React / toolbox /contact.py
yxma's picture
Add react_toolbox: VBTS utilities (reference/contact mask/approx depth/viz/calibration/actions) + quickstart + demo montage. MIT.
8420318 verified
Raw
History Blame Contribute Delete
2.11 kB
"""Contact detection for markerless GelSight.
Provides the classical, calibration-free contact pipeline:
diff-from-reference (L2) -> threshold -> largest connected component -> mask,
plus the exact contact scalars the dataset ships (intensity/area/mixed) so
users can reproduce / recompute them.
"""
from __future__ import annotations
import numpy as np
from .reference import l2_diff
TAU = 8.0 # dataset default L2 threshold (uint8 scale)
def contact_metrics(frame, reference, tau: float = TAU):
"""Reproduce the dataset's per-frame contact scalars.
d = ||frame - ref||_2 (per pixel, over RGB)
intensity = mean(d)
area = mean(d > tau)
mixed = mean(d * (d > tau))
Matches data/<task>/meta/*.parquet tactile_*_{intensity,area,mixed}.
"""
d = l2_diff(frame, reference)
above = d > tau
return {
"intensity": float(d.mean()),
"area": float(above.mean()),
"mixed": float((d * above).mean()),
}
def contact_mask(frame, reference, tau: float = TAU, largest_only: bool = True,
min_area_px: int = 50):
"""Binary contact mask (H, W) bool.
diff-from-reference L2 > tau, then (optionally) keep only the largest
connected component and drop specks < min_area_px. No calibration needed.
"""
import cv2
d = l2_diff(frame, reference)
mask = (d > tau).astype(np.uint8)
if int(mask.sum()) < min_area_px:
return np.zeros(mask.shape, bool)
if not largest_only:
return mask.astype(bool)
n, lbl, stats, _ = cv2.connectedComponentsWithStats(mask, connectivity=8)
if n <= 1:
return np.zeros(mask.shape, bool)
areas = stats[1:, cv2.CC_STAT_AREA] # skip background (label 0)
keep = int(areas.argmax()) + 1
out = lbl == keep
return out if out.sum() >= min_area_px else np.zeros(mask.shape, bool)
def contact_centroid(mask):
"""(row, col) centroid of a contact mask, or None if empty."""
ys, xs = np.nonzero(mask)
if len(ys) == 0:
return None
return float(ys.mean()), float(xs.mean())