Download toolbox/reference.py from yxma/React: direct link, hf CLI and curl.
- Browser
- Download file 2.85 kB
-
https://huggingface.co/datasets/yxma/React/resolve/fd3fe6bfe5e7783d1685886905ef0f63d74e43c4/toolbox/reference.py
- Command line
-
hf download hf://datasets/yxma/React@fd3fe6bfe5e7783d1685886905ef0f63d74e43c4/toolbox/reference.py
-
curl -L -o reference.py https://huggingface.co/datasets/yxma/React/resolve/fd3fe6bfe5e7783d1685886905ef0f63d74e43c4/toolbox/reference.py
2.85 kB
| """Reference (background) frame handling + difference image for GelSight. | |
| GelSight Mini is markerless, so background subtraction is clean (no markers | |
| to occlude). Convention follows Sparsh (Meta, CoRL 2024): a single no-contact | |
| reference frame + signed per-channel difference. | |
| All frames are RGB uint8 (H, W, 3) — the convention returned by the dataset's | |
| video decoder. | |
| """ | |
| from __future__ import annotations | |
| import numpy as np | |
| def get_reference(frames, mode: str = "p01", intensity=None, window: int = 30): | |
| """Pick a no-contact reference frame from a sequence. | |
| frames: (T, H, W, 3) uint8 OR a callable i->frame for lazy access with | |
| an explicit `intensity` array. | |
| mode: | |
| "first" -> frames[0] | |
| "p01" -> the 1st-percentile-intensity (quietest) frame; needs | |
| `intensity` (T,) array of per-frame contact intensity, | |
| else computes a cheap proxy = mean abs deviation from | |
| the temporal median. | |
| "running_avg" -> mean of the `window` lowest-intensity frames. | |
| Returns an (H, W, 3) uint8 reference. | |
| """ | |
| arr = np.asarray(frames) | |
| T = arr.shape[0] | |
| if mode == "first": | |
| return arr[0].copy() | |
| if intensity is None: | |
| med = np.median(arr.reshape(T, -1).astype(np.float32), axis=0) | |
| intensity = np.abs(arr.reshape(T, -1).astype(np.float32) - med).mean(axis=1) | |
| intensity = np.asarray(intensity, np.float32) | |
| if mode == "p01": | |
| thr = np.percentile(intensity, 1) | |
| idx = int(np.where(intensity <= thr)[0][0]) if (intensity <= thr).any() else int(intensity.argmin()) | |
| return arr[idx].copy() | |
| if mode == "running_avg": | |
| k = min(window, T) | |
| order = np.argsort(intensity)[:k] | |
| return arr[order].mean(axis=0).round().astype(np.uint8) | |
| raise ValueError(f"unknown reference mode: {mode}") | |
| def difference(frame, reference, signed: bool = True): | |
| """Sparsh-style difference image. | |
| signed=True -> (frame - ref)/255 + 0.5, clipped to [0,1], scaled to uint8. | |
| Mid-gray = no change; preserves direction of deformation. | |
| signed=False -> |frame - ref| as uint8 (magnitude only). | |
| frame/reference: (H, W, 3) uint8. Returns (H, W, 3) uint8. | |
| """ | |
| f = frame.astype(np.float32) | |
| r = reference.astype(np.float32) | |
| if signed: | |
| d = (f - r) / 255.0 + 0.5 | |
| return (np.clip(d, 0, 1) * 255).astype(np.uint8) | |
| return np.clip(np.abs(f - r), 0, 255).astype(np.uint8) | |
| def l2_diff(frame, reference): | |
| """Per-pixel L2 distance across RGB channels (H, W) float32. | |
| This is the quantity the dataset's contact scalars are built on: | |
| d[h,w] = ||frame[h,w,:] - ref[h,w,:]||_2. | |
| """ | |
| return np.sqrt(((frame.astype(np.float32) - reference.astype(np.float32)) ** 2).sum(axis=2)) | |