgb', 'depth')}
+        data.append(lab)
+        if 'cam_pose' in lab:
+            camps[i] = np.linalg.inv(lab['cam_pose'].reshape(4, 4))
+            cv[i] = True
+    if cv.any():
+        view.tensor('extrinsic', 'transform/se3/v1', 'camera/extrinsic.pt', camps, cv,
+                    ['T', 'row', 'column'], loc, convention='T_camera_from_world',
+                    source_convention='T_world_from_camera')
+    motion, mv = np.full((n, 2, 61), np.nan, np.float32), np.zeros((n, 2), bool)
+    points, pv = np.full((n, 2, 21, 3), np.nan, np.float32), np.zeros((n, 2, 21), bool)
+    object_ids = set()
+    for i, lab in enumerate(data):
+        if 'hand_pose_mano' in lab:
+            hands = lab['hand_pose_mano'].reshape(2, 62)
+            for side, row in enumerate(hands):
+                if row[0] not in (0, 1):
+                    raise ValueError('unexpected H2O MANO presence flag')
+                if row[0] == 1:
+                    motion[i, side] = np.r_[row[4:52], row[52:62], row[1:4]]
+                    mv[i, side] = True
+        if 'hand_pose' in lab:
+            hands = lab['hand_pose'].reshape(2, 64)
+            for side, row in enumerate(hands):
+                if row[0] not in (0, 1):
+                    raise ValueError('unexpected H2O keypoint presence flag')
+                if row[0] == 1:
+                    points[i, side] = row[1:].reshape(21, 3)
+                    pv[i, side] = True
+        for key in ('obj_pose', 'obj_pose_rt'):
+            if key in lab and lab[key][0] != 0:
+                object_ids.add(int(lab[key][0]))
+    if any('hand_pose_mano' in lab for lab in data):
+        if len(models) == 2:
+            view.tensor('mano', 'mano/axis_angle61/v1', 'annotation/mano/motion.pt', motion, mv,
+                        ['T', 'hand', 'parameter'], loc, model_profile=model_profile(models))
+        else:
+            view.absent('mano', 'Source parameters exist; --mano-dir must identify both model profiles', 'dependency_missing')
+    if any('hand_pose' in lab for lab in data):
+        view.tensor('keypoints_xyz', 'keypoints/xyz/v1', 'annotation/keypoints/xyz.pt', points, pv,
+                    ['T', 'hand', 'keypoint', 'coordinate'], loc, keypoint_schema='mano21/v1', unit='m')
+    for oid in sorted(object_ids):
+        if oid not in OBJECTS:
+            raise ValueError(f'unknown H2O object class {oid}')
+        entity = OBJECTS[oid]
+        view.object(entity, loc, category_id=oid, source_object_id=oid)
+        poses, ov = np.full((n, 4, 4), np.nan, np.float32), np.zeros(n, bool)
+        pts, kv = np.full((n, 21, 3), np.nan, np.float32), np.zeros((n, 21), bool)
+        for i, lab in enumerate(data):
+            if 'obj_pose_rt' in lab and lab['obj_pose_rt'][0] == oid:
+                poses[i] = lab['obj_pose_rt'][1:].reshape(4, 4)
+                ov[i] = True
+            if 'obj_pose' in lab and lab['obj_pose'][0] == oid:
+                pts[i] = lab['obj_pose'][1:].reshape(21, 3)
+                kv[i] = True
+        if ov.any():
+            view.tensor('pose_' + entity, 'object/rigid/v1', f'annotation/object/{entity}/pose.pt',
+                        poses, ov, ['T', 'row', 'column'], loc, convention='T_camera_from_object', unit='m')
+        if kv.any():
+            view.tensor('points_' + entity, 'object_keypoints/camera/v1',
+                        f'annotation/object/{entity}/keypoints/xyz.pt', pts, kv,
+                        ['T', 'keypoint', 'coordinate'], loc, keypoint_schema='H2O_official_object21',
+                        keypoint_names=['center'] + [f'corner_{i}' for i in range(8)] + [f'edge_midpoint_{i}' for i in range(12)],
+                        source_order='official obj_point_order.jpg: center, 8 corners, 12 edge midpoints', unit='m')
+        if oid in assets:
+            directory, obj, aliases = assets[oid]
+            asset_loc = locator(asset_archive, f'object/{entity}/{obj.name}')
+            if aliases:
+                asset_loc['texture_aliases'] = aliases
+            mesh_asset(view, entity, directory, obj, asset_loc)
+        else:
+            view.absent('mesh_' + entity, 'Source object.zip unavailable', 'not_downloaded')
+    if any('depth' in rows[f] for f in ids):
+        width, height = view.meta['camera']['image_size']
+        def depths():
+            for f in ids:
+                item = rows[f].get('depth')
+                if item:
+                    with Image.open(item[0]) as im:
+                        yield np.asarray(im).astype(np.float32) / 1000.0
+                else:
+                    yield np.full((height, width), np.nan, np.float32)
+        view.raster('depth', depths(), None, loc | {'source_field': 'component_sources.depth',
+            'definition': 'https://github.com/taeinkwon/h2oplayer/blob/main/utils.py',
+            'conversion': 'registered millimetre camera-Z to metres; no viewer range clipping'},
+            codec_threads=codec_threads)
+    else:
+        view.absent('depth', 'No source depth frames')
+    view.absent('segment', 'H2O provides no native segmentation; rendering deferred to v2.0')
+    view.absent('relations', 'No source contact/grasp annotation')
+    return view.finish()
+
+
+def convert_archive(archive, args, out, splits, models, assets, asset_archive):
+    selected_views = set(args.views.split(',')) if args.views else None
+    subjects = {'subject' + s.strip() for s in args.subjects.split(',')} if args.subjects else None
+    current, rows, intrinsics, done = None, {}, {}, 0
+    with tempfile.TemporaryDirectory(prefix='h2o-source-') as wd:
+        work = Path(wd)
+        def flush():
+            nonlocal done
+            if not rows:
+                return
+            seq = '/'.join(current)
+            reports, pending = {}, {}
+            for cam, frames in sorted(rows.items()):
+                existing = out / 'data' / seq.replace('/', '_') / cam
+                if (existing / 'metainfo.json').is_file():
+                    old_frames = [json.loads(line) for line in (existing / 'camera/frames.jsonl').read_text().splitlines()]
+                    if [f['source_frame_id'] for f in old_frames] != list(range(min(frames), max(frames) + 1)):
+                        raise ValueError('existing view has different sampling; choose a new release-id')
+                    for frame in old_frames:
+                        incoming = frames.get(frame['source_frame_id'], {})
+                        previous = frame.get('component_sources', {})
+                        if set(previous) != set(incoming) or any(previous[k].get('sha256') != v[1].get('sha256') for k, v in incoming.items()):
+                            raise ValueError('duplicate H2O view differs from the published source; choose a new release-id')
+                else:
+                    pending[cam] = frames
+            if pending:
+                with sequence_dir(out, seq.replace('/', '_')) as target:
+                    for cam, frames in pending.items():
+                        reports[cam] = convert_view(target / cam, args, seq, cam, frames,
+                            intrinsics.get(cam), splits, models, assets, asset_archive)
+                build_record(out, seq.replace('/', '_'), reports)
+            for staged in list(work.iterdir()):
+                if staged.is_dir():
+                    shutil.rmtree(staged)
+                else:
+                    staged.unlink()
+            done += 1
+            print(f'{"Converted" if pending else "Verified existing"} {seq}', flush=True)
+        # Uncompressed CASA supports seeking over large frames; gzip is single-pass.
+        mode = 'r:' if archive.suffix == '.tar' else 'r|gz'
+        with tarfile.open(archive, mode) as tf:
+            for member in tf:
+                match = MEMBER.fullmatch(member.name)
+                if not member.isfile() or not match:
+                    continue
+                subj, scene, seq, cam, tail = match.groups()
+                if subjects and subj not in subjects:
+                    continue
+                key = (subj, scene, seq)
+                if key != current:
+                    flush()
+                    if args.limit and done >= args.limit:
+                        break
+                    current, rows, intrinsics = key, {}, {}
+                if selected_views and cam not in selected_views:
+                    continue
+                frame = FRAME.fullmatch(tail)
+                if tail != 'cam_intrinsics.txt' and not frame:
+                    continue
+                if frame and args.max_frames and int(frame[2]) >= args.max_frames:
+                    continue
+                data = tf.extractfile(member).read()
+                dest = safe_member(work, member.name)
+                dest.parent.mkdir(parents=True, exist_ok=True)
+                dest.write_bytes(data)
+                item = (dest, locator(archive, member.name, data))
+                if tail == 'cam_intrinsics.txt':
+                    intrinsics[cam] = item
+                else:
+                    kind, f = frame[1], int(frame[2])
+                    row = rows.setdefault(cam, {}).setdefault(f, {})
+                    if kind in row:
+                        raise ValueError('duplicate H2O frame')
+                    row[kind] = item
+            else:
+                flush()
+    return done
+
+
+def main(argv=None):
+    p = parser('H2O', __doc__)
+    p.set_defaults(release_id='shared-v1.0')
+    p.add_argument('--workers', type=int, choices=range(1, 5), default=1,
+                   help='Independent subject archives in parallel, 1..4; benchmark before using >2')
+    p.add_argument('--codec-threads', type=int, choices=(1, 2, 4, 8, 10), default=4,
+                   help='FFV1 slice threads per encode; default 4, use 10 for the current full run')
+    p.add_argument('--subjects')
+    p.add_argument('--include-casa', action='store_true')
+    p.add_argument('--archive', type=Path, help='Convert one explicit source archive')
+    args = p.parse_args(argv)
+    if args.include_casa and args.workers != 1:
+        p.error('CASA can share sequence IDs with subject archives; convert it sequentially in a separate release')
+    archives = [args.archive.expanduser()] if args.archive else sorted(
+        {a for root in roots(args.root) for a in root.glob('subject*_v1_1.tar.gz')})
+    if args.subjects and not args.archive:
+        requested = {'subject' + s.strip() for s in args.subjects.split(',')}
+        archives = [arc for arc in archives if arc.name.split('_', 1)[0] in requested]
+    if args.include_casa:
+        casa = find_file(args.root, 'h2o_CASA.tar')
+        if casa:
+            archives.append(casa)
+    if not archives:
+        p.error('no H2O archives found')
+    out = release(args, 'H2O')
+    source_mapping = next(p for p in (Path(__file__).resolve().parents[1] / 'docs/h2o-RAW_TO_SHARED.md',
+                                     Path(__file__).resolve().parents[1] / 'doc/RAW_TO_SHARED.md',
+                                     out / 'doc/RAW_TO_SHARED.md') if p.is_file())
+    mapping = out / 'doc/RAW_TO_SHARED.md'
+    if mapping.exists() and mapping.read_bytes() != source_mapping.read_bytes():
+        raise ValueError('RAW_TO_SHARED.md changed; choose a new release-id')
+    if not mapping.exists():
+        shutil.copyfile(source_mapping, mapping)
+    torch.set_num_threads(1)
+    models, splits = mano_models(args.mano_dir), load_splits(args.root)
+    with tempfile.TemporaryDirectory(prefix='h2o-assets-') as wd:
+        assets, asset_archive = load_assets(args.root, Path(wd))
+        def convert_one(archive):
+            return convert_archive(archive, args, out, splits, models, assets, asset_archive)
+        if args.workers == 1:
+            for archive in archives:
+                convert_one(archive)
+        else:
+            with ThreadPoolExecutor(max_workers=args.workers) as pool:
+                for archive, count in zip(archives, pool.map(convert_one, archives), strict=True):
+                    print(f'Archive {archive.name}: {count} sequences', flush=True)
+    return 0
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())
+
+# TODO(v2.0): 以下旧 MANO 网格和渲染分割代码仅保留备查；v1.0 不生成标注。
+# def load_mano_model(path: Path):
+#     """MPI MANO pkl → numpy 字典。shapedirs 在 pkl 内为 chumpy Select（对 (778,3,20) 基
+#     底 flat 选取前 10 个 shape 分量；idxs 已核实等价于 base[:,:,:10]）。"""
+#     with open(path, "rb") as f:
+#         d = _ManoUnpickler(f, encoding="latin1").load()
+#     sd = d["shapedirs"]
+#     base = np.asarray(sd.a.x, np.float64)                       # (778, 3, 20)
+#     shapedirs = base.reshape(-1)[np.asarray(sd.idxs)].reshape(base.shape[0], 3, -1)
+#     jreg = d["J_regressor"]
+#     jreg = np.asarray(jreg.todense()) if hasattr(jreg, "todense") else np.asarray(jreg)
+#     return {"v_template": np.asarray(d["v_template"], np.float64),
+#             "shapedirs": shapedirs,
+#             "posedirs": np.asarray(d["posedirs"], np.float64),
+#             "weights": np.asarray(d["weights"], np.float64),
+#             "J_regressor": np.asarray(jreg, np.float64),
+#             "kintree": np.asarray(d["kintree_table"], np.int64),
+#             "faces": np.asarray(d["f"], np.int64)}
+#
+#
+# def rodrigues(r):
+#     """axis-angle (N,3) → 旋转矩阵 (N,3,3)。公式 I + sinθ·K + (1−cosθ)·K²（已对 scipy 核验）。"""
+#     r = np.asarray(r, np.float64).reshape(-1, 3)
+#     th = np.linalg.norm(r, axis=1)
+#     R = np.zeros((len(r), 3, 3))
+#     nz = th > 1e-12
+#     R[~nz] = np.eye(3)
+#     k = r[nz] / th[nz, None]
+#     K = np.zeros((len(k), 3, 3))
+#     K[:, 0, 1], K[:, 0, 2], K[:, 1, 2] = -k[:, 2], k[:, 1], -k[:, 0]
+#     K[:, 1, 0], K[:, 2, 0], K[:, 2, 1] = k[:, 2], -k[:, 1], k[:, 0]
+#     s, c = np.sin(th[nz])[:, None, None], np.cos(th[nz])[:, None, None]
+#     R[nz] = np.eye(3) + s * K + (1 - c) * (K @ K)
+#     return R
+#
+#
+# def _mano_chain(model, betas, pose):
+#     """返回 (v_shaped, J, Rs, G)：G 为 16 个关节的全��变换（平移列 = 关节位置）。"""
+#     v = model["v_template"] + np.einsum("nkc,c->nk", model["shapedirs"], betas)
+#     J = model["J_regressor"] @ v
+#     Rs = rodrigues(np.asarray(pose, np.float64).reshape(16, 3))
+#     kt = model["kintree"]
+#     G = np.zeros((16, 4, 4))
+#     for j in range(16):
+#         p = kt[0, j] if j else -1
+#         loc = np.eye(4)
+#         loc[:3, :3] = Rs[j]
+#         loc[:3, 3] = J[j] - (J[p] if j else 0.0)
+#         G[j] = loc if not j else G[p] @ loc
+#     return v, J, Rs, G
+#
+#
+# def mano_forward(model, betas, pose, transl):
+#     """MANO LBS 前向（经典蒙皮 M_j = G_j·G0_j⁻¹；对 hand_pose 21 关节实证 0.0 mm）。
+#     pose: 48 axis-angle（[0:3] 全局）；返回相机系米顶点 (778,3)。"""
+#     v, J, Rs, G = _mano_chain(model, betas, pose)
+#     lrot = (Rs[1:] - np.eye(3)).ravel()
+#     v_posed = v + np.einsum("nkc,c->nk", model["posedirs"], lrot)
+#     kt = model["kintree"]
+#     G0 = np.zeros((16, 4, 4))
+#     for j in range(16):
+#         p = kt[0, j] if j else -1
+#         loc = np.eye(4)
+#         loc[:3, 3] = J[j] - (J[p] if j else 0.0)
+#         G0[j] = loc if not j else G0[p] @ loc
+#     M = np.einsum("nj,jab->nab", model["weights"], G @ np.linalg.inv(G0))
+#     out = np.einsum("nab,nb->na", M[:, :3, :3], v_posed) + M[:, :3, 3]
+#     return out + np.asarray(transl, np.float64)
+#
+#
+# def mano_chain_joints(model, betas, pose, transl):
+#     """16 个关节位置（相机系米）= 链平移列；自检用。"""
+#     _, _, _, G = _mano_chain(model, betas, pose)
+#     return G[:, :3, 3] + np.asarray(transl, np.float64)
+#
+#
+# def rasterize(verts_cam, faces, labels, W, H, fx, fy, cx, cy):
+#     z = verts_cam[:, 2]
+#     P = faces
+#     ok = (z[P] > 0.02).all(1)                 # 任一顶点在相机后 → 丢弃该三角形
+#     P, labels = P[ok], labels[ok]
+#     if len(P) == 0:
+#         return np.zeros((H, W), np.uint16), np.zeros(0, np.int64)
+#     x = verts_cam[:, 0] / z * fx + cx
+#     y = verts_cam[:, 1] / z * fy + cy
+#     xa = np.maximum(0, np.minimum(np.floor(x[P[:, 0]]), np.minimum(
+#         np.floor(x[P[:, 1]]), np.floor(x[P[:, 2]]))).astype(np.int64))
+#     xb = np.minimum(W - 1, np.maximum(np.ceil(x[P[:, 0]]), np.maximum(
+#         np.ceil(x[P[:, 1]]), np.ceil(x[P[:, 2]]))).astype(np.int64))
+#     ya = np.maximum(0, np.minimum(np.floor(y[P[:, 0]]), np.minimum(
+#         np.floor(y[P[:, 1]]), np.floor(y[P[:, 2]]))).astype(np.int64))
+#     yb = np.minimum(H - 1, np.maximum(np.ceil(y[P[:, 0]]), np.maximum(
+#         np.ceil(y[P[:, 1]]), np.ceil(y[P[:, 2]]))).astype(np.int64))
+#     xa, ya = np.minimum(xa, xb), np.minimum(ya, yb)   # 完全出画的三角形 → 空 bbox
+#     wdt, hgt = xb - xa + 1, yb - ya + 1
+#     cnt = wdt * hgt
+#     tot = int(cnt.sum())
+#     tri = np.repeat(np.arange(len(P)), cnt)
+#     k = np.arange(tot) - (np.cumsum(cnt) - cnt)[tri]
+#     px = xa[tri] + k % wdt[tri] + 0.5
+#     py = ya[tri] + k // wdt[tri] + 0.5
+#     ax, ay = x[P[tri, 0]], y[P[tri, 0]]
+#     bx, by = x[P[tri, 1]], y[P[tri, 1]]
+#     cx_, cy_ = x[P[tri, 2]], y[P[tri, 2]]
+#     area = ((x[P[:, 1]] - x[P[:, 0]]) * (y[P[:, 2]] - y[P[:, 0]])
+#             - (y[P[:, 1]] - y[P[:, 0]]) * (x[P[:, 2]] - x[P[:, 0]]))
+#     areat = area[tri]
+#     e0 = (bx - ax) * (py - ay) - (by - ay) * (px - ax)
+#     e1 = (cx_ - bx) * (py - by) - (cy_ - by) * (px - bx)
+#     e2 = (ax - cx_) * (py - cy_) - (ay - cy_) * (px - cx_)
+#     s = np.sign(areat)
+#     inside = (np.abs(areat) > 1e-9) & ((e0 * s >= 0) & (e1 * s >= 0) & (e2 * s >= 0))
+#     px, py, tri = px[inside], py[inside], tri[inside]
+#     at = areat[inside]
+#     w0, w1, w2 = e1[inside] / at, e2[inside] / at, e0[inside] / at
+#     zz = w0 * z[P[tri, 0]] + w1 * z[P[tri, 1]] + w2 * z[P[tri, 2]]
+#     pix = py.astype(np.int64) * W + px.astype(np.int64)
+#     order = np.lexsort((zz, pix))             # 同像素按深度升序 → 首条即最近
+#     pix, tri = pix[order], tri[order]
+#     first = np.unique(pix, return_index=True)[1]
+#     mask = np.zeros(H * W, np.uint16)
+#     win = tri[first]
+#     mask[pix[first]] = labels[win] + 1        # 临时标签 +1（0=背景）
+#     return mask.reshape(H, W), np.bincount(labels[win] + 1)
+#
+#
+# # ============================================================================
+# # ffmpeg 写出与核验（与 convert_dexycb.py 同构）
+# # ============================================================================
+#             verts, faces_all, labels = [], [], []
+#             for ti, (tid, kind) in enumerate(tracks):
+#                 if kind == "hand_left" and b["mano"][0] > 0.5:
+#                     verts.append(mano_forward(mano_models["left"], b["mano"][52:62],
+#                                               b["mano"][4:52], b["mano"][1:4]))
+#                     faces_all.append(mano_models["left"]["faces"])
+#                 elif kind == "hand_right" and b["mano"][62] > 0.5:
+#                     verts.append(mano_forward(mano_models["right"], b["mano"][114:124],
+#                                               b["mano"][66:114], b["mano"][63:66]))
+#                     faces_all.append(mano_models["right"]["faces"])
+#                 elif isinstance(kind, tuple) and int(b["objp"][0]) == kind[1]:
+#                     M = np.asarray(b["objrt"][1:17], np.float64).reshape(4, 4)
+#                     V0, F0 = meshes[kind[1]]
+#                     verts.append(V0 @ M[:3, :3].T + M[:3, 3])
+#                     faces_all.append(F0)
+#                 else:
+#                     continue
+#                 labels.append(np.full(len(faces_all[-1]), ti))
+#             if verts:
+#                 V = np.concatenate(verts)
+#                 offs = np.cumsum([0] + [len(v) for v in verts[:-1]])
+#                 F = np.concatenate([fa + o for fa, o in zip(faces_all, offs)])
+#                 mask, counts = rasterize(V, F, np.concatenate(labels),
+#                                          IMG_W, IMG_H, fx, fy, cx, cy)
+#             else:
+#                 mask, counts = np.zeros((IMG_H, IMG_W), np.uint16), np.zeros(0, np.int64)
+#             lut = np.zeros(n_tracks + 1, np.uint16)
+#             for ti in range(n_tracks):
+#                 present = ti + 1 < len(counts) and counts[ti + 1] > 0
+#                 if present:
+#                     presence[tracks[ti][0]][f] = True
+#                     if active_id[ti] is None:
+#                         pid = next(i for i in range(1, 65535) if i not in used)
+#                         used.add(pid)
+#                         active_id[ti], seg_start[ti] = pid, f
+#                         frame_maps[ti].append({"id": pid})
+#                     lut[ti + 1] = active_id[ti]
+#                 elif active_id[ti] is not None:
+#                     frame_maps[ti][-1]["frames"] = [seg_start[ti], f - 1]
+#                     segments[ti].append([seg_start[ti], f - 1])
+#                     used.discard(active_id[ti])
+#                     active_id[ti] = None
+#             mw.write(lut[mask])
+
+
+# TODO(v2.0): 旧物体尺寸派生代码；v1.0 仅保留源物体关键点。
+#                 pts = np.asarray(b["objp"][1:64], np.float64).reshape(21, 3)
+#                 half = np.abs((pts[1:9] - pts[0]) @ M[:3, :3]).max(0)
+#                 obj_recs.append({
+#                     "track_id": f"o{classes.index(c)}", "frame": f,
+#                     "mesh_ref": f"h2o/{OBJ_CLASSES[c]}.obj",
+#                     "pose": {"quat": rotmat_to_quat_wxyz(M[:3, :3])[0].tolist(),
+#                              "trans": M[:3, 3].tolist()},
+#                     "bbox_center3d": pts[0].tolist(),
+#                     "bbox_size": (2 * half).tolist()})

H2O/shared-v1.0/script/provenance/previous-run/script/provenance/previous-run/script/convert_hoi4d_v1.0.py

@@ -0,0 +1,595 @@
+#!/usr/bin/env python3
+"""HOI4D native observations and annotations → CONTRACT.md v1.0."""
+from collections import Counter
+from concurrent.futures import ThreadPoolExecutor
+from contextlib import ExitStack
+from functools import lru_cache
+import gzip
+import io
+import json
+import os
+from pathlib import Path
+import pickle
+import re
+import shutil
+import shlex
+import posixpath
+import subprocess
+import tempfile
+import tarfile
+import threading
+import zipfile
+
+# Each worker owns one Python thread; prevent native libraries spawning CPU pools.
+for key in ('OMP_NUM_THREADS', 'OPENBLAS_NUM_THREADS', 'MKL_NUM_THREADS', 'NUMEXPR_NUM_THREADS'):
+    os.environ[key] = '1'
+
+import numpy as np
+from PIL import Image
+from scipy.spatial.transform import Rotation
+
+import torch
+
+from warehouse_v1_0 import (View, build_record, check_mesh, decode_args, find_file, json_write, locator,
+    mano_models, model_profile, parser, probe, release, run, safe_member, sequence_dir)
+
+CATEGORIES = {
+    1:  dict(cad="rigid/ToyCar", groups=[("main", [1], "Toy Car", 1)],
+             hands={2: "right"}, objmap={"Toycar": "main"}),
+    2:  dict(cad="rigid/Mug", groups=[("a", [1], "Mug", 2), ("b", [3], "Mug", 2)],
+             hands={2: "right"}, objmap={"Watercup": "a"}),
+    3:  dict(cad="articulated/Laptop", groups=[("main", [1, 3], "Laptop", 3)],
+             hands={2: "right"},
+             objmap={"Laptopdisplay": "main", "Laptopkeyboard": "main"}),
+    4:  dict(cad="articulated/StorageFurniture",
+             groups=[("main", [1, 3, 4], "Storage Furniture", 4),
+                     ("bottle", [5], "Bottle", 5)],
+             hands={2: "right"},
+             objmap={"Lockerbody": "main", "Lockersldingdoor": "main",
+                     "Lockerdrawer": "main", "Safebox": "main"}),
+    5:  dict(cad="rigid/Bottle", groups=[("main", [1], "Bottle", 5),
+                                         ("mug", [3], "Mug", 2)],
+             hands={2: "right"}, objmap={"bottleddrinks": "main"}),
+    6:  dict(cad="articulated/Safe", groups=[("main", [1, 3], "Safe", 6),
+                                              ("bottle", [4], "Bottle", 5)],
+             hands={2: "right"}, objmap={"Safebox": "main", "Safedoor": "main"}),
+    7:  dict(cad="rigid/Bowl", groups=[("main", [1], "Bowl", 7),
+                                       ("ball", [3], "Ball", None)],
+             hands={2: "right"}, objmap={"bowl": "main"}),
+    8:  dict(cad="articulated/Bucket",
+             groups=[("a", [1, 3], "Bucket", 8), ("b", [4, 5], "Bucket", 8)],
+             hands={2: "right", 6: "left"},
+             objmap={"bucket": "a", "Buckethandle": "a"}),
+    9:  dict(cad="articulated/Scissors", groups=[("main", [1, 3], "Scissors", 9)],
+             hands={2: "right"},
+             objmap={"scissors": "main", "scissors2": None}),   # scissors2：场景道具，无掩码
+    11: dict(cad="articulated/Pliers", groups=[("main", [1, 3], "Pliers", 11)],
+             hands={2: "right"}, objmap={"Pliersleft": "main", "Pliersright": "main"}),
+    12: dict(cad="rigid/Kettle", groups=[("main", [1], "Kettle", 12),
+                                         ("mug", [3], "Mug", 2)],
+             hands={2: "right"}, objmap={"kettle": "main"}),
+    13: dict(cad="rigid/Knife", groups=[("main", [1], "Knife", 13),
+                                        ("apple", [3], "Apple", None)],
+             hands={2: "right"}, objmap={"knife": "main"}),
+    14: dict(cad="articulated/TrashCan",
+             groups=[("main", [1, 3], "Trash Can", 14), ("bottle", [4], "Bottle", 5)],
+             hands={2: "right"}, objmap={"Dustbinbase": "main", "Dustbincover": "main"}),
+    17: dict(cad="articulated/Lamp", groups=[("main", [1, 3, 4, 5], "Lamp", 17)],
+             hands={2: "right", 6: "left"},
+             objmap={"Lampbase": "main", "Tablelampfirstjoint": "main",
+                     "Tablelampsecondjoint": "main", "Thirdjointoftablelamp": "main"}),
+    18: dict(cad="articulated/Stapler",
+             groups=[("main", [1, 3, 4], "Stapler", 18), ("paper", [5], "Paper", None)],
+             hands={2: "right", 6: "left"},       # 6=左手：官方 label.csv 缺列，实测修正
+             objmap={"Staplerbase": "main", "Staplercover": "main"}),
+    20: dict(cad="rigid/Chair", groups=[("main", [1], "Chair", 20)],
+             hands={2: "right", 3: "left"}, objmap={"chair": "main"}),
+}
+
+
+# Official HOI4D definitions/motion segmentation/label.csv; no spatial side inference.
+REGION_NAMES = {1: {1: 'Toy Car', 2: 'Right Hand'}, 2: {1: 'Mug', 2: 'Right Hand', 3: 'Mug'}, 3: {1: 'Screen', 2: 'Right Hand', 3: 'Keyboard'}, 4: {1: 'Door', 2: 'Right Hand', 3: 'Body of the Cabinet', 4: 'Drawer', 5: 'Bottle'}, 5: {1: 'Bottle', 2: 'Right Hand', 3: 'Mug'}, 6: {1: 'Body of the Safe', 2: 'Right Hand', 3: 'Door', 4: 'Bottle'}, 7: {1: 'Bowl', 2: 'Right Hand', 3: 'Ball'}, 8: {1: 'Body of the Bucket', 2: 'Right Hand', 3: 'Handle', 4: 'Body of the Bucket', 5: 'Handle', 6: 'Left Hand'}, 9: {1: 'Left Part of the Scissors', 2: 'Right Hand', 3: 'Right Part of the Scissors'}, 11: {1: 'Left Part of the Pliers', 2: 'Right Hand', 3: 'Right Part of the Pliers'}, 12: {1: 'Kettle', 2: 'Right Hand', 3: 'Mug'}, 13: {1: 'Knife', 2: 'Right Hand', 3: 'Apple'}, 14: {1: 'Body of the Trash Can', 2: 'Right Hand', 3: 'Lid', 4: 'Bottle'}, 17: {1: 'Base', 2: 'Right Hand', 3: 'Tube', 4: 'Top Link', 5: 'Bottom Link', 6: 'Left Hand'}, 18: {1: 'Base', 2: 'Right Hand', 3: 'Slot', 4: 'Lid', 5: 'Paper'}, 20: {1: 'Chair', 2: 'Right Hand', 3: 'Left Hand'}}
+
+
+class SplitFiles(io.RawIOBase):
+    """Read concatenated source gzip parts without concatenating 119 GB on disk."""
+    def __init__(self, paths):
+        self.paths = iter(paths)
+        self.current = None
+
+    def readable(self):
+        return True
+
+    def readinto(self, buffer):
+        while True:
+            if self.current is None:
+                path = next(self.paths, None)
+                if path is None:
+                    return 0
+                self.current = path.open('rb')
+            n = self.current.readinto(buffer)
+            if n:
+                return n
+            self.current.close()
+            self.current = None
+
+    def close(self):
+        if self.current is not None:
+            self.current.close()
+        super().close()
+
+
+def extract_depths(root, sequences, work):
+    directory = find_file(root, 'DEPTH')
+    if directory is None:
+        return {}, []
+    parts = sorted(directory.glob('HOI4D_depth_video.tar.gz*'), key=lambda p: int(p.name.split('.gz')[1] or 0))
+    if not parts:
+        return {}, []
+    wanted = {f'HOI4D_depth_video/{seq}/align_depth/depth_video.avi': seq for seq in sequences}
+    found = {}
+    with SplitFiles(parts) as joined, io.BufferedReader(joined) as inp, gzip.GzipFile(fileobj=inp) as gz:
+        with tarfile.open(fileobj=gz, mode='r|') as tf:
+            for member in tf:
+                if member.isfile() and member.name in wanted:
+                    seq = wanted[member.name]
+                    dest = work / (seq.replace('/', '_') + '.avi')
+                    with tf.extractfile(member) as src, dest.open('wb') as dst:
+                        shutil.copyfileobj(src, dst)
+                    found[seq] = (dest, {'files': [str(p.resolve()) for p in parts], 'member': member.name})
+                    if len(found) == len(wanted):
+                        break
+    return found, parts
+
+
+def palette_mask(data):
+    with Image.open(io.BytesIO(data)) as im:
+        arr = np.asarray(im)
+    if arr.ndim == 2:
+        return arr.astype(np.uint16)
+    if arr.shape[-1] != 3:
+        raise ValueError('unsupported HOI4D mask representation')
+    # VOC's 256-color palette interleaves eight label bits into the high RGB bits.
+    # Invert it directly instead of sorting every pixel with np.unique.
+    if np.any(arr & np.array([31, 31, 63], dtype=np.uint8)):
+        raise ValueError('source mask has colors outside official VOC palette')
+    mask = np.zeros(arr.shape[:2], dtype=np.uint16)
+    for bit in range(8):
+        mask |= ((arr[..., bit % 3] >> (7 - bit // 3)) & 1).astype(np.uint16) << bit
+    return mask
+
+
+# Four workers each reuse annotation and hand indices instead of rebuilding the ZIP directory.
+@lru_cache(maxsize=8)
+def index_frames(z):
+    index = {}
+    for name in z.namelist():
+        parent, _, filename = name.rpartition('/')
+        stem, dot, suffix = filename.rpartition('.')
+        if not dot or not stem.isdigit() or suffix not in ('png', 'json', 'pickle'):
+            continue
+        entries = index.setdefault((parent + '/', '.' + suffix), {})
+        f = int(stem)
+        if f in entries:
+            raise ValueError(f'duplicate frame number under {parent}')
+        entries[f] = name
+    return index
+
+
+def frame_members(z, prefix, suffix):
+    return index_frames(z).get((prefix, suffix), {})
+
+
+def convert_sequence(seq, args, target, rgbz, annz, hpz, calz, cadz, paths, models, depth):
+    # ponytail: cache only one sequence's encoded annotation bytes; no decoded masks or CAD cache.
+    @lru_cache(maxsize=None)
+    def read_annotation(archive, name):
+        return archive.read(name)
+
+    category = int(re.search(r'/C(\d+)/', seq)[1])
+    definition = CATEGORIES.get(category)
+    if definition is None:
+        raise ValueError(f'category mapping unavailable: {category}')
+    prefix = f'HOI4D_annotations/{seq}'
+    mask_members = frame_members(annz, prefix + '/2Dseg/mask/', '.png')
+    shifted = frame_members(annz, prefix + '/2Dseg/shift_mask/', '.png')
+    if not mask_members:
+        mask_members = shifted
+    pose_members = frame_members(annz, prefix + '/objpose/', '.json')
+    hand_members = {side: frame_members(hpz, f'Hand_pose/handpose_{side}_hand/{seq}/', '.pickle')
+                    for side in ('left', 'right')} if hpz else {'left': {}, 'right': {}}
+    with tempfile.TemporaryDirectory(prefix='hoi4d-source-') as wd:
+        work = Path(wd)
+        member = f'HOI4D_release/{seq}/align_rgb/image.mp4'
+        video = work / 'rgb.mp4'
+        with rgbz.open(member) as src, video.open('wb') as dst:
+            shutil.copyfileobj(src, dst)
+        info = probe(video)
+        if 'nb_frames' not in info:
+            info['nb_frames'] = json.loads(run(['ffprobe', '-v', 'error', '-count_frames', '-select_streams', 'v:0',
+                '-show_entries', 'stream=nb_read_frames', '-of', 'json', str(video)], stdout=subprocess.PIPE).stdout)['streams'][0]['nb_read_frames']
+        n = int(info['nb_frames'])
+        if args.max_frames:
+            n = min(n, args.max_frames)
+        if shifted and mask_members != shifted and any(f not in mask_members and f in shifted for f in range(n)):
+            raise ValueError('shift_mask has frames absent from mask; correspondence unverified')
+        frames = []
+        for f in range(n):
+            sources = {}
+            if f in mask_members:
+                sources['segment'] = locator(paths['annotations'], mask_members[f], read_annotation(annz, mask_members[f]))
+            if f in pose_members:
+                sources['object'] = locator(paths['annotations'], pose_members[f], read_annotation(annz, pose_members[f]))
+            for side in ('left', 'right'):
+                if f in hand_members[side]:
+                    sources['hand_' + side] = locator(paths['hand'], hand_members[side][f], read_annotation(hpz, hand_members[side][f]))
+            frames.append(dict(source_frame_id=f, source_locator=locator(paths['release'], member, frame=f),
+                component_sources=sources, annotation_matches={k: dict(source_frame_id=f, source_timestamp_ns=None,
+                    clock_domain=None, match_method='index_map', delta_ns=None) for k in sources}))
+        view = View(target, 'HOI4D', args.release_id, seq.replace('/', '_'), 'ego', frames)
+        view.valid_rules = {
+            'mano': 'Matched side pickle contains poseCoeff, beta, trans and corresponding model available; zero angles are not missing',
+            'keypoints': 'Matched side pickle contains kps2D; independently provided source points, no visibility gate',
+            'pose': 'Matched source object record contains rotation and center; record identity retained',
+            'dimensions': 'Matched source object record contains dimensions; size/mesh relation remains source_unspecified',
+            'segment': 'Source mask or shift_mask present with verified palette codes and entity/part mapping',
+        }
+        view.full_jpeg = getattr(args, 'jpeg_full_check', False)
+        view.meta['view_type'] = 'ego'
+        view.meta['sampling']['selection'] = {'max_frames': args.max_frames, 'source_sequence': seq}
+        view.rgb([], locator(paths['release'], member), video=video, count=n)
+        intr_name = f'camera_params/{seq.split("/")[0]}/intrin.npy'
+        if calz and intr_name in calz.namelist():
+            view.camera(np.load(io.BytesIO(calz.read(intr_name))), locator(paths['camera'], intr_name))
+        extr_name = prefix + '/3Dseg/output.log'
+        if extr_name in annz.NameToInfo:
+            lines = [line for line in annz.read(extr_name).decode().splitlines() if line.strip()]
+            if len(lines) % 5:
+                raise ValueError('invalid Open3D camera trajectory')
+            extr, valid = np.full((n, 4, 4), np.nan, np.float32), np.zeros(n, bool)
+            for i in range(0, len(lines), 5):
+                idx = int(lines[i].split()[0])
+                if idx < n:
+                    if valid[idx]:
+                        raise ValueError('duplicate trajectory index')
+                    extr[idx] = np.linalg.inv(np.loadtxt(io.StringIO('\n'.join(lines[i + 1:i + 5]))))
+                    valid[idx] = True
+            view.tensor('extrinsic', 'transform/se3/v1', 'camera/extrinsic.pt', extr, valid,
+                ['T', 'row', 'column'], locator(paths['annotations'], extr_name),
+                convention='T_camera_from_world', source_convention='Open3D LOG camera-to-world; inverted')
+        loc = {'frames_ref': 'camera/frames.jsonl', 'source_field': 'component_sources'}
+        motion, mv = np.full((n, 2, 61), np.nan, np.float32), np.zeros((n, 2), bool)
+        points, pv = np.full((n, 2, 21, 2), np.nan, np.float32), np.zeros((n, 2, 21), bool)
+        for slot, side in enumerate(('left', 'right')):
+            for f, name in hand_members[side].items():
+                if f >= n:
+                    continue
+                data = pickle.loads(read_annotation(hpz, name), encoding='latin1')
+                if all(k in data for k in ('poseCoeff', 'beta', 'trans')):
+                    motion[f, slot] = np.r_[np.asarray(data['poseCoeff']).reshape(48),
+                        np.asarray(data['beta']).reshape(10), np.asarray(data['trans']).reshape(3)]
+                    mv[f, slot] = True  # Zero angles are a valid pose, not a missing sentinel.
+                if 'kps2D' in data:
+                    points[f, slot] = np.asarray(data['kps2D']).reshape(21, 2)
+                    pv[f, slot] = True
+        if mv.any():
+            if all(side in models for slot, side in enumerate(('left', 'right')) if mv[:, slot].any()):
+                view.tensor('mano', 'mano/axis_angle61/v1', 'annotation/mano/motion.pt', motion, mv,
+                    ['T', 'hand', 'parameter'], loc, model_profile=model_profile(models))
+            else:
+                view.absent('mano', 'Source MANO exists; provide --mano-dir to fix model identity', 'dependency_missing')
+        if pv.any():
+            view.tensor('keypoints_uv', 'keypoints/uv/v1', 'annotation/keypoints/uv.pt', points, pv,
+                ['T', 'hand', 'keypoint', 'coordinate'], loc, keypoint_schema='mano21/v1', unit='pixel',
+                source_method='Upstream fitted MANO projection already present in kps2D')
+        objects = {}
+        groups = {key: (vals, name, cls) for key, vals, name, cls in definition['groups']}
+        def ensure_object(key, source_loc, category_id=None, frame=None, source_id=None):
+            if key not in objects:
+                view.object(key, source_loc, scope='frame' if frame is not None else 'sequence',
+                    scope_id=f'{seq}:{frame}' if frame is not None else None,
+                    category_id=category_id, source_object_id=source_id)
+                objects[key] = {}
+        mappings = []
+        mask_valid = [f in mask_members for f in range(n)]
+        if any(mask_valid):
+            width, height = view.meta['camera']['image_size']
+            def masks():
+                for f in range(n):
+                    labels = []
+                    if f in mask_members:
+                        mask = palette_mask(read_annotation(annz, mask_members[f]))
+                        for value in np.unique(mask):
+                            value = int(value)
+                            if value in (0, 255):
+                                if value == 255:
+                                    raise ValueError('HOI4D 255 region semantics unverified')
+                                continue
+                            if REGION_NAMES[category].get(value) in ('Left Hand', 'Right Hand'):
+                                entity, part, cls = 'hand_' + REGION_NAMES[category][value].split()[0].lower(), None, 'hand'
+                            else:
+                                group = next((k for k, (vals, _, _) in groups.items() if value in vals), None)
+                                entity = group or f'unknown_region_{value}'
+                                cls = groups[group][2] if group else None
+                                part = f'region_{value}' if group and len(groups[group][0]) > 1 else None
+                                ensure_object(entity, frames[f]['component_sources']['segment'], cls)
+                                if part:
+                                    objects[entity].setdefault('parts', {})[part] = {'part_id': part, 'parent_id': None,
+                                        'has_model': False, 'mesh_ref': None, 'source_region_code': value, 'name': REGION_NAMES[category].get(value)}
+                            labels.append(dict(mask_value=value, category_id=cls, entity_id=entity,
+                                               part_id=part, source_instance_id=value,
+                                               source_attributes={'region_name': REGION_NAMES[category].get(value)}))
+                    else:
+                        mask = np.zeros((height, width), np.uint16)
+                    mappings.append(dict(frame_index=f, labels=labels))
+                    yield mask
+            view.raster('segment', masks(), mask_valid, loc, mappings)
+            view.meta['components']['segment']['coverage'] = {'kind': 'unknown'}
+            if any(label['part_id'] for row in mappings for label in row['labels']):
+                view.meta['components']['segment']['segmentation_type'] = 'part'
+        else:
+            view.absent('segment', 'No native segmentation for this source sequence')
+        records = {}
+        for f, name in pose_members.items():
+            if f >= n:
+                continue
+            data = json.loads(read_annotation(annz, name))
+            # isEffective is not used to erase existing numerical annotations.
+            source_records = data.get('dataList', data.get('objects'))
+            if source_records is None:
+                raise ValueError('unrecognized source object annotation structure')
+            if category == 2 and source_records and (len(source_records) != 1 or
+                    source_records[0].get('id') != 1 or source_records[0].get('label') != 'Watercup'):
+                raise ValueError('C2 native pose differs from the verified one-object target; cannot map to mask region 1')
+            label_counts = Counter(record['label'] for record in source_records)
+            per_frame = set()
+            for record_index, record in enumerate(source_records):
+                label = record['label']
+                group = definition['objmap'].get(label)
+                # Pinned official ObjTracking reads dataList[0] with segmentation label 1.
+                source_primary = category == 2 and record_index == 0 and record.get('id') == 1 and group == 'a'
+                ambiguous = group is not None and sum(g[2] == groups[group][2] for g in groups.values()) > 1 and not source_primary
+                ambiguous = ambiguous or (label_counts[label] > 1 and not source_primary) or group is None
+                entity = f'pose_frame_{f}_{record_index}' if ambiguous else group
+                ensure_object(entity, locator(paths['annotations'], name, record=record_index),
+                    groups[group][2] if group else None, frame=f if ambiguous else None, source_id=record.get('id'))
+                if ambiguous:
+                    view.build['unresolved'].append({'component': entity, 'status': 'source_semantics_unverified',
+                        'reason': 'Source pose retained with frame identity; no guessed association to same-class mask instances'})
+                label_key = (entity, label)
+                if label_key in per_frame:
+                    raise ValueError(f'ambiguous repeated object label {label}; source identity must be resolved')
+                per_frame.add(label_key)
+                records.setdefault(label_key, {})[f] = (record, locator(paths['annotations'], name, record=record.get('id')))
+        if category == 2 and any(entity == 'a' for entity, _ in records):
+            view.build['validation'].append({'component': 'C2_source_pose_mask_identity',
+                'rule': 'Official ObjTracking process_data.py reads dataList[0] with semantic mask label 1',
+                'source_revision': 'afa7bf37b02b7e33cf245acd08b98c207ccc175c',
+                'source_pose_id': 1, 'mask_value': 1, 'matched_frames': sum(len(by_frame) for (entity, _), by_frame in records.items() if entity == 'a')})
+        for (entity, label), by_frame in records.items():
+            part_labels = [k[1] for k in records if k[0] == entity]
+            articulated = definition['cad'].startswith('articulated/') and entity == definition['groups'][0][0]
+            part = 'pose_' + re.sub(r'[^a-zA-Z0-9_]', '_', label) if articulated or len(part_labels) > 1 else None
+            base = f'annotation/object/{entity}'
+            if part:
+                base += '/parts/' + part
+                objects[entity].setdefault('parts', {})[part] = dict(part_id=part, parent_id=None,
+                    has_model=False, mesh_ref=None, source_label=label)
+            poses, valid = np.full((n, 4, 4), np.nan, np.float32), np.zeros(n, bool)
+            dimensions = np.full((n, 3), np.nan, np.float32)
+            dv = np.zeros(n, bool)
+            for f, (record, _) in by_frame.items():
+                if all(k in record for k in ('rotation', 'center')):
+                    poses[f] = np.eye(4)
+                    poses[f, :3, :3] = Rotation.from_euler('XYZ', [record['rotation'][k] for k in ('x', 'y', 'z')]).as_matrix()
+                    poses[f, :3, 3] = [record['center'][k] for k in ('x', 'y', 'z')]
+                    valid[f] = True
+                if 'dimensions' in record:
+                    dimensions[f] = [record['dimensions'][k] for k in ('length', 'width', 'height')]
+                    dv[f] = True
+            key = entity + ('_' + part if part else '')
+            view.tensor('pose_' + key, 'transform/se3/v1' if part else 'object/rigid/v1', base + '/pose.pt', poses, valid,
+                ['T', 'row', 'column'], loc, convention='T_camera_from_part' if part else 'T_camera_from_object',
+                unit='m', source_label=label, source_euler_order='XYZ',
+                definition='https://github.com/hoi4d/ObjTracking/blob/main/process_data.py')
+            if dv.any():
+                view.tensor('dimensions_' + key, 'object/dimensions/v1', base + '/dimensions.pt', dimensions, dv,
+                    ['T', 'dimension'], loc, dimensions_semantics='source_unspecified', unit='m', scale_applied_to_mesh=False)
+        if cadz:
+            primary = definition['groups'][0][0]
+            if category != 2 and sum(g[2] == category for g in groups.values()) > 1:
+                primary = 'source_cad_instance'
+            instance = int(re.search(r'/N(\d+)/', seq)[1])
+            mesh_prefix = f'HOI4D_CAD_Model_for_release/{definition["cad"]}/{instance:03d}'
+            candidates = [name for name in cadz.namelist() if not name.endswith('/') and
+                          (name == mesh_prefix + '.obj' or name.startswith(mesh_prefix + '/'))]
+            if candidates:
+                # Retain OBJ material/texture dependencies beside their original names.
+                queue = list(candidates)
+                for name in queue:
+                    if Path(name).suffix.lower() not in ('.obj', '.mtl'):
+                        continue
+                    for line in cadz.read(name).decode().splitlines():
+                        fields = shlex.split(line, comments=True)
+                        if not fields:
+                            continue
+                        refs = fields[1:] if fields[0] == 'mtllib' else ([fields[-1]] if
+                            fields[0].startswith('map_') or fields[0] in ('bump', 'norm', 'disp', 'decal') else [])
+                        for ref in refs:
+                            dependency = posixpath.normpath(posixpath.join(posixpath.dirname(name), ref))
+                            if dependency not in cadz.NameToInfo:
+                                raise ValueError(f'missing CAD dependency: {dependency}')
+                            if dependency not in candidates:
+                                candidates.append(dependency)
+                                queue.append(dependency)
+                ensure_object(primary, locator(paths['cad'], mesh_prefix), category)
+                asset_dir = view.path / f'annotation/object/{primary}/assets'
+                asset_refs = []
+                for name in candidates:
+                    relative = posixpath.relpath(name, mesh_prefix if definition['cad'].startswith('articulated/') else posixpath.dirname(mesh_prefix))
+                    dest = safe_member(asset_dir, relative)
+                    dest.parent.mkdir(parents=True, exist_ok=True)
+                    dest.write_bytes(cadz.read(name))
+                    ref = dest.relative_to(view.path).as_posix()
+                    if dest.suffix.lower() == '.obj':
+                        asset_key = re.sub(r'[^a-zA-Z0-9_]', '_', relative)
+                        view.component('mesh_' + primary + '_' + asset_key, 'mesh/obj/v1', ref, ['vertex'],
+                            locator(paths['cad'], name), unit='m', valid=True,
+                            scale_applied_to_mesh=False, dimensions_semantics='source_unspecified')
+                        asset_refs.append(ref)
+                        if definition['cad'].startswith('articulated/'):
+                            part_id = 'asset_' + asset_key
+                            objects[primary].setdefault('parts', {})[part_id] = dict(part_id=part_id,
+                                parent_id=None, has_model=True, mesh_ref=ref, source_asset=relative,
+                                note='Source asset; correspondence to motion part labels remains unverified')
+                for asset_ref in asset_refs:
+                    mesh_check = check_mesh(view.path / asset_ref, asset_dir)
+                    view.build['validation'].append({'component': asset_ref, **mesh_check,
+                        'unit_basis': 'official ObjTracking loads source canonical CAD without scaling in metre camera coordinates',
+                        'geometric_acceptance': 'physical dimensions-to-mesh relation remains source_unspecified'})
+                metadata_path = view.path / f'annotation/object/{primary}/metainfo.json'
+                metadata = json.loads(metadata_path.read_text())
+                metadata.update(has_model=bool(asset_refs), asset_refs=asset_refs)
+                if definition['cad'].startswith('rigid/') and len(asset_refs) == 1:
+                    metadata['mesh_ref'] = asset_refs[0]
+                json_write(metadata_path, metadata)
+                if definition['cad'].startswith('articulated/'):
+                    view.build['unresolved'].append({'component': 'object_part_asset_association',
+                        'status': 'source_semantics_unverified',
+                        'reason': 'All source assets and motion labels retained, but no guessed correspondence or joint_state'})
+        for entity, info in objects.items():
+            if 'parts' in info:
+                ref = f'annotation/object/{entity}/articulation.json'
+                json_write(view.path / ref, {'parts': list(info['parts'].values()), 'joints': [],
+                    'hierarchy_semantics': 'semantic_containment',
+                    'note': 'Unassociated source region, motion and asset nodes; original CAD trees retained in assets'})
+                view.component('articulation_' + entity, 'object/articulated/v1', ref, [], loc)
+                meta_path = view.path / f'annotation/object/{entity}/metainfo.json'
+                metadata = json.loads(meta_path.read_text())
+                metadata['articulation_ref'] = ref
+                json_write(meta_path, metadata)
+        if depth is not None:
+            dpath, dloc = depth
+            dmeta = probe(dpath)
+            if [dmeta['width'], dmeta['height']] != view.meta['camera']['image_size']:
+                raise ValueError('HOI4D depth grid differs from RGB')
+            if dmeta['pix_fmt'] != 'gray16le':
+                raise ValueError('unexpected HOI4D depth pixel format')
+            width, height = view.meta['camera']['image_size']
+            raw = work / 'depth.raw'
+            limits = ['-frames:v', str(n)] if args.max_frames else []
+            run(decode_args(dpath, 'gray16le') + limits + ['-y', str(raw)])
+            if raw.stat().st_size != n * width * height * 2:
+                raise ValueError('depth/RGB frame count or grid mismatch')
+            def depths():
+                with raw.open('rb') as inp:
+                    for _ in range(n):
+                        yield np.frombuffer(inp.read(width * height * 2), '<u2').reshape(height, width).astype(np.float32) / 1000
+            view.raster('depth', depths(), None, dloc | {'conversion': 'registered millimetre camera-Z to metres'})
+        else:
+            if find_file(args.root, 'DEPTH'):
+                raise ValueError('DEPTH source exists, but no matching depth video was found')
+            view.absent('depth', 'Depth source not downloaded', 'not_downloaded')
+        return view.finish()
+
+
+def main(argv=None):
+    p = parser('HOI4D', __doc__)
+    p.set_defaults(release_id='shared-v1.0')
+    # Four workers reached 21 observed OS threads including timing; stay below the 24-thread budget.
+    p.add_argument('--workers', type=int, choices=range(1, 5), default=1,
+                   help='Independent views in parallel, 1..4; codec/BLAS threads fixed at 1')
+    p.add_argument('--tasks', help='Comma-separated substrings of source sequence paths')
+    p.add_argument('--sessions', help='Comma-separated sessions')
+    p.add_argument('--categories', help='Comma-separated category numbers')
+    args = p.parse_args(argv)
+    if args.views and args.views != 'ego':
+        p.error('HOI4D source camera ID is ego')
+    paths = {key: find_file(args.root, name) for key, name in {
+        'release': 'HOI4D_release.zip', 'annotations': 'HOI4D_annotations.zip', 'hand': 'HOI4D_Hand_pose.zip',
+        'camera': 'camera_params.zip', 'cad': 'HOI4D_CAD_Model_for_release.zip'}.items()}
+    if not paths['release'] or not paths['annotations']:
+        p.error('HOI4D release and annotation ZIPs are required')
+    out = release(args, 'HOI4D')
+    mapping = Path(__file__).resolve().parent.parent / 'docs/hoi4d-RAW_TO_SHARED.md'
+    published_mapping = out / 'doc/RAW_TO_SHARED.md'
+    if published_mapping.exists() and published_mapping.read_bytes() != mapping.read_bytes():
+        raise ValueError('RAW_TO_SHARED.md changed; choose a new release-id')
+    if not published_mapping.exists():
+        shutil.copyfile(mapping, published_mapping)
+    torch.set_num_threads(1)
+    models = mano_models(args.mano_dir)
+    with ExitStack() as stack, tempfile.TemporaryDirectory(prefix='hoi4d-depth-') as wd:
+        archives = {key: stack.enter_context(zipfile.ZipFile(path)) if path else None for key, path in paths.items()}
+        sequences = sorted(name.removeprefix('HOI4D_release/').removesuffix('/align_rgb/image.mp4')
+                           for name in archives['release'].namelist() if name.endswith('/align_rgb/image.mp4'))
+        if args.tasks:
+            sequences = [s for s in sequences if any(token in s for token in args.tasks.split(','))]
+        if args.sessions:
+            sequences = [s for s in sequences if s.split('/')[0] in args.sessions.split(',')]
+        if args.categories:
+            sequences = [s for s in sequences if re.search(r'/C(\d+)/', s)[1] in args.categories.split(',')]
+        if args.limit:
+            sequences = sequences[:args.limit]
+        if not sequences:
+            p.error('no matching source sequences')
+        depths, _ = extract_depths(args.root, sequences, Path(wd))
+        worker = threading.local()
+        worker_stacks = []
+        def init_worker():
+            owned = ExitStack()
+            try:
+                # ZipFile.read mutates a shared seek position: never share a handle across workers.
+                worker.archives = {key: owned.enter_context(zipfile.ZipFile(path)) if path else None
+                                   for key, path in paths.items()}
+                index_frames(worker.archives['annotations'])
+                if worker.archives['hand']:
+                    index_frames(worker.archives['hand'])
+            except BaseException:
+                owned.close()
+                raise
+            worker_stacks.append(owned)
+        def convert_one(seq):
+            source = archives if args.workers == 1 else worker.archives
+            name = seq.replace('/', '_')
+            with sequence_dir(out, name) as target:
+                report = convert_sequence(seq, args, target / 'ego', source['release'], source['annotations'],
+                    source['hand'], source['camera'], source['cad'], paths, models, depths.get(seq))
+            build_record(out, name, {'ego': report})
+            return seq
+        if args.workers == 1:
+            for seq in map(convert_one, sequences):
+                print(f'Converted {seq}', flush=True)
+        else:
+            try:
+                with ThreadPoolExecutor(max_workers=args.workers, initializer=init_worker) as pool:
+                    for seq in pool.map(convert_one, sequences):
+                        print(f'Converted {seq}', flush=True)
+            finally:
+                for owned in worker_stacks:
+                    owned.close()
+    return 0
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())
+
+# TODO(v2.0): 旧 bbox IoU 实例重分配属于推断，注释保留，不在 v1.0 执行。
+#                         # 同类多实例（C8 双桶 / C2 双杯）：2dBox 可用时按 IoU 复核归属
+#                         cls = next(x[3] for x in cat["groups"] if x[0] == gk)
+#                         sibs = [g for g in cat["groups"]
+#                                 if g[0] != gk and cls is not None and g[3] == cls]
+#                         box = (d.get("2dBox") or {}).get("camera1")
+#                         box = (box[0][0], box[0][1], box[1][0], box[1][1]) \
+#                             if box and len(box) == 2 else None
+#                         if box is not None and sibs:
+#                             gvals = next(x[1] for x in cat["groups"] if x[0] == gk)
+#                             best, best_iou = gk, box_iou(box, bbox_of(mask, gvals))
+#                             for s in sibs:
+#                                 if ("obj", s[0]) not in tid_of_group:
+#                                     continue
+#                                 i = box_iou(box, bbox_of(mask, s[1]))
+#                                 if i > best_iou + 0.1 and i > 0.15:
+#                                     best, best_iou = s[0], i
+#                             if best != gk:
+#                                 gk, iou_fixups = best, iou_fixups + 1

H2O/shared-v1.0/script/provenance/previous-run/script/provenance/previous-run/script/convert_visor_v1.0.py

@@ -0,0 +1,271 @@
+#!/usr/bin/env python3
+"""VISOR source polygons and contact → CONTRACT.md v1.0, without rasterization or side inference."""
+from collections import Counter
+import hashlib
+import json
+import re
+import shutil
+import tempfile
+from pathlib import Path
+import zipfile
+
+import numpy as np
+import torch
+
+from warehouse_v1_0 import (View, build_record, find_file, jsonl_write, locator, parser,
+                            release, roots, sequence_dir, source)
+
+FRAME = re.compile(r'frame_(\d+)')
+HAND = {300: 'left', 301: 'right'}
+GLOVE = {303: 'left', 304: 'right'}
+STATES = {'hand-not-in-contact': 'negative', 'glove-not-in-contact': 'negative',
+          'none-of-the-above': 'target_unannotated', 'inconclusive': 'unknown'}
+
+
+def hand_side(ann):
+    if ann['class_id'] in HAND:
+        return HAND[ann['class_id']]
+    if ann['class_id'] not in (60, 303, 304):
+        return None
+    raw = ann.get('on_which_hand')
+    raw = raw if isinstance(raw, list) else [raw]
+    sides = set()
+    for item in raw:
+        if item == 'left hand':
+            sides.add('left')
+        if item in ('right hand', 'rigth hand'):
+            sides.add('right')
+    if len(sides) == 1:
+        return next(iter(sides))
+    if not sides and ann['class_id'] in GLOVE:
+        return GLOVE[ann['class_id']]
+    return None
+
+
+def convert_video(video, split, annotation, ann_loc, rgb_zip, rgb_loc, mapping, args, target):
+    indexed = sorted(enumerate(annotation['video_annotations']), key=lambda row: int(FRAME.search(row[1]['image']['name'])[1]))
+    if args.max_frames:
+        indexed = indexed[:args.max_frames]
+    if len({row['image']['name'] for _, row in indexed}) != len(indexed):
+        raise ValueError('duplicate VISOR frame')
+    keys = lambda row, ann: (row['image']['subsequence'], ann['class_id'], ann['name'])
+    collisions = set()
+    for _, row in indexed:
+        collisions.update(key for key, number in Counter(keys(row, a) for a in row['annotations']).items() if number > 1)
+    frames, paths = [], []
+    with tempfile.TemporaryDirectory(prefix='visor-rgb-') as wd, zipfile.ZipFile(rgb_zip) as z:
+        work = Path(wd)
+        for original_index, row in indexed:
+            name = row['image']['name']
+            data = z.read(name)
+            path = work / Path(name).name
+            path.write_bytes(data)
+            paths.append(path)
+            loc = rgb_loc | {'nested_member': name, 'sha256': hashlib.sha256(data).hexdigest()}
+            frames.append(dict(source_frame_id=int(FRAME.search(name)[1]), source_locator=loc,
+                source_image=row['image'], original_frame_name=mapping.get(name),
+                annotation_matches={'segment': dict(source_frame_id=name, source_timestamp_ns=None,
+                    clock_domain=None, match_method='index_map', delta_ns=None,
+                    source_locator=ann_loc | {'record': f'video_annotations[{original_index}]'})}))
+        view = View(target, 'EPIC-KITCHENS-VISOR', args.release_id, video, 'ego', frames, sampling='sparse', split=split)
+        view.valid_rules = {
+            'segment': 'Selected sparse annotation frame present; all source polygons finite with at least three vertices and entity mapping valid',
+            'relations': 'At least one source in_contact_object value present; positive targets resolve in the same frame; unknown is an explicit usable status',
+        }
+        view.full_jpeg = getattr(args, 'jpeg_full_check', False)
+        view.meta.update(view_type='ego', split_scheme='VISOR_GroundTruth_SparseAnnotations', split_source_locator=ann_loc)
+        view.meta['sampling']['selection'] = {'representation': 'official sparse annotations', 'max_frames': args.max_frames}
+        view.rgb(paths, rgb_loc)
+    entities = {}
+    polygons, mappings, relations, relation_valid = [], [], [], []
+    for f, (original_index, row) in enumerate(indexed):
+        frame_loc = ann_loc | {'record': f'video_annotations[{original_index}]'}
+        id_map, labels, polys = {}, [], []
+        for code, ann in enumerate(row['annotations'], 1):
+            if code >= 65535:
+                raise ValueError('too many source instances')
+            sid = ann['id']
+            if sid in id_map:
+                raise ValueError('duplicate source instance ID within a frame')
+            side = hand_side(ann)
+            if side is not None:
+                entity = 'hand_' + side
+            else:
+                key = keys(row, ann)
+                identity = key + (sid, row['image']['name']) if key in collisions else key
+                entity = 'object_' + hashlib.sha256(json.dumps(identity).encode()).hexdigest()[:20]
+                if entity not in entities:
+                    scope = 'frame' if key in collisions else 'subsequence'
+                    scope_id = f'{video}:{row["image"]["name"]}' if scope == 'frame' else f'{video}:{key[0]}'
+                    view.object(entity, frame_loc | {'source_instance_id': sid}, scope, scope_id,
+                        category_id=ann['class_id'], name=ann['name'],
+                        association_basis='source_frame_instance' if scope == 'frame' else 'official_subsequence_entity_name')
+                    entities[entity] = True
+            id_map[sid] = entity
+            for polygon in ann.get('segments', []):
+                p = np.asarray(polygon)
+                if p.ndim != 2 or p.shape[1] != 2 or len(p) < 3 or not np.isfinite(p).all():
+                    raise ValueError(f'invalid source polygon {sid}')
+            polys.append(dict(mask_value=code, polygons=ann.get('segments', [])))
+            labels.append(dict(mask_value=code, category_id=ann['class_id'], entity_id=entity,
+                               part_id=None, source_instance_id=sid,
+                               source_attributes={k: v for k, v in ann.items() if k != 'segments'}))
+        rr = []
+        provided = False
+        for ann in row['annotations']:
+            raw = ann.get('in_contact_object')
+            values = raw if isinstance(raw, list) else [raw]
+            for value in values:
+                if value is None:
+                    continue
+                provided = True
+                status = STATES.get(value)
+                obj = None
+                if status is None:
+                    if value not in id_map:
+                        raise ValueError(f'unresolved source contact target {value} in {video} frame {f}')
+                    status, obj = 'positive', id_map[value]
+                rr.append(dict(relation_type='contact', subject_id=id_map[ann['id']], object_id=obj,
+                    status=status, source_value=value, source_locator=frame_loc | {'source_instance_id': ann['id']},
+                    provenance=source('Resolve source contact ID against instances in this frame')))
+        polygons.append(dict(frame_index=f, instances=polys))
+        mappings.append(dict(frame_index=f, labels=labels))
+        relations.append(dict(frame_index=f, relations=rr))
+        relation_valid.append(provided)
+    jsonl_write(view.path / 'annotation/segment/polygons.jsonl', polygons)
+    jsonl_write(view.path / 'annotation/segment/instances.jsonl', mappings)
+    torch.save(torch.ones(len(frames), dtype=torch.bool, device='cpu'), view.path / 'annotation/segment/mask_valid.pt')
+    view.component('segment', 'segment/instance_polygons/v1', 'annotation/segment/polygons.jsonl', ['T'], ann_loc,
+        valid_ref='annotation/segment/mask_valid.pt', valid_axes=['T'], segmentation_type='instance',
+        mapping_ref='annotation/segment/instances.jsonl', coverage={'kind': 'unknown', 'exhaustive_ref': 'annotation/segment/instances.jsonl'})
+    view.relations(relations, relation_valid, ann_loc)
+    for component in ('depth', 'mano', 'keypoints_uv', 'keypoints_xyz', 'intrinsic', 'extrinsic'):
+        view.absent(component, 'Not provided by VISOR sparse source annotations')
+    return view.finish()
+
+
+def main(argv=None):
+    p = parser('EPIC-KITCHENS-VISOR', __doc__)
+    p.add_argument('--videos', help='Comma-separated video names')
+    p.add_argument('--splits', default='train,val')
+    args = p.parse_args(argv)
+    wanted = set(args.videos.split(',')) if args.videos else None
+    splits = set(args.splits.split(','))
+    if not splits <= {'train', 'val'}:
+        p.error('only annotated train/val splits are available')
+    out = release(args, 'EPIC-KITCHENS-VISOR')
+    archive = find_file(args.root, 'EPIC-KITCHENS VISOR.zip')
+    with tempfile.TemporaryDirectory(prefix='visor-source-') as wd:
+        work = Path(wd)
+        if archive:
+            with zipfile.ZipFile(archive) as z:
+                suffix = 'GroundTruth-SparseAnnotations/annotations/'
+                members = [n for n in z.namelist() if suffix in n and n.endswith('.json')]
+                prefix = members[0].split(suffix)[0]
+                fm = json.loads(z.read(prefix + 'frame_mapping.json'))
+                done = 0
+                for member in sorted(members):
+                    split, filename = member.split(suffix)[1].split('/')
+                    video = Path(filename).stem
+                    if split not in splits or wanted and video not in wanted:
+                        continue
+                    data = z.read(member)
+                    rgb_member = prefix + f'GroundTruth-SparseAnnotations/rgb_frames/{split}/{video.split("_")[0]}/{video}.zip'
+                    rgb_zip = work / (video + '.zip')
+                    with z.open(rgb_member) as inp, rgb_zip.open('wb') as dest:
+                        shutil.copyfileobj(inp, dest)
+                    with sequence_dir(out, video) as dest:
+                        report = convert_video(video, split, json.loads(data), locator(archive, member, data),
+                            rgb_zip, locator(archive, rgb_member), fm.get(video, {}), args, dest / 'ego')
+                    build_record(out, video, {'ego': report})
+                    rgb_zip.unlink()
+                    done += 1
+                    print(f'Converted {video}', flush=True)
+                    if args.limit and done >= args.limit:
+                        break
+        else:
+            root = next((r for r in roots(args.root) if (r / 'GroundTruth-SparseAnnotations').is_dir()), None)
+            if root is None:
+                p.error('VISOR outer ZIP or extracted source directory not found')
+            fm = json.loads((root / 'frame_mapping.json').read_text()) if (root / 'frame_mapping.json').exists() else {}
+            done = 0
+            for split in sorted(splits):
+                for ann in sorted((root / 'GroundTruth-SparseAnnotations/annotations' / split).glob('*.json')):
+                    video = ann.stem
+                    if wanted and video not in wanted:
+                        continue
+                    rgb = root / f'GroundTruth-SparseAnnotations/rgb_frames/{split}/{video.split("_")[0]}/{video}.zip'
+                    with sequence_dir(out, video) as dest:
+                        report = convert_video(video, split, json.loads(ann.read_text()), locator(ann), rgb,
+                            locator(rgb), fm.get(video, {}), args, dest / 'ego')
+                    build_record(out, video, {'ego': report})
+                    done += 1
+                    if args.limit and done >= args.limit:
+                        return 0
+    if not done:
+        p.error('no matching annotated videos found')
+    return 0
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())
+
+# TODO(v2.0): 旧距离猜侧和单层掩码生成仅注释保留；v1.0 不执行。
+#                     if side is None and len(frame_hands) == 1:
+#                         # 空间门槛：脱下放在远处的手套不并入（实测双峰 0..25px / ≥44px）
+#                         hb = frame_hands[0][1]
+#                         gb = _bbox_of(polys)
+#                         if hb and gb and _bbox_dist(gb, hb) <= GLOVE60_MERGE_BBOX_DIST:
+#                             side = frame_hands[0][0]
+#                             stats["glove60_merged_into_sole_hand"] += 1
+#                     if side is None:
+#                         stats["dropped_glove60_unresolved"] += 1
+#                         id_disp[mid] = ("dropped", "glove60")
+#                         continue
+# def rasterize_and_encode(vdir, frames, frame_draw, zip_path, n, stats):
+#     """流式读 zip 内 jpg + 光栅化多边形 → rgb.mkv / mask.mkv；返回 (w, h)。"""
+#     from PIL import Image, ImageDraw
+#
+#     with zipfile.ZipFile(zip_path) as zf:
+#         member_of = {}
+#         for nm in zf.namelist():
+#             member_of[os.path.basename(nm)] = nm
+#         names = [im["name"] for im in frames]
+#         missing = [nm for nm in names if nm not in member_of]
+#         if missing:
+#             raise RuntimeError(f"{len(missing)} annotated frames missing in zip, "
+#                                f"e.g. {missing[:3]}")
+#         stats["zip_extra_members"] = len(member_of) - len(names)
+#
+#         # 首帧定分辨率（官方 1920×1080，实测每帧校验）
+#         img0 = Image.open(io.BytesIO(zf.read(member_of[names[0]]))).convert("RGB")
+#         w, h = img0.size
+#
+#         bin_img = Image.new("L", (w, h), 0)
+#         bin_draw = ImageDraw.Draw(bin_img)
+#         canvas = np.zeros((h, w), np.uint16)
+#         with VideoPipe(vdir / "mask.mkv", w, h, FPS, "mask") as mw, \
+#                 VideoPipe(vdir / "rgb.mkv", w, h, FPS, "rgb") as rw:
+#             for f in range(n):
+#                 canvas[:] = 0
+#                 for pid, polys in frame_draw[f]:
+#                     bin_draw.rectangle((0, 0, w, h), fill=0)
+#                     for poly in polys:
+#                         bin_draw.polygon([(float(x), float(y)) for x, y in poly], fill=255)
+#                     arr = np.asarray(bin_img)
+#                     canvas[arr > 0] = pid          # 后画覆盖先画：物体→手（modal）
+#                 mw.write(canvas)
+#                 img = img0 if f == 0 else Image.open(
+#                     io.BytesIO(zf.read(member_of[names[f]]))).convert("RGB")
+#                 if img.size != (w, h):
+#                     raise RuntimeError(f"frame {names[f]}: size {img.size} != {(w, h)}")
+#                 rw.write(np.asarray(img))
+#             n_mask, n_rgb = mw.n, rw.n
+#     if (n_rgb, n_mask) != (n, n):
+#         raise RuntimeError(f"wrote rgb={n_rgb} mask={n_mask} != {n}")
+#     got = (count_frames(vdir / "rgb.mkv"), count_frames(vdir / "mask.mkv"))
+#     if got != (n, n):
+#         raise RuntimeError(f"encoded rgb={got[0]} mask={got[1]} != {n} (decode verify)")
+#     return w, h
+#
+#

H2O/shared-v1.0/script/provenance/previous-run/script/provenance/previous-run/script/provenance/operations/prepare-w4-c10.py

@@ -0,0 +1,52 @@
+"""Preserve completed r1 sequences without re-encoding or touching source files."""
+import datetime
+import json
+import os
+from pathlib import Path
+import shutil
+
+old = Path('/data/xuzhefeng/Datasets/uni-hoi-h2o-full-20260929-r1')
+new = Path('/data/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-w4-c10')
+snapshot = Path('/tmp/h2o-w4-c10-recovery')
+new.mkdir()
+shutil.copytree(snapshot / 'scripts', new / 'frozen-code/scripts')
+shutil.copytree(snapshot / 'docs', new / 'frozen-code/docs')
+shutil.copytree(snapshot / 'tests', new / 'frozen-code/tests')
+release = new / 'H2O/shared-v1.0'
+for directory in ('data', 'doc', 'script', 'script/provenance/r1', 'script/provenance/operations'):
+    (release / directory).mkdir(parents=True, exist_ok=True)
+history = release / 'script/provenance/r1'
+shutil.copytree(old / 'H2O/shared-v1.0/script', history / 'script')
+shutil.copytree(old / 'frozen-code/docs', history / 'docs')
+for name in ('full.log', 'full.status', 'run-full.sh'):
+    shutil.copy2(old / name, history / name)
+records = []
+for record in sorted((old / 'H2O/shared-v1.0/doc').glob('*.build.json')):
+    build = json.loads(record.read_text())
+    seq = build['sequence_id']
+    assert Path(seq).name == seq and seq not in ('.', '..')
+    assert set(build['views']) == {f'cam{i}' for i in range(5)}, seq
+    data = old / 'H2O/shared-v1.0/data' / seq
+    frames = {}
+    for cam in build['views']:
+        meta = json.loads((data / cam / 'metainfo.json').read_text())
+        frames[cam] = sum(1 for _ in (data / cam / 'camera/frames.jsonl').open())
+        assert frames[cam] > 0
+    shutil.copytree(data, release / 'data' / seq, copy_function=os.link)
+    shutil.copy2(record, release / 'doc' / record.name)
+    records.append({'sequence_id': seq, 'frames_by_camera': frames,
+                    'original_script_sha256': build['scripts']})
+manifest = {'time': datetime.datetime.now().astimezone().isoformat(),
+            'source_run': str(old), 'destination_run': str(new),
+            'reason': 'User requested four archive workers, ten threads per encoder, no CPU quota',
+            'reuse': 'Hard-linked immutable completed views; frames and component source SHA256 rechecked during resume',
+            'excluded': 'Unpublished .tmp directories; historical 62-sequence output; failed initial full run',
+            'original_runtime_provenance': 'script/provenance/r1', 'sequences': records}
+(release / 'doc/RESTART_PROVENANCE.json').write_text(json.dumps(manifest, indent=2) + '\n')
+for source, target in (('/tmp/h2o-prepare-w4-c10.py', 'provenance/operations/prepare-w4-c10.py'),
+                       ('/tmp/h2o-reconfigure-w4-c10.sh', 'reconfigure-w4-c10.sh'),
+                       ('/tmp/h2o-run-w4-c10.sh', 'run-full.sh')):
+    shutil.copy2(source, release / 'script' / target)
+shutil.copy2('/tmp/h2o-run-w4-c10.sh', new / 'run-full.sh')
+shutil.copy2('/tmp/h2o-w4-c10-frozen-tests.log', history.parent / 'worker4-codec10-regression.log')
+print(f'REUSED_SEQUENCES={len(records)} REUSED_VIEWS={sum(len(r["frames_by_camera"]) for r in records)}', flush=True)

H2O/shared-v1.0/script/provenance/previous-run/script/provenance/previous-run/script/provenance/operations/retry-w4-c10.sh

@@ -0,0 +1,11 @@
+#!/usr/bin/env bash
+set -euo pipefail
+base=/data/xuzhefeng/Datasets/uni-hoi-h2o-full-20260930-w4-c10
+script="$base/H2O/shared-v1.0/script"
+mkdir -p "$script/provenance/operations" "$script/provenance/startup-failure"
+mv "$script/prepare-w4-c10.py" "$script/provenance/startup-failure/prepare-w4-c10.py"
+cp /tmp/h2o-prepare-w4-c10.py "$script/provenance/operations/prepare-w4-c10.py"
+mv "$base/full.status" "$script/provenance/startup-failure/full.status"
+mv "$base/full.log" "$script/provenance/startup-failure/full.log"
+cp /tmp/h2o-retry-w4-c10.sh "$script/provenance/operations/retry-w4-c10.sh"
+systemd-run --user --scope --unit=h2o-full-w4c10-r2-20260930 -p CPUAccounting=yes /bin/bash "$base/run-full.sh"

H2O/shared-v1.0/script/provenance/previous-run/script/provenance/previous-run/script/provenance/r1/docs/CONTRACT.md

@@ -0,0 +1,457 @@
+# 手物数据仓库统一约定 v1.0
+
+格式版本：`hand-warehouse/1.0`。存储约定：`lossless-video-2026-09-18`。
+
+范围限于手物交互重建。按本任务纳入的源模态转换，不要求每个数据集填满所有组件。缺失模态、尚未下载、语义未核实分别报告，不能互相替代。
+
+本文定义各字段的含义与使用规则，机器可读注册表见 [schema.json](schema.json)。两者互相补充；如有冲突，应先核实并统一定义。
+
+## 1. 目录
+
+```text
+hand-datasets/
+  <dataset_name>/
+    raw/
+      doc/                         # 官方说明、来源与校验记录
+      script/                      # 官方下载脚本、必要包装脚本
+      data/                        # 保留原始包，不原地解压或改写
+    <dataset_version>/
+      doc/                         # 发布说明、schema、来源清单
+      script/                      # 转换、校验脚本及运行配置
+      data/
+        <sequence_id>/
+          <camera_id>/
+            metainfo.json
+            observation/
+              rgb.mkv
+              rgb_valid.pt
+              depth.mkv           # float32 深度的可逆视频编码
+              depth_valid.mkv     # 逐像素 bool 的无损视频编码
+            camera/
+              frames.jsonl
+              intrinsic.pt
+              intrinsic_valid.pt
+              extrinsic.pt
+              extrinsic_valid.pt
+            annotation/
+              mano/
+                motion.pt
+                motion_valid.pt
+                visibility.pt
+              keypoints/
+                uv.pt
+                uv_valid.pt
+                xyz.pt
+                xyz_valid.pt
+                visibility.pt    # 源数据有时提供
+              object/
+                <object_id>/
+                  assets/          # 该物体的网格、材质、纹理及部件资产
+                    <mesh_id>.obj
+                  pose.pt
+                  pose_valid.pt
+                  visibility.pt
+                  metainfo.json    # 物体身份、类别、mesh、尺寸与部件引用
+                  articulation.json
+                  keypoints/
+                  vertex_part_ids.pt
+                  joint_state.pt
+                  joint_state_valid.pt
+                  parts/<part_id>/pose.pt
+                  parts/<part_id>/pose_valid.pt
+              segment/
+                mask.mkv
+                mask_valid.pt
+                instances.jsonl
+              contact/
+                relations.jsonl
+                <subject_id>__<object_id>/  # 对应实体间的距离/接触区域
+```
+
+目录展示可能存在的组件，不要求填满。具体文件、字段和缺省规则分别见后续对应小节。原始包的解压发生在临时工作区，raw 内原始包保持不变。
+
+## 2. 相机视图的 metainfo.json
+
+### 2.1 顶层字段与发布信息
+
+每个相机视图必须有 `metainfo.json`。以下字段除另行说明外均为必填：
+
+| 字段 | 填写内容 |
+|---|---|
+| schema_version | 结构版本，本版为 `hand-warehouse/1.0` |
+| dataset_id | 数据集的逻辑标识 |
+| release_id | 本次数据发布版本，对应目录中的 dataset_version |
+| sequence_id | 所属序列或独立图片样本的标识 |
+| camera_id | 当前相机视图标识 |
+| sample_type | `image` 或 `video`，规则见 2.2 |
+| num_frames | 当前视图帧数 T，所有随帧组件按同一时间轴组织 |
+| hand_slots | 固定为 `["left", "right"]`；解剖学左右的使用规则见 5.1 |
+| sampling | 采样类型、源 FPS、帧索引引用及选择策略，见 2.2 |
+| camera | 当前相机模型与图像尺寸，见 2.3 |
+| components | 实际提供的组件及其文件、schema、有效性和来源，见 2.5 |
+| split | `train/val/test/unknown`，划分协议见 2.4 |
+| view_type | 可选，`ego/exo/unknown` |
+
+`dataset_version` 是数据发布版本；`schema_version` 是结构版本，两者独立。原始包校验摘要、官方来源和转换配置应记录在发布清单中。解压发生在临时工作区，原始包保持不变。
+
+### 2.2 图片、视频与 sampling
+
+`sample_type=image`：允许单图 T=1、`sampling.type=single`，也允许将独立图片批量装入一个视频容器，T>1 时 `sampling.type=independent`、`source_fps=null`。分片大小按数据集选择；FreiHAND 默认每片最多 4096 张，训练与评估分开，尾片不补帧。该数量不是其他数据集的强制要求；同批分辨率、相机模型类型及划分协议一致。每行 frames 必须有稳定且唯一的 `sample_id`、原始文件定位；内参、手形状等仍逐帧保留，不能因打包而共享错误的标定或身份。`sequence_id` 在这里仅是批次标识，不代表真实序列；不得按相邻帧推断跟踪、同步或接触延续。源 camera_id、image_variant、split 等随样本变化时放在 frames 的对应字段中。真实采集时间未知则 timestamp_ns=null。合成背景变体不视为真实多相机。
+
+`sample_type=video`：`sampling.type=dense/sparse`；`source_fps` 是真实采集 FPS，未知为 null。非连续取样必须保留源帧号和已知源时间，不可用采样后索引除以 FPS 冒充源时间。可另记采样策略与步长。
+
+shared 可以仅存所选 RGB 帧对应的数据。OakInk2 图像/动捕不同采样率时用 annotation_matches 明确映射；不要求将未采样动捕点硬塞进 T。多视角身份和源同步关系要保留；选择了哪些帧/相机及原因写在 sampling 中。
+
+`sampling.frames_ref` 指向 `camera/frames.jsonl`。帧号、逐帧时间、同步与标注匹配记录的结构见 4.1；MKV 封装时间的约定见 3.1。
+
+### 2.3 camera 模型声明
+
+`camera` 至少包含 `model`、`schema_id`、`image_size:[W,H]`。模型未知使用 `unknown` 和 `null`，不猜成针孔。实际内参、畸变及外参数组放在 `camera/`，各模型的固定结构见第 4 节。
+
+### 2.4 split 与划分协议
+
+保留上游划分，不将同一采集序列的相邻帧随意拆到不同集合。
+
+有官方划分时必填 split_scheme 和 split_source_locator，split_scheme 区分如 Dex-YCB s0/s1、ARCTIC P1/P2 等选定协议；不知道时 split=unknown 并给理由，不凭空分配。每发布可选择一种协议；存在逐帧差异时 frames 中的 split 覆盖视图默认值，并登记其来源。
+
+### 2.5 components 组件登记
+
+组件表的键是本发布内唯一名称，例如 `rgb`、`depth`、`mano`、`keypoints_uv`。每个已提供组件声明以下信息；模态特有参数在第 3–5 节对应小节中定义。
+
+| 字段 | 约定 |
+|---|---|
+| available | 本发布是否提供该组件；缺省也表示未提供 |
+| schema_id | 已登记的具体存储与语义类型，不得临时创造未登记名称 |
+| data_ref | 数据文件路径，相对当前相机视图根目录，使用 POSIX 路径 |
+| axes | 数组各维度的含义 |
+| valid_ref / valid_axes | 有效性文件及其维度；语义由所属组件规定 |
+| storage / valid_storage | 视频数据／视频有效性的存储配置；包含已登记 profile、实际 codec、pixel_format、encoded_size，见 3.1；非视频组件按原数组规则 |
+| provenance | source / format_conversion / derived，以及转换或生成方法 |
+| source_locator | 原文件、压缩包成员和记录／帧的定位信息 |
+
+不存在的组件允许缺省或写作 `{ "available": false, "reason": "..." }`，不用引用不存在的文件。静态网格可声明标量有效性；具体见 5.4。观测的 valid 见第 3 节，标注的 valid 与 visibility 区分见 5.1。
+
+`provenance.kind` 使用 source/format_conversion/derived，`method` 说明转换或生成方法；发布清单补充脚本版本、参数、输入摘要。`source_locator` 定位到原始文件/压缩包成员和记录、帧号；允许引用有完整溯源清单的前一发布，但链条最终应能回到 raw。不得仅写一个无法定位记录的数据集名字。
+
+`.pt` 仅保存 CPU Tensor，不保存 Python 自定义对象；浮点默认 float32，有效性为 bool，浮点无效位置填 NaN。所有随帧数组保留首维 T，单帧也不 squeeze；静态资产按其 schema 单独声明。单位、坐标及固定形状见各组件小节。
+
+机器注册表 `schema.json` 中，`components` 固定解码后的数据类型、形状、单位和语义；`storage_profiles` 固定视频内的编码／还原方式。不得临时创造未登记的 schema_id 或 storage.profile。同一语义允许选择其登记的不同无损存储 profile。
+
+## 3. observation 观测数据
+
+### 3.1 rgb.mkv 与 rgb_valid.pt
+
+RGB、深度和栅格 mask 在 shared 中统一使用 MKV 视频容器，不展开为逐帧图片或图像 tar。MANO、关键点、位姿等数值标注仍用数组；分割的重叠多边形仍用 JSONL，不为了视频化而有损栅格化。单图可为一帧视频，独立图批量封装规则见 2.2。
+
+**无损指共享层不增加信息损失。** 源 JPEG／H.264 等本身可能已是有损编码，原样保留并不恢复已丢失信息；不得重新以有损 JPEG、H.264、H.265、AV1 参数编码。源视频／图片的色彩、位深、分辨率和几何关系不得被静默改变。RGB schema `rgb/uint8/v1` 的统一读取结果是 `[T,H,W,3] uint8`、RGB 通道，有效性仍为 `rgb_valid.pt [T] bool`。超出该 schema 的高位深或 HDR 源必须报告并扩展定义，不能降到 8 位冒充无损。
+
+#### 3.1.1 允许的 RGB 存储 profile
+
+| storage.profile | 编码方式 | 适用规则 |
+|---|---|---|
+| mkv/mjpeg_copy/v1 | 源 JPEG 压缩帧直接封装成 MJPEG/MKV | 不解码重编码；保持每张 JPEG 码流字节；普通 8 位 JPEG 且 MKV/解码器支持时使用 |
+| mkv/source_copy/v1 | 已有视频的 RGB 对应码流直接封装为 MKV | 仅源编码能被 MKV 和选定解码器支持时使用，不伪造帧；需维持正确的呈现顺序与依赖帧 |
+| mkv/h264_rgb_lossless/v1 | libx264rgb 的真正无损模式 | 必须 RGB 路径、CRF=0，并通过全帧 RGB 字节比较；不能替换成默认 libx264 YUV420 路径 |
+| mkv/ffv1_rgb/v1 | FFV1 的 RGB 可逆编码路径 | 作为通用候选；解码回 RGB 后必须逐字节相同，不做色度降采样 |
+
+优先尝试源码流直接封装；不能封装或需要选择稀疏帧时，比较 FFV1 与无损 RGB H.264。对完整待发布片段，只在通过无损验收的候选中选择文件较小者，不假设某个编码器总最小；编码时间和随机读取代价记入构建记录。中间候选不作为 shared 多份常驻。H.264 跨帧压缩允许较长 GOP，必须记录 GOP／关键帧间隔，随机取帧需从前一关键帧正确解码，不能把 seek 返回的近似帧当目标帧。
+
+普通源视频可以保留它已有的有损码流，但不允许新增有损重编码。源是 VRS 等记录容器时，先查明图像流实际编码：已有兼容 JPEG／视频包可直接提取封装，否则解码后使用真正无损编码；不把整个 VRS 改后缀当 MKV。PNG 不要求直接塞进 MKV，可选上表两种无损重编码方式。
+
+#### 3.1.2 视频配置与校验
+
+每个视频组件在 metainfo 中登记 `storage`，至少包含：
+
+| 字段 | 内容 |
+|---|---|
+| profile | 上述登记的存储 profile；语义仍由组件 schema_id 指定 |
+| codec | 文件中实际 codec，例如 mjpeg、h264、ffv1；source_copy 必须写实际 codec |
+| pixel_format | 实际编码像素格式；不得仅凭编码器参数猜测 |
+| encoded_size | 视频物理像素尺寸 `[width,height]`；深度可能不同于逻辑尺寸，见 3.2 |
+| decoder | 解码器名称、版本，以及颜色矩阵／范围／旋转处理等实际像素解释配置；未知的源色彩声明 unknown，不擅自称为 sRGB |
+| encoder_parameters | 重编码时的编码器、版本、无损模式和 GOP 等参数；直接封装时记录 stream_copy |
+| source_lossy | 源是否已有损：true / false / unknown；不是共享层再有损编码的许可 |
+
+无损重新编码的 RGB 必须使用同一固定解码配置，比较**全部帧**的统一 RGB 字节、数量及顺序。MJPEG-copy 不重新编码，生产校验采用 `jpeg-packets-all-rgb-sample16/v1`：从 MKV 逐包取回，对全部源 JPEG 的 SHA256、数量、顺序和图像尺寸进行精确核对；每视图另用同一固定解码配置比较最多 16 帧的源与目标 RGB 字节（含首尾与四分位位置，其余由固定种子抽取）。记录策略、实际索引、种子与结果；不足 16 帧时全查。试点与最终完整序列随机抽查，对选中视图的全部帧复核解码 RGB。历史已做全帧像素检查的证明继续保留。
+
+逐包一致与像素抽查的范围必须分别声明，不能将抽查写成全量像素检查。不能用 PSNR 高、视觉相似、编码器参数称为 lossless 等替代精确检查。同时保留源编码的色彩标记和必要元数据；原图有 EXIF/色彩/方向等元数据时保留 raw 和必要来源信息；不能静默自动旋转图像而不更新像素对应与标定。以上加速策略仅适用于原 JPEG 码流直接封装，不适用于深度、标签或重新编码的 RGB。
+
+所有共享视频的呈现帧顺序与统一 frame_index 一一对应，每个栅格组件恰好 T 帧。MKV 的 `storage_fps` 只服务封装，声明 `timebase_role=index_only`；source_copy 可以保留源 PTS 以避免改坏码流，但真实时间与标注对齐仍以 frames 为准。未知真实时间保留 null。缺失某帧组件时保留索引位置并使用该组件的无效占位／valid 标识，不删帧；不能以重复相邻图像冒充已知观测。
+
+### 3.2 depth.mkv 与 depth_valid.mkv
+
+深度语义仍是 `depth/z_float32/v1`：统一读取结果 `[T,H,W] float32`，米制相机 Z 分量，`depth_type=z`。有效深度有限且大于 0，无效位置为 NaN。`depth_origin=measured/rendered/unknown` 必填，来源不明时 unknown。默认与 RGB 共用像素网格及相机参考系，不处理未注册 RGB-D。射线距离必须依据源定义转换成 z，不能只改声明。
+
+#### 3.2.1 float32 的可逆视频编码
+
+采用 `mkv/depth_float32_bits/v1`：**不是把深度量化成 uint16**，而是将规范化后每个 IEEE-754 float32 的完整 32 位拆成两个 uint16，使用 FFV1 gray16le 保存。视频物理尺寸为 `[2W,H]`：每行前 W 列存原 float32 位串的低 16 位，后 W 列存高 16 位。读取时先合并为 uint32 位串，再按 float32 重新解释，不能数值 cast 成 float32。
+
+```text
+u = reinterpret_uint32(depth_float32)
+encoded[:, :, :W] = uint16(u & 65535)
+encoded[:, :, W:] = uint16(u >> 16)
+u_restored = uint32(low) | (uint32(high) << 16)
+depth_float32 = reinterpret_float32(u_restored)
+```
+
+统一采用 little-endian 原始字节接口。`camera.image_size` 始终是逻辑 RGB 尺寸 `[W,H]`，不能改成编���视频的 2W；`storage.encoded_size=[2W,H]`。包括 NaN 在内的 float32 位串必须全帧完全恢复，不允许改成 float16、整数毫米、归一化灰度或 RGB 伪彩色。这里的无损验收基准是完成既定米制／坐标转换后的标准 float32 数组，不声称消除 float64→float32 等原约定转换本身的舍入。
+
+#### 3.2.2 逐像素有效性
+
+`depth_valid.mkv` 使用 `mkv/bool_ffv1/v1`，FFV1 gray8（FFmpeg pixel_format=`gray`），物理尺寸 `[W,H]`。像素只能为 0 或 1，统一解码为 `[T,H,W] bool`，逐像素含义不变。这样不再把庞大的 bool 数组保存为未压缩 `.pt`；标注的帧级／手级 valid 仍保留原有小数组。
+
+登记示例（W、H 以下取 640、480；decoder/encoder_parameters 的实际版本由转换器填写）：
+
+```json
+{
+  "available": true,
+  "schema_id": "depth/z_float32/v1",
+  "data_ref": "observation/depth.mkv",
+  "axes": ["T", "height", "width"],
+  "valid_ref": "observation/depth_valid.mkv",
+  "valid_axes": ["T", "height", "width"],
+  "unit": "m",
+  "depth_type": "z",
+  "depth_origin": "measured",
+  "invalid_value": "NaN",
+  "storage": {
+    "profile": "mkv/depth_float32_bits/v1",
+    "codec": "ffv1",
+    "pixel_format": "gray16le",
+    "encoded_size": [1280, 480],
+    "source_lossy": false,
+    "decoder": {"name": "ffmpeg", "version": "填写实际版本"},
+    "encoder_parameters": {"name": "ffv1", "level": 3, "version": "填写实际版本"}
+  },
+  "valid_storage": {
+    "profile": "mkv/bool_ffv1/v1",
+    "codec": "ffv1",
+    "pixel_format": "gray",
+    "encoded_size": [640, 480],
+    "source_lossy": false,
+    "decoder": {"name": "ffmpeg", "version": "填写实际版本"},
+    "encoder_parameters": {"name": "ffv1", "version": "填写实际版本"}
+  },
+  "provenance": {"kind": "format_conversion", "method": "源深度转米制 z float32；逐位打包为 FFV1"},
+  "source_locator": {"file": "raw/data/depth.zip", "member": "depth/0001.png"}
+}
+```
+
+深度和有效性视频都必须逐帧解码验收：float32 位串相同、bool 相同、T 相同，并检查 valid=true 的正有限值和 valid=false 的 NaN。读取器必须按 profile 还原，不能把 depth.mkv 的编码像素直接当深度值。
+
+## 4. camera 相机、帧索引与标定
+
+### 4.1 frames.jsonl：索引、时间与对应关系
+
+`camera/frames.jsonl` 每行对应统一帧号 `frame_index=0..T-1`，是所有组件的唯一行索引。至少包含 `frame_index`、`timestamp_ns`、`clock_domain`、`source_locator`；原始帧号/文件名保存在 `source_frame_id`，不能被新帧号替代。每行还可有 `annotation_matches`，按组件键保存 `source_frame_id`、`source_timestamp_ns`、`clock_domain`、`match_method` 和 `delta_ns`；方法为 exact/index_map/nearest/interpolated，索引映射无时间依据时 delta_ns=null。nearest 必须记录允许误差；interpolated 必须记录两端源标识和权重。delta_ns 仅在同一时间域中计算为源标注时间减图像时间。多相机共享 `sequence_id`；已知同步关系必须用 `sync_id` 或明确的源帧/时间映射表达，不能仅凭相同数组索引认定同步。
+
+每个视图必须提供此文件。源时间使用整数纳秒及 `clock_domain`，未知写 JSON `null`；不同时间域不能直接比较。逐帧 split 可以覆盖 metainfo 的视图默认划分，并保留来源，见 2.4。
+
+### 4.2 坐标系与标定通则
+
+| 项目 | 统一约定 |
+|---|---|
+| 相机坐标 | 右手系，X 向右、Y 向下、Z 向前，长度单位米 |
+| 齐次变换 | 列向量，`p_target = T_target_from_source @ p_source` |
+| 图像坐标 | `(u,v)`；左上像素中心为 `(0,0)`，u 向右、v 向下，单位 pixel |
+| 固定标定 | 本版展开为 T 份，同一 schema 不混用静态与动态形状 |
+| 分辨率或模型变化 | 分开存储为视图／片段，不在同一组件中混装 |
+| 未知标定 | 相机内外参允许缺省，不填单位阵冒充已知标定 |
+
+### 4.3 intrinsic.pt 与畸变参数
+
+| model / schema_id | intrinsic.pt 的固定结构 |
+|---|---|
+| pinhole / camera/pinhole/v1 | `[T,3,3]`，标准 K，最后一行为 `[0,0,1]`；限无畸变或已校正图像 |
+| fisheye624 / camera/fisheye624/v1 | `[T,15]`，`[f,cx,cy,k0,k1,k2,k3,k4,k5,p0,p1,s0,s1,s2,s3]`，采用 Project Aria Fisheye624 投影定义 |
+| opencv_fisheye4 / camera/opencv_fisheye4/v1 | `intrinsic.pt [T,3,3]` 保存 K；同组件 `distortion_ref` 指向 `[T,4]`，顺序 `[k1,k2,k3,k4]`，使用 OpenCV fisheye 等距角度多项式；已由 ReInterHand 实际原始相机参数验证 |
+| rational8 / camera/rational8/v1 | `intrinsic.pt [T,3,3]` 保存 K；同组件必需 `distortion_ref` 指向 `[T,8]`，顺序 `[k1,k2,p1,p2,k3,k4,k5,k6]` |
+| unknown / null | 不提供可用 intrinsic，不声称可以投影 |
+
+有畸变的其他模型必须注册新 schema，固定参数名、顺序和投影公式；不能塞进针孔 K 后遗失畸变项。`metainfo.camera.schema_id` 必须与 intrinsic 组件一致。同一模型的别名不可隐式改变参数布局。
+
+`opencv_fisheye4`：对前方点令 `x=X/Z, y=Y/Z, r=sqrt(x²+y²), theta=atan(r)`，`theta_d=theta*(1+k1*theta²+k2*theta⁴+k3*theta⁶+k4*theta⁸)`；`u=fx*x*theta_d/r+cx, v=fy*y*theta_d/r+cy`。光轴 r=0 时使用连续极限 `theta_d/r=1`。四系数不能按普通径向畸变解释，也不能补零冒充 Fisheye624。
+
+rational8 对归一化坐标 x=X/Z、y=Y/Z，令 r²=x²+y²，径向项为 `(1+k1*r²+k2*r⁴+k3*r⁶)/(1+k4*r²+k5*r⁴+k6*r⁶)`；再加 p1/p2 切向项并经 K 得到像素。`distortion_ref` 与 intrinsic 共用 `[T]` valid，保存到 camera/distortion.pt。不得把这种畸变三维坐标当作真实相机系 xyz 保存。
+
+内参有效性为 `intrinsic_valid.pt [T] bool`；有畸变时使用对应 `distortion_ref`。同一模型的参数名和顺序固定。
+
+### 4.4 extrinsic.pt 与 extrinsic_valid.pt
+
+`extrinsic.pt` 为 `[T,4,4]`，schema `transform/se3/v1`，方向固定 `T_camera_from_world`，有效性 `[T]`。世界坐标含义及源坐标转换记录在 provenance 中；只有相机坐标标注不要求虚构世界坐标。旋转块须正交且 determinant≈1。
+
+## 5. annotation 标注数据
+
+### 5.1 标注通则：手槽、valid 与 visibility
+
+手槽固定为 0=left、1=right，指解剖学左右，不按画面位置或可见性重排。本阶段只处理左右手两槽，不处理更多手。实体引用使用 `hand_left`、`hand_right`；手槽不证明跨采集片段的个人身份。
+
+#### 5.1.1 valid：标注是否可用于监督
+
+标注的 `valid=true` 表示：有对应实际标注、通过本 schema 数值检查，且通过来源明确提供并适用于该组件的有效性/QA 条件。`valid=false` 可由缺标注、无法建立帧对应、源无效占位或官方 QA 不通过导致；不等于画面没有手或手不可见。各组件独立判断，MANO 按 `[T,2]` 标记左右手，关键点按 `[T,2,K]` 标记。
+
+**2026-09-20 将标注 valid 从“仅表示存在”修订为“可用于监督”，字段形状不变。** 公式为 `source_annotation_present AND numerical_checks_pass AND applicable_official_QA_pass`。没有官方 QA 时省略最后一项，不伪造 QA；不能把未知当成已通过已存在的 QA。适配器在组件 `valid_policy` 中记录具体规则、来源、聚合/广播作用域。缺失官方必需 QA 的记录须明确处理为未知或暂停验收，不能静默通过。无效浮点项仍填 NaN，原始数据保留在 raw；被排除原值可用于 doc 中明确标识的诊断渲染，不能冒充 shared 有效监督。
+
+本地筛选须依据明确的规则，并在 `valid_policy.local_excluded_frame_indices`、排除原因和正式报告中记录。单个可疑投影作为问题记录，不直接硬编码排除该帧。已确认的曝光异常须明确作用范围，不能把一个序列的“前两帧有问题”套到所有序列。RGB 编码完好与几何监督可用是两件事，允许保留原图及帧索引而屏蔽对应标注；未发现问题不等于完成全部人工审核，试点待确认时不能作为已验收发布使用。
+
+源字段必须按官方语义映射，不能直接按名字复制：
+
+| 来源 | 官方含义 | 映射原则 |
+|---|---|---|
+| [H2O](https://github.com/taeinkwon/h2odataset) 的手标注首位 | 0 未标注、1 已标注 | 可作为对应手标注存在性的依据，不推断未标注原因 |
+| [HOT3D](https://github.com/facebookresearch/hot3d#using-mask-files) 的 pose_available、visible、mask_qa_pass | 分别表达姿态存在、可见性、对画面中可见手物位姿的人工 QA | 逐手/物体记录决定存在性，再 AND 当前相机当前帧 QA；QA 是整帧聚合判定，保守排除该帧全部手物几何监督，不声称每个实体均有错误；源可见性基于位姿推导时，其自身 valid 同样受 QA 约束 |
+| [Dex-YCB](https://github.com/NVlabs/dex-ycb-toolkit#loading-dataset-and-visualizing-samples) 的全零 pose_m、全 -1 joints | 不可见的手或缺标注共用占位表示 | 对应组件 valid=false；不能从该值判定 visibility=false，也不能声称原因是动捕缺失 |
+| [ARCTIC](https://github.com/zc-alexfan/arctic/blob/master/docs/data/data_doc.md) 的 is_valid/left_valid/right_valid | 是否有效用于官方评估 | 核实每项筛选依据；不把可见性/评估选择直接等同 QA。本次原始 mano.npy 不含这些派生评估标记，按原记录与数值检查生成 valid，并声明没有额外官方 QA 输入 |
+
+源 valid=false 若代表评估筛选或其他条件，应检查实际标注和源约定后映射；不能一律清空坐标。反之，占位零或 -1 只有在源官方明确规定时才判为缺失，不泛化成所有数据的判断规则。
+
+`valid=true` 的标注应通过其 schema 的数值与形状检查；检查失败必须报告或修复，不能静默隐藏。QA 不通过应报告排除数量与原因，并保留可追溯信息；不能把这类无效称为缺标注。RGB、深度等观测有效性仍按各自 schema；手物 QA 不自动否定 RGB、静态相机内参或物体资产。相机外参是否过滤须依据其自身来源规则；HOT3D 本次保留源外参用于溯源和诊断，不能用其存在性证明 QA 失败帧的几何监督正确。
+
+#### 5.1.2 visibility：可见性属性
+
+附加质量属性只保留 `visibility`，存放在对应模态或标注集内，不设 quality 目录。同一份可见性仅存一次；uv 与 xyz 共用时引用同一个文件。
+
+visibility 通过所属组件 attributes 登记 data_ref、axes、dtype、visibility_type、meaning、provenance、source_locator；未知值通过属性 valid_ref 标识，全已知可声明 valid=true。比例与布尔值不相互强转，也不按任意阈值二值化。没有 MANO 时，整手可见性挂在 keypoints 组件，使用独立的 whole_hand_visibility.pt，明确 meaning=whole_hand，不与逐关键点可见性混用。
+
+`valid` 回答“这项标注能否用于监督”，`visibility` 回答“在当前图像中是否可见（或可见多少）”。不可见的手或关键点仍可有有效标注；不能因 visibility=0 就清空坐标。源没有 visibility 时缺省或标为未知，不由 valid 推断。
+
+| 情况 | 手标注 valid | visibility |
+|---|---|---|
+| RGB 已开始录制，动捕尚未开始，没有其他对应手标注 | false | 以独立源可见性标注为准；没有则未知 |
+| 手被遮挡，但源有标注且通过适用有效性检查 | true | 若源声明不可见，则 false 或源可见比例 |
+| 手可见且有对应标注，通过适用 QA | true | 若源有此信息，则 true 或源可见比例 |
+| 手可见，但该帧漏标或动捕中断 | false | 可以为 true，但必须有独立来源依据 |
+| 有手部参数，但官方 QA 明确不通过 | false | 不据此断言不可见；源可见比例若基于不可靠位姿，其属性 valid 也为 false |
+
+visibility 自己的 valid_ref 表示该可见性标注是否可用，不复用 MANO 或关键点坐标的 valid_ref；明确记录源存在性及适用的 QA 条件。上述例子不要求凭图像自动生成可见性。
+
+不设 quality 组件或目录，不另存拟合误差、曝光、截断比例等通用质量数组。适用的官方 QA 合并进对应 valid，在 `valid_policy` 记录来源与规则；正式校验报告保留原始 QA 时间区间及排除统计。原始 QA 文件仍在 raw，官方 split 保留。
+
+### 5.2 mano：参数与整手可见性
+
+MANO schema `mano/axis_angle61/v1`：`motion.pt [T,2,61]`，顺序为完整 16 关节 axis-angle 48、shape betas 10、全局平移 3。姿态使用弧度，平移米，位于当前相机系。有效性 `[T,2]`；缺少一侧整槽 NaN。PCA 必须先按对应模型展开，不能将 PCA 系数直接填入 axis-angle。
+
+`model_profile` 必须固定 MANO 版本、`flat_hand_mean=true`、左手 shapedirs 修正状态。平移表示模型输出顶点所加的全局平移，不是某个腕关节的坐标；源语义不同须转换。shape 无量纲。模型权重不作为共享数据附带上传。
+
+`mano/visibility.pt [T,2]` 保存源整手可见比例时使用 float32 [0,1]，`visibility_type=ratio`；只有实际采用 0–100 的百分比才除以100，不能只凭表头判断。HOT3D 的 `visibility_ratio[%]` 实值已经是0–1，官方读取器原样读取，必须保留原值。通过 MANO 组件 attributes 登记，使用独立有效性，不复用 motion_valid。
+
+手部共享表示仅保留 MANO 参数与选定关键点，不发布独立手表面几何。需要渲染或接触计算时由 MANO 参数生成，几何引用使用固定 MANO 模型配置与顶点顺序；不影响 object 下的物体模型和材质资产。
+
+### 5.3 keypoints：二维点、三维点与逐点可见性
+
+关键点独立于 MANO 参数存在；没有 motion 也可有关键点。`uv.pt [T,2,K,2]` 与 `xyz.pt [T,2,K,3]`，有效性 `[T,2,K]`，分别采用 `keypoints/uv/v1` 和 `keypoints/xyz/v1`。K 由 `keypoint_schema` 固定，不能在同一数组中逐帧变化；同体系统一顺序。数值使用像素坐标或米制相机系。
+
+| keypoint_schema | K 与固定索引 |
+|---|---|
+| mano21/v1 | 21；wrist 后依次 thumb/index/middle/ring/little，每指由近端至指尖 |
+| rhd21/v1 | 21；保持 wrist，然后五指由近端至指尖；原始每指 tip→palm 反序。只统一顺序，不宣称解剖学位置等价于 MANO |
+
+具体名称与源索引映射见 `schema.json`。当前只登记 MANO21 与 RHD21；不登记 MediaPipe、ReInterHand capture21 或 fit24 中间关键点。
+
+每个发布中的每个视图只提供一套标准手关键点，直接放在 keypoints 下，不设 annotation_set_id 字段或目录。uv 与 xyz 是这套关键点的二维与三维表示，不是两套标注。provenance 分别记录二者来自原始标注还是计算结果；不要求已有 2D 必须由当前 xyz 重新投影覆盖。
+
+ReInterHand 的 Original keypoints���Original fits keypoints 及带前臂关键点不纳入 shared；如需手关键点，从选定 MANO 参数生成 MANO21 并记录模型/指尖回归约定。其他数据集默认使用选定的官方关键点，不额外生成平行的 MANO 派生集合。后续主动更换生成方法或结果时发布新的 dataset_version，不在同一版本内增设方法分支。
+
+manual/projection/fitting 只记录在 provenance，不作为目录或类型分支。不得通过坐标为零、是否出画或 visibility 判断标注是否存在。派生标注仅在实际生成并通过数值检查后声明 valid=true；没有生成的项为 false，不伪造坐标。
+
+`keypoints/visibility.pt [T,2,K]` 保存源逐点可见性，布尔表示使用 `visibility_type=boolean`。uv 与 xyz 共用时引用同一份文件。整手比例属性没有 MANO 可挂载时使用 `whole_hand_visibility.pt [T,2]`，`meaning=whole_hand`；它与逐点可见性分开，通用登记规则见 5.1。
+
+不使用 annotation_set_id；ReInterHand 中间关键点集合不进入 shared，raw 保留来源文件。
+
+### 5.4 object：身份、资产、位姿与几何
+
+#### 5.4.1 metainfo.json：物体身份与跟踪作用域
+
+手用 `hand_left`、`hand_right` 及槽位 0、1 表示，不再另存 entities.json。物体身份登记在 `annotation/object/<object_id>/metainfo.json`，必填 `object_id`、`identity_scope`、`scope_id`、`source_locator`；可含 `category_id`、`source_object_id`、`mesh_ref`、`has_model`、部件引用。即使只有分割、没有 pose/mesh，也可只有这份物体元信息。
+
+`identity_scope` 为 frame/subsequence/sequence，`scope_id` 指定具体作用域；跨相机共享身份只在已有源对应证据时使用。VISOR 按官方 subsequence 与实体名关联时记录 `association_basis=official_subsequence_entity_name`，不能声称这是全视频物理实例跟踪。无需与 object_id 一一对应的标准 track_id；官方独立轨迹标识存在时保留为 source_track_id，并声明作用域。手槽不证明跨采集片段的个人身份。
+
+#### 5.4.2 assets：网格、材质与纹理
+
+资产属于物体，统一存放在 `annotation/object/<object_id>/` 内，可用 `assets/` 子目录组织网格、材质、纹理和部件资产。序列层不设共享 assets。所有引用采用视图根目录为基准；其中物体资产组件的 `data_ref` 与部件的 `mesh_ref` 采用相对于当前视图目录的 POSIX 路径，例如 `annotation/object/cup_01/assets/cup.obj`，且必须指向所属物体目录内的资产。OBJ 内部的材质、纹理引用遵循文件格式自身的相对路径规则，依赖文件一并保留。不同相机视图允许各自保存物体资产，不要求跨视图去重；下载物体时包含其资产依赖。
+
+网格 schema 为 `mesh/obj/v1`，顶点米制，静态有效性可直接声明标量 `valid`，或通过 `valid_ref` 指向标量 bool Tensor；不按 T 重复静态资产。由物体／部件通过引用关联，不仅凭文件名猜测。
+
+#### 5.4.3 pose.pt、pose_valid.pt 与 visibility.pt
+
+刚体 schema 为 `object/rigid/v1`，`pose.pt [T,4,4]` 固定为 `T_camera_from_object`，平移米制，有效性 `[T]`。物体坐标原点、轴向必须与 mesh 一致；旋转块为正交矩阵且 determinant≈1。
+
+`object/<id>/visibility.pt [T]` 保存源物体可见比例，float32 [0,1]，`visibility_type=ratio`。通过所属物体组件 attributes 登记，独立表达其有效性。
+
+#### 5.4.4 articulation.json、部件位姿与 joint_state.pt
+
+关节物体：schema `object/articulated/v1`，`articulation.json` 登记 `parts` 与 `joints`。每个 part 有 `part_id`、`parent_id`（根为 null）、`has_model`、`mesh_ref`；has_model=false 时 mesh_ref=null，结构节点允许没有网格。父链不得成环。每个 joint 有唯一 `joint_id`、`parent_part`、`child_part`、`type`；fixed 单位 null、revolute 弧度、prismatic 米。有源限位时保存 `limits:[lower,upper]`；没有时 null，不伪造上下限。部件姿态统一为 `T_camera_from_part [T,4,4]`，不与父局部变换混装。
+
+若提供 `joint_state.pt [T,J]`，必须登记固定 joint_order、轴在父部件系的表达、零位变换与正方向，valid 为 `[T,J]`；缺少这些运动学定义时只发布部件位姿与已有来源信息，不伪造可驱动的关节模型。关节值与部件位姿同时存在时需验证一致性。
+
+部件层级应声明语义：纯语义包含树允许 `joints=[]`，不将包含关系推断为机械关节。关节状态 schema 为 `object/joint_state/v1`；转动用弧度、移动用米，混合类型按 joint_order 对应的关节定义解释。
+
+#### 5.4.5 dimensions、几何关键点与顶点部件标签
+
+`dimensions.pt [T,3]` 采用 `object/dimensions/v1`，有效性为 `[T]`。
+
+物体/部件 `dimensions` 为 `[length,width,height]` 米，轴顺序与该 mesh 的局部坐标关联，登记 `dimensions_semantics` 和来源。HOI4D dimensions 与 mesh 的具体尺度关���未验证前标为 source_unspecified，保留源值并暂停依赖该关系的几何验收，不能默认为待施加缩放。需要施加缩放时明确 scale_applied_to_mesh，禁止重复应用。
+
+源物体关键点放 `object/<id>/keypoints/xyz.pt [T,K,3]`（相机系）或 `template_xyz.pt [K,3]`（物体系），schema 分别为 object_keypoints/camera/v1 和 object_keypoints/template/v1。明确 keypoint_schema、keypoint_names、单位及有效性；模板点配合 pose 可投影，无须重复生成全部帧。H2O 中心/角点/棱中点，以及 ARCTIC 部件点分别登记具体源顺序，不套用手的 21 点 schema。每个发布选择一套源几何点并记录选择，不另建方法分支。
+
+`vertex_part_ids.pt [V] int64` 绑定 `mesh_ref` 与 topology_id，登记整数到 part_id 的映射；未知用 -1。ARCTIC 不同拓扑的资产分别绑定，不能共用未经转换的顶点标签。部件层级中 is_instance/has_model 等源属性保留，几何文件仍全部归 object。
+
+### 5.5 segment：分割类别、编码与覆盖范围
+
+每个发布选择一套任务使用的分割表示，不建立方法／标注集分支。没有分割就不创建 segment 目录，不为其他模态创建空的实体表。
+
+| schema_id | 文件与数值语义 |
+|---|---|
+| segment/uint16_labels/v1 | mask.mkv；整数区域编号，通过 instances.jsonl 解释 |
+| segment/instance_polygons/v1 | polygons.jsonl；实例多边形，可重叠；当前不支持孔洞拓扑 |
+| segment/foreground_intensity/v1 | mask.mkv；0–255 前景强度，不是区域编号或经过校准的概率 |
+
+#### 5.5.1 uint16 标签图与 instances.jsonl
+
+组件声明 `segmentation_type=semantic/instance/part`；同一组件固定一种类型。`mask.mkv` 解码为 `[T,H,W] uint16`，schema `segment/uint16_labels/v1`，使用 `mkv/uint16_ffv1/v1`，FFV1 gray16le 无损；只有所有区域编码均 ≤255 时允许 gray8（pixel_format=gray），读取后转 uint16，数值不变。禁止调色板和有损编码。有效性 `[T]`，表示该帧有可用分割标注。数值 0 固定为 unassigned，65535 固定为 ignore，1–65534 为区域编码，不能直接等同 entity_id。
+
+`instances.jsonl` 每帧一行，包含 `frame_index` 与 `labels`。每条标签包含 `mask_value`、`category_id`、`entity_id`、`part_id`；无意义或未知字段为 null。semantic 允许 entity_id=null；instance 必须引用 hand_left/hand_right 或登记的 object_id；part 还须关联登记部件。源实例 ID 和其他属性分别保存在 `source_instance_id`、`source_attributes`。mask 中每个非保留编码必须有且仅有一个映射。
+
+FreiHAND 的试点按本任务需要，统一到手／背景类别，保留每个原生区域编码；不要求给原生细分区域猜测手指名称。需要细部件监督时，尚未核实的部件名称必须保持未知，不能因粗粒度手类别已核实便声称部件语义也已统一。
+
+#### 5.5.2 重叠实例多边形
+
+每发布选择一套任务使用的分割表示，不建立方法/标注集分支。VISOR 的实际试点存在重叠实例多边形，单层标签图不能无损表达。对此补充 `segment/instance_polygons/v1`：`segment/polygons.jsonl` 每帧包含 `frame_index` 和 `instances`，每实例包含 `mask_value` 与 `polygons`（列表，每个多边形为源图像像素坐标顶点序列，保留原始精度）。身份和类别仍由同一 `instances.jsonl` 映射，有效性仍为 `[T]`，通过 schema 区分物理编码。允许多实例重叠，不采用后画覆盖前画。消费者按需栅格化为独立实例二值 mask；不能声称合并的单层 mask 无损。
+
+#### 5.5.3 灰度前景强度
+
+ReInterHand 的官方渲染 mask 为双手合并的灰度前景掩码，实测存在 0–255 的中间值，不是左右手实例编号。登记 `segment/foreground_intensity/v1`：仍存 `segment/mask.mkv`，解码为 `[T,H,W] uint16`，使用 `mkv/uint16_ffv1/v1`，此类数据优先编码为 FFV1 gray8（pixel_format=gray），读取后转 uint16。有效值范围 0–255，归一化值为 `float32(value)/255`；0 为背景端点、255 为前景端点。组件声明 `foreground_category=hand`、`entity_id=null` 和双手合并语义，不使用区域编号映射。这里保留原始渲染 mask 强度，不把它称为经过校准的概率。若任务需要二值 mask，阈值化属于任务层派生操作，记录阈值；共享层不丢弃中间灰度。不同 schema 的数值含义不同，不能将该文件按整数类别图解码。
+
+FreiHAND 训练集的 JPEG mask 使用 `segment/foreground_intensity/v1`，逐像素保留 JPEG 解码后的灰度，`entity_id=hand_right`。该 schema 不限定双手合并：单手或合并范围由 `entity_id` 与 `coverage` 明确声明。评估集整数区域图仍用 `segment/uint16_labels/v1`，两者不能仅凭文件名使用同一种解码语义。
+
+#### 5.5.4 coverage 与有效性
+
+`coverage.kind=full/selected_entities/partial/unknown`。full 指声明���标注范围完整；selected_entities 还应列出目标范围。无论哪种，0 本身不自动成为背景；背景要有明确类别映射，未标注区域不能用于构造负样本。保留源逐类/逐实例 exhaustive 信息，未知不可改写为完整。
+
+三种表示的帧级有效性均为 `[T]`，表示该帧是否提供此分割标注；对应 valid_ref 在组件中登记，仍用小型 bool `.pt` 数组。栅格 mask 必须登记 storage，格式见 3.1.2，所有帧的整数编码或强度逐值完全相同。编码语义由 schema_id 决定，读取器不能仅依据 mask.mkv 文件名选择解码方式。
+
+### 5.6 contact：接触、抓取、区域与距离
+
+#### 5.6.1 relations.jsonl 与 relations_valid.pt
+
+`annotation/contact/relations.jsonl`，schema contact/relations/v1，每帧一行含 frame_index、relations。每条记录含 relation_type（contact/grasp）、subject_id（手槽或登记实体）、object_id、status、source_locator、provenance。status 固定为 positive/negative/unknown/target_unannotated：positive 必须有可解析目标，negative 表示源明确否定该关系，unknown 为不确定，target_unannotated 为确有接触但目标未标注。非 positive 的 object_id 可 null，不能用 null 冒充“不接触”。稀疏源无记录的帧不自动补 negative。
+
+VISOR 的对象 mask ID 必须先按同帧 source_instance_id 解析成 object_id；hand-not-in-contact、none-of-the-above、inconclusive 分别映射 negative、target_unannotated、unknown，保留原始字符串。Dex ycb_grasp_ind 通过 ycb_ids 映射到登记物体，以 relation_type=grasp 保存；不得当作逐顶点接触。
+
+relations 组件的 valid_ref 指向 `relations_valid.pt [T] bool`，表示该帧是否有可用源关系标注。有效帧中可以存在 status=unknown 的明确不确定标注；无源标注的帧为 valid=false、relations=[]。组件仍声明 available/provenance/source_locator。
+
+#### 5.6.2 实体对的顶点区域与距离
+
+已发布的接触区域/距离按参与关系的 subject_id 与 object_id 组织，每对实体保留选定的一套结果，以精确 geometry_ref/topology_id 关联到手或物体/部件；不要求每套数据都有。顶点区域 schema contact/vertex_region/v1：bool `[T,V]` 加同形 valid；顶点距离 schema contact/vertex_distance/v1：float32 `[T,V]`、米、同形 valid，声明 source/target geometry_ref、distance_definition（signed/unsigned、测量到点集或表面等明确源定义）。geometry_ref 必须指向已登记 mesh 或由固定 MANO 模型配置确定的拓扑；拓扑/顶点数变化时按拓扑拆分片段。其他源区域表示只有在实际接入并登记 schema 后才可转换，不能随意栅格化宣称等价。
+
+顶点区域／距离已有存储定义，但尚未完成真实输入试点；接入时必须提供源语义与几何验收证据，不因已经登记 schema 就视为通过。
+
+## 附录：任务范围
+
+不强制收录 2D bbox、完整高频动捕、动作程序、身体 SMPL-X、点云或网络特征。范围内的接触、visibility、选定手关键点、物体几何不能因不便处理而静默丢弃。分割采用的物理编码由第 5.5 节决定，存在重叠时不能以单层标签图声称无损。
+
+不设置 entities_ref/segment/entities.json，不要求与 object_id 重复的 track_id。MediaPipe 不在当前关键点定义中；旧版本迁移必须保留可追溯关系，不能直接删表。手部共享表示不含独立 hand_mesh。
+
+schema.json 是普通注册表，不是 JSON Schema 文档或可执行验收器；定义存在不代表真实数据已验证。验收须保留实测范围与未决项。
+
+本存储约定只更新标准，不代表既有 FFV1／depth.pt 试点已迁移。新发布必须显式写 storage 配置；旧产物按其旧定义读取，不能只改文件后缀或版本标识伪装为合规。