VLMdataset / realcase /106 /move_replace_mesh_v1 /code /prepare_selected.py
FrontRear's picture
Add reviewed experimental move+replace conditions 106 move_replace_mesh_v1
44a519e verified
Raw History Blame Contribute Delete
3.58 kB
import json,cv2,numpy as np,hashlib,os
from pathlib import Path
ROOT=Path('/root/autodl-tmp/kimi_realcase_20261109')
INV=json.loads((ROOT/'inventory_sources.json').read_text())
CONFIG=json.loads((ROOT/os.environ.get('BATCH_CONFIG','first_batch_config.json')).read_text())
W,H=1280,704
def cover(fr,nearest=False):
h,w=fr.shape[:2];scale=max(W/w,H/h);nw,nh=round(w*scale),round(h*scale)
interp=cv2.INTER_NEAREST if nearest else cv2.INTER_AREA
fr=cv2.resize(fr,(nw,nh),interpolation=interp);x=(nw-W)//2;y=(nh-H)//2
return fr[y:y+H,x:x+W],{'source_wh':[w,h],'scale':scale,'resized_wh':[nw,nh],'crop_xy':[x,y],'output_wh':[W,H]}
def mask_score(r):
best=None
for m in r['mask_videos']:
vals=[s['nonzero_px'] or 0 for s in m['samples']];score=float(np.median(vals))
if best is None or score>best[0]:best=(score,m['file'])
if not best:raise RuntimeError('no dataset mask prompt candidate')
return best[1]
manifest=[]
for cfg in CONFIG:
r=INV['records'][cfg['inventory_index']]
assert r['source_sha256'] and r['probe']['frames_metadata']>=49
case=cfg['case_id'];cd=ROOT/'cases'/case;sd=ROOT/'samples'/case
if cd.exists():
existing=json.loads((cd/'case.json').read_text());manifest.append(existing)
print(json.dumps({'case':case,'status':'already_preserved'}),flush=True)
continue
for sub in ['frames','segmentation/lamp0']:(cd/sub).mkdir(parents=True,exist_ok=False)
sd.mkdir(parents=True)
src=r['source_path'];mask_file=cfg.get('mask_prompt_video') or mask_score(r);mask_src=mask_file if Path(mask_file).is_absolute() else str(Path(r['raw_dir'])/mask_file)
static_mask=mask_src.lower().endswith(('.png','.jpg','.jpeg'))
cap=cv2.VideoCapture(src);mcap=None if static_mask else cv2.VideoCapture(mask_src)
fps=float(cap.get(5));n=int(cap.get(7));mfps=10.0 if static_mask else float(mcap.get(5));mn=1 if static_mask else int(mcap.get(7))
frame_map=[];transform=None
for i in range(49):
si=min(int(round(i*fps/10.0)),n-1);mi=min(int(round(i*mfps/10.0)),mn-1)
cap.set(cv2.CAP_PROP_POS_FRAMES,si);ok,fr=cap.read()
if static_mask:
mfr=cv2.imread(mask_src);mok=mfr is not None
else:
mcap.set(cv2.CAP_PROP_POS_FRAMES,mi);mok,mfr=mcap.read()
if not ok or not mok:raise RuntimeError(f'decode failed {case} frame {i} source {si} mask {mi}')
fr,transform=cover(fr);mfr,_=cover(mfr,nearest=True)
mask=(mfr.mean(axis=2)>127).astype(np.uint8)*255
cv2.imwrite(str(cd/'frames'/f'{i:03}.png'),fr)
cv2.imwrite(str(cd/'segmentation/lamp0'/f'{i:03}.png'),mask)
frame_map.append({'output_frame':i,'source_frame':si,'mask_source_frame':mi,'source_time_seconds':si/fps,'mask_time_seconds':mi/mfps})
cap.release()
if not static_mask:mcap.release()
cfg2=dict(cfg);cfg2.update({'source_path':src,'source_sha256':r['source_sha256'],'source_size':r['source_size'],'raw_dir':r['raw_dir'],'mask_prompt_video':mask_src,'mask_prompt_semantics':'dataset-derived candidate prompt only; completeness and identity require visual review','source_probe':r['probe'],'frame_map':frame_map,'cover_transform':transform,'status':'prepared'})
(cd/'case.json').write_text(json.dumps(cfg2,ensure_ascii=False,indent=2))
manifest.append(cfg2)
print(json.dumps({'case':case,'frames':49,'mask_prompt':mask_file,'source_sha256':r['source_sha256']}),flush=True)
(ROOT/os.environ.get('BATCH_OUTPUT','first_batch_prepared.json')).write_text(json.dumps(manifest,ensure_ascii=False,indent=2))