React / preprocess /README.md
yxma's picture
preprocess: add detect/curation/previews modules (curate CLI; ports verified bit-exact against published outputs)
f91b300 verified
|
Raw
History Blame Contribute Delete
5.27 kB

react_preprocess — building the React release

The producer side of the dataset. toolbox/ (react_toolbox) reads the published data; this package is what turns raw rig recordings into it, and ships alongside the data so the release is reproducible.

python -m react_preprocess build --task pushT [--with-depth]
python -m react_preprocess audit --task pushT
python -m react_preprocess backfill-flags --task pushT
python -m react_preprocess verify-flags --task pushT --against h5

Pipeline

recording.h5
   │  h5io      read timestamps/poses, resolve tactile↔camera alignment
   │  tactile   pass 1: pick the p01 no-contact reference
   │            pass 2: contact metrics + new-frame flags + encode
   │  encode    H.264 yuv444p CRF18 (RGB), FFV1 gray16le (depth)
   │  meta      per-frame parquet
   ▼
release/<task>/{videos,depth,meta}/<date>/episode_NNN/…
Module Responsibility
config paths, camera mapping, encoding and contact constants
h5io source reading, pose alignment, tactile time alignment
contact contact metrics, p01 reference, duplicate-frame detection
encode ffmpeg writers
tactile two-pass GelSight processing
meta parquet assembly and index columns
detect bad-interval detectors + clean-span complement
curation per-task bad_frames.json / segments.json / episodes.jsonl
previews preview policy (calibration choice, trim, world offset, layout)
pipeline per-episode orchestration
backfill recover flags for already-published parquet
publish mirror data + code to the Hub

previews holds policy only — the panel renderer needs rig-local calibration the release does not ship, so it stays in twm/scripts/build_release_previews.py as a thin adapter over previews.plan(). Port checks: detect/curation reproduce the published bad_frames.json and segments.json for all 36 episodes with zero differences; the preview adapter re-renders pushT/episode_000 bit-identically (first-frame MAD 0.00, same 900 frames).

Tactile time alignment

How a GelSight frame is paired with a camera frame depends on the recording:

legacy (≤ 2026-06-18) timestamped (2026-06-27 →)
Pairing by tick index nearest capture timestamp
Systematic lag ~15 frames (0.5 s) removed at the source
Constant shift needed yes no — would double-correct

The rig used to decode full 8 MP MJPG frames on the capture thread (~71 ms each), so tactile really ran at ~8 fps while rows were written at 30 Hz. That produced both the lag and heavy frame duplication. The rig now decodes at reduced scale and records gelsight/<side>/timestamps.

h5io.TactileAlignment.needs_legacy_shift is the guard: it is False for timestamped recordings, so a latency shift can never be applied twice.

tactile_*_is_new

The GelSight Mini tops out at 18.75 fps while parquet rows are written at 30 Hz, so some rows necessarily repeat the previous tactile frame. Legacy recordings repeat far more. These boolean columns mark the rows that are genuinely fresh readings:

df = pq.read_table("episode_000.parquet").to_pandas()
fresh = df[df.tactile_left_is_new]        # train tactile dynamics on these

Measured — legacy over the whole published release (480 080 rows, 36 episodes), fixed rig over test/2026-06-29/episode_001 (1 294 rows, both sensors):

capture rate duplicate rows effective rate longest frozen run
legacy ~8 fps (decode-bound) 71.8 % 8.5 fps 30 frames (1.0 s)
fixed rig 19.3 fps 39.5–41.5 % 17.6–18.2 fps 5 frames (0.17 s)

The residual ~40 % on the fixed rig is irreducible: a 19 fps sensor sampled onto a 30 Hz row clock must repeat roughly a third of its rows. Only the legacy excess above that was a bug.

How the flags are recovered for already-published data. A repeated frame yields a bit-identical contact triple, so a row is fresh exactly when its triple differs from the previous row's — no video decode required. verify-flags --against h5 checks this against source pixels: on all seven audited episodes (4 pushT + 3 motherboard) it reproduces the source frame-by-frame with 0 mismatches in 899 frames each, and independently recovers the +15 shift baked into the release (every other offset in 0..20 disagrees on >33 % of frames, so the detection is unambiguous).

Checking against the published MP4s can only ever be approximate — H.264 is lossy, so a duplicated frame does not decode back to identical pixels. --against video therefore compares with a tolerance and reports the observed separation rather than asserting exactness.

Conventions

  • Frame i of every MP4 == parquet row i == source frame trim_offset + i
  • Cameras: cam0 → view_right, cam1 → view_left, cam2 → view_middle (verified against calibration serials)
  • Tactile frames are RGB in HDF5 and converted to BGR only for ffmpeg
  • Depth is uint16 millimetres, 0 = no return
  • The 2026-05-19 motherboard world-origin offset is baked into stored poses

Paths come from REACT_DATA_ROOT / REACT_STAGE_ROOT when set, so the package runs off the rig.