File size: 3,281 Bytes
0ea7dca | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 | """Exact PNG-derived BOP visible/valid metrics, preserving raw writer metadata.
Only derived annotation fields change; RGB, depth, masks, poses and camera K do
not. Full silhouette area/box remain the writer's padded-image measurement,
except the area is bounded below by the saved, clipped amodal-mask area.
"""
import json
from pathlib import Path
import cv2
import numpy as np
from scripts.audit_handal40_render import sha
def canonicalize(root):
scene=root/'bop_data/handal/train_pbr/000000';path=scene/'scene_gt_info.json'
if (root/'visible_count_canonicalization.json').exists():
receipt=json.loads((root/'visible_count_canonicalization.json').read_text())
if sha(path)!=receipt['canonical_info_sha256']:raise ValueError('previously canonicalized metadata changed')
return receipt
raw=path.read_bytes();info=json.loads(raw);changes=[]
for frame,rows in info.items():
depth=cv2.imread(str(scene/'depth'/f'{int(frame):06d}.png'),cv2.IMREAD_UNCHANGED)
if depth is None:raise ValueError('missing depth')
for i,row in enumerate(rows):
mask=cv2.imread(str(scene/'mask'/f'{int(frame):06d}_{i:06d}.png'),0)
vis=cv2.imread(str(scene/'mask_visib'/f'{int(frame):06d}_{i:06d}.png'),0)
if mask is None or vis is None or mask.shape!=depth.shape or vis.shape!=depth.shape:raise ValueError('mask/depth dimensions')
if not set(np.unique(mask))<={0,255} or not set(np.unique(vis))<={0,255}:raise ValueError('non-binary mask')
m=mask>0;v=vis>0
if np.any(v&~m):raise ValueError('visible pixels outside amodal silhouette')
visible=int(v.sum());valid=int((m&(depth>0)).sum());all_pixels=max(row['px_count_all'],int(m.sum()))
if not 0<=visible<=valid<=all_pixels:raise ValueError('invalid BOP pixel count ordering')
differences=[abs(visible-row['px_count_visib']),abs(valid-row['px_count_valid']),abs(all_pixels-row['px_count_all'])]
if max(differences)>8:raise ValueError('raster-derived count adjustment exceeds8 pixels; investigate')
x,y,w,h=cv2.boundingRect(vis)
bbox=[x,y,w-1,h-1] if visible else [-1,-1,-1,-1]
after={**row,'px_count_all':all_pixels,'px_count_valid':valid,'px_count_visib':visible,
'visib_fract':visible/all_pixels if all_pixels else 0.,'bbox_visib':bbox}
if after!=row:changes.append({'frame':int(frame),'instance':i,'before':dict(row),'after':after})
rows[i]=after
archive=root/'writer_raw';archive.mkdir(exist_ok=True)
original=archive/'scene_gt_info.json'
if original.exists() and original.read_bytes()!=raw:raise ValueError('raw writer metadata already differs')
original.write_bytes(raw)
path.write_text(json.dumps(info,indent=2)+'\n')
result={'method':'saved PNG visible/valid counts and visible bbox; full amodal lower bound',
'max_allowed_pixel_count_correction':8,'raw_info_sha256':sha(original),
'canonical_info_sha256':sha(path),'implementation_sha256':sha(Path(__file__)),
'changed_instances':len(changes),'changes':changes}
(root/'visible_count_canonicalization.json').write_text(json.dumps(result,indent=2)+'\n')
return result
|