constructelligence's picture
Upload completion.py with huggingface_hub
27db08f verified
Raw History Blame Contribute Delete
4.07 kB
"""Is this wall finished, and if not, where does the robot go back to?
Coverage from the argmax mask is the wrong number to sign a wall off on: the
first mixed test set showed the argmax declaring 81% of unfinished walls done.
Here a pixel counts as done only when the model calls it painted *and* its
calibrated P(painted) clears the checkpoint's ``completion_threshold``, which
``train.py`` picks on validation to keep false completes under a target
(``validation.MAX_FALSE_COMPLETE``). Uncertain wall counts as not done.
A checkpoint without a threshold (older checkpoints, or one where no threshold
met the target) never reports ``complete``: the status is ``not_calibrated`` and
a person signs the wall off.
Everything left over becomes touch-up regions, largest first, and a binary mask
of confidently painted pixels that ``robot_planner.py --observe`` can merge into
its progress grid. numpy + scipy only.
"""
import numpy as np
from scipy import ndimage
from validation import COMPLETE
MAX_TOUCH_UP_REGIONS = 20
def completion_summary(wall, painted, painted_probability, threshold, uncertain=None, min_region_px=64):
"""Return ``(summary, confident_painted_mask, touch_up_mask)`` for one frame.
``wall`` and ``painted`` are boolean masks (painted is a subset of wall);
``painted_probability`` is the calibrated P(painted) per pixel.
"""
wall = np.asarray(wall, bool)
painted = np.asarray(painted, bool) & wall
uncertain = np.zeros_like(wall) if uncertain is None else np.asarray(uncertain, bool) & wall
total = int(wall.sum())
summary = {"completion_threshold": threshold, "complete_at_fraction": COMPLETE, "min_region_px": min_region_px}
if not total:
empty = np.zeros_like(wall)
return {**summary, "status": "no_wall_detected", "confident_painted_fraction": None,
"touch_up_regions": [], "touch_up_regions_total": 0}, empty, empty
if threshold is None:
confident = np.zeros_like(wall) # nothing is trusted as done without a validated threshold
else:
confident = painted & (np.asarray(painted_probability) >= threshold)
remaining = wall & ~confident
fraction = float(confident.sum()) / total
if threshold is None:
status = "not_calibrated"
else:
status = "complete" if fraction >= COMPLETE else "incomplete"
labels, count = ndimage.label(remaining, structure=np.ones((3, 3), bool))
regions = []
if count:
sizes = np.bincount(labels.ravel(), minlength=count + 1)
slices = ndimage.find_objects(labels)
for index in np.argsort(-sizes[1:]) + 1:
if sizes[index] < min_region_px:
break
ys, xs = slices[index - 1]
region = labels[ys, xs] == index
cy, cx = ndimage.center_of_mass(region)
regions.append({
"bbox_xyxy": [xs.start, ys.start, xs.stop - 1, ys.stop - 1],
"center_xy": [round(float(cx) + xs.start, 1), round(float(cy) + ys.start, 1)],
"pixels": int(sizes[index]),
# Why it is not done: never painted, model unsure, or painted but below the threshold.
"unpainted_px": int((region & wall[ys, xs] & ~painted[ys, xs] & ~uncertain[ys, xs]).sum()),
"uncertain_px": int((region & uncertain[ys, xs]).sum()),
"low_confidence_painted_px": int((region & painted[ys, xs]).sum()),
})
touch_up = np.zeros_like(wall)
for region in regions[:MAX_TOUCH_UP_REGIONS]:
x0, y0, x1, y1 = region["bbox_xyxy"]
touch_up[y0:y1 + 1, x0:x1 + 1] |= remaining[y0:y1 + 1, x0:x1 + 1]
summary.update({
"status": status,
"confident_painted_fraction": round(fraction, 5),
"argmax_painted_fraction": round(float(painted.sum()) / total, 5),
"remaining_fraction": round(float(remaining.sum()) / total, 5),
"touch_up_regions": regions[:MAX_TOUCH_UP_REGIONS],
"touch_up_regions_total": len(regions),
})
return summary, confident, touch_up