SceneBench / code /generate.py
EgoF0102's picture
SceneBench v1: frozen source-derived test benchmark and verified artifacts
7735593 verified
Raw History Blame Contribute Delete
23.8 kB
import argparse, random, itertools, collections, json, math
from pathlib import Path
from common import *
CVS_OPTIONS=[
['0 — Two separate connecting structures cannot be established.','1 — Two structures are partly distinguishable, but overlap or limited clarity prevents a clear assessment.','2 — Two connecting structures are clearly distinguishable.'],
['0 — The cystic plate is not visible or requires further dissection.','1 — It is visible, but exposure, overlap, or viewing angle limits the assessment.','2 — It is clearly exposed to approximately the lower third of the gallbladder.'],
['0 — Tissue or technical limitations prevent assessment of the triangle and its structures.','1 — The triangle is only partly clear or its view remains limited.','2 — The triangle is fully cleared and its structures are clearly visible.'],['No','Yes']]
CVS_QUESTIONS=['How clearly are exactly two structures seen connecting to the gallbladder?','How clearly is the cystic plate exposed at the lower third of the gallbladder?','How clearly has the hepatocystic triangle been cleared to reveal its structures?','Is the critical view of safety (CVS) achieved in this image?']
class Generator:
def __init__(self,s,config):
self.s=s;self.cfg=config;self.rng=random.Random(config['seed']);self.pool=collections.defaultdict(list);self.seen=set();self.reject=collections.Counter()
self.all_targets=sorted({t[2] for d in s.ivt.values() for ts in d.values() for t in ts if t[2]!='null_target'})
self.inst_verbs={i:sorted({t[1] for d in s.ivt.values() for ts in d.values() for t in ts if t[0]==i and t[1]!='null_verb'}) for i in INSTRUMENTS}
def opts(self,answer,universe):
neg=[x for x in universe if x!=answer]
return [answer]+self.rng.sample(neg,3) if len(neg)>=3 else None
def add(self,task,v,frames,q,choices,answer,query,answer_class=None,episode=None,marker=None,tags=None):
if not choices or answer not in choices or len(set(choices))!=len(choices):return
if isinstance(frames,int):frames=[frames]
ident=digest([task,v,frames,query])[:20]
if ident in self.seen:return
self.seen.add(ident)
self.pool[task].append(dict(candidate_id=ident,task=task,family=family(task),video=v,frames=frames,question=q,choices=choices,answer=answer,answer_class=str(answer_class if answer_class is not None else answer),query=query,episode=episode or f'{v}:{task}:{frames[0]//30}',marker=marker,tags=tags or {}))
def run_episode(self,v,f,ts):return '|'.join(f'{v}:{"/".join(t)}:{self.s.action_starts.get((v,f,t),f)}' for t in sorted(ts))
def frame(self,v,f):
s=self.s;g=s.graphs.get((v,f));phase=int(s.phase[v][f]['Phase_gt']);raw=s.ivt[v].get(f,set());ts={t for t in raw if t[1]!='null_verb' and t[2]!='null_target'}
# Current-phase task covers complete phase inventory; exclude phase boundaries by 3 s.
if phase and f>=3 and f+3<len(s.phase[v]) and all(int(s.phase[v][k]['Phase_gt'])==phase for k in range(f-3,f+4)):
ans=PHASES[phase];self.add('PR01',v,f,'Which surgical phase is shown in this image?',self.opts(ans,[PHASES[i] for i in range(1,7)]),ans,{'op':'phase','phase':phase},phase)
if not g:return
objs=s.clean_objects(v,f);u=s.unique(v,f);present=s.tools(v,f)
if present is None:return
# Require graph detections and expert category presence to agree for tool queries.
toolids=[i for i,o in objs.items() if o['type']=='instrument' and o['component'] in INSTRUMENTS and o['component'] in present]
rawtoolids=[i for i,o in enumerate(g['objects']) if o['type']=='instrument' and o['component'] in INSTRUMENTS]
complete_tools=set(rawtoolids)==set(toolids) and {g['objects'][i]['component'] for i in toolids}==(present&set(INSTRUMENTS))
for i in toolids:
o=objs[i];a=o['component'];ans=title(a)
self.add('P01',v,f,'What type of instrument is indicated by box A?',self.opts(ans,[title(t) for t in INSTRUMENTS]),ans,{'op':'object_identity','id':i,'component':a},a,marker={'kind':'box','object_id':i,'bbox':o['bbox'],'label':'A'})
for a in ANATOMIES:
if a not in u:continue
i=u[a];o=objs[i];b=o['bbox'];point=o['center']
# Marker placement remains a visual-review gate, not automatic anatomy truth.
ans=title(a)
self.add('P02',v,f,'Which anatomical structure is indicated by arrow A?',self.opts(ans,[title(t) for t in ANATOMIES]),ans,{'op':'object_identity','id':i,'component':a},a,marker={'kind':'arrow','object_id':i,'bbox':b,'point':point,'label':'A','review_required':True})
# Tool absence is defined by original Tool_gt, not missing SG detections.
for a in INSTRUMENTS:
val=a in present
if val and a not in {objs[i]['component'] for i in toolids}:continue
if not val and any(o['component']==a for o in g['objects']):continue
self.add('P03',v,f,f'Is a {display(a)} visible in this image?',['Yes','No'],'Yes' if val else 'No',{'op':'tool_presence','tool':a},'Yes' if val else 'No',tags={'tool':a})
if complete_tools:
for a in INSTRUMENTS:
ids=[i for i in toolids if objs[i]['component']==a];n=len(ids)
if n>3:continue
# Duplicate overlapping detections are not counted as separate tools.
bad=False
for i,j in itertools.combinations(ids,2):
b=objs[i]['bbox'];c=objs[j]['bbox'];inter=max(0,min(b[2],c[2])-max(b[0],c[0]))*max(0,min(b[3],c[3])-max(b[1],c[1]));union=(b[2]-b[0])*(b[3]-b[1])+(c[2]-c[0])*(c[3]-c[1])-inter
if union and inter/union>.5:bad=True
if not bad:self.add('P04',v,f,f'How many separate {display(a)} instruments are visible?',['0','1','2','3'],str(n),{'op':'count_tools','tool':a,'ids':ids},n,tags={'tool':a})
for a in INSTRUMENTS:
if a not in u or u[a] not in toolids:continue
ans=s.quadrant(v,f,u[a])
if ans:self.add('P05',v,f,f'In which image quadrant is the center of the {display(a)} bounding box?',QUADS,ans,{'op':'absolute_quadrant','id':u[a]},ans,tags={'object':a})
# IVT three-way querying, with full matching answer-set uniqueness.
for inst,verb,target in sorted(ts):
if inst not in present or inst not in u or u[inst] not in toolids:continue
ep=self.run_episode(v,f,[(inst,verb,target)])
aa={t[1] for t in ts if t[0]==inst and t[2]==target}
if len(aa)==1:
ans=GERUND[verb];self.add('R01',v,f,f'What action is the {display(inst)} performing on the {display(target)}?',self.opts(ans,[GERUND[t] for t in VERBS]),ans,{'op':'ivt_verb','instrument':inst,'target':target},verb,ep,tags={'instrument':inst,'verb':verb,'target':target})
aa={t[0] for t in ts if t[1]==verb and t[2]==target}
if len(aa)==1:
ans=title(inst);self.add('R02',v,f,f'Which instrument is {GERUND[verb].lower()} the {display(target)}?',self.opts(ans,[title(t) for t in INSTRUMENTS]),ans,{'op':'ivt_instrument','verb':verb,'target':target},inst,ep,tags={'verb':verb,'target':target})
aa={t[2] for t in ts if t[0]==inst and t[1]==verb}
if len(aa)==1:
ans=title(target);self.add('R03',v,f,f'Which target is the {display(inst)} {GERUND[verb].lower()}?',self.opts(ans,[title(t) for t in self.all_targets]),ans,{'op':'ivt_target','instrument':inst,'verb':verb},target,ep,tags={'instrument':inst,'verb':verb})
# Quadrants are mutually exclusive; every axis is supported by both directed arrays.
pairs=list(itertools.permutations([a for a in u if a in ANATOMIES or a in INSTRUMENTS],2));self.rng.shuffle(pairs)
emitted=collections.Counter()
for a,b in pairs:
ia,ib=u[a],u[b];A=a in INSTRUMENTS;B=b in INSTRUMENTS
if A and ia not in toolids or B and ib not in toolids:continue
task='R06' if A and B else 'R05' if A and not B else 'R04' if not A and not B else None
if not task or emitted[task]>=2:continue
ans=s.quadrant(v,f,ia,ib)
if ans:
self.add(task,v,f,f'Where is the center of the {display(a)} bounding box relative to the center of the {display(b)} bounding box?',QUADS,ans,{'op':'relative_quadrant','subject':ia,'reference':ib},ans,tags={'subject':a,'reference':b});emitted[task]+=1
if phase==0:return
facts=[t for t in sorted(ts) if t[0] in u and t[2] in u and u[t[0]] in toolids and s.action_edge(v,f,*t)]
unique_facts=[t for t in facts if len({x[2] for x in ts if x[:2]==t[:2]})==1]
# Two independent actions -> ordered target pair, including partial-match distractors.
for x,y in itertools.combinations(unique_facts,2):
if x[0]==y[0]:continue
a,b=x[2],y[2];a2=self.rng.choice([t for t in ANATOMIES if t!=a]);b2=self.rng.choice([t for t in ANATOMIES if t!=b]);fmt=lambda p,q:f'({title(p)}, {title(q)})'
ans=fmt(a,b);opts=[ans,fmt(a2,b),fmt(a,b2),fmt(a2,b2)]
self.add('C01',v,f,f'Which ordered pair gives the structure {PARTICIPLE[x[1]]} by the {display(x[0])} first and the structure {PARTICIPLE[y[1]]} by the {display(y[0])} second?',opts,ans,{'op':'ordered_action_targets','queries':[list(x[:2]),list(y[:2])]},f'{a}|{b}',self.run_episode(v,f,[x,y]),tags={'ivts':[x,y]});break
refs=[a for a in u if a in ANATOMIES];self.rng.shuffle(refs)
# Spatially identify one tool, then follow its unique target.
for ref in refs:
for p in self.rng.sample(list(INV),4):
if len(toolids)<2 or any(not s.consistent(v,f,i,u[ref],p) for i in toolids):continue
ids=[i for i in toolids if s.spatial(v,f,i,u[ref],p)]
if len(ids)!=1:continue
inst=objs[ids[0]]['component'];targets={t[2] for t in facts if t[0]==inst}
if len(targets)!=1 or targets!={t[2] for t in ts if t[0]==inst}:continue
target=next(iter(targets));ans=title(target)
self.add('C02',v,f,f'Which structure is being acted on by the instrument whose bounding-box center is {self.phrase(p)} the {display(ref)} center?',self.opts(ans,[title(t) for t in ANATOMIES]),ans,{'op':'spatial_tool_target','reference':u[ref],'predicate':p},target,tags={'instrument':inst,'reference':ref});break
else:continue
break
for inst,verb,target in unique_facts:
for ref in refs:
if ref==target:continue
ans=s.quadrant(v,f,u[target],u[ref])
if ans:
self.add('C03',v,f,f'Where is the bounding-box center of the structure the {display(inst)} is {GERUND[verb].lower()} relative to the {display(ref)} center?',QUADS,ans,{'op':'action_target_quadrant','instrument':inst,'verb':verb,'reference':u[ref]},ans,self.run_episode(v,f,[(inst,verb,target)]));break
# Two spatial predicates intersect on exactly one known anatomy; each predicate has alternatives.
anatomies=[a for a in ANATOMIES if a in u]
refnames=[a for a in u if a in ANATOMIES or a in INSTRUMENTS and u[a] in toolids]
possibilities=list(itertools.combinations(refnames,2));self.rng.shuffle(possibilities);done=False
if len(anatomies)>=4:
for r1,r2 in possibilities[:12]:
for p,q in self.rng.sample(list(itertools.product(INV,repeat=2)),16):
universe=[a for a in anatomies if s.consistent(v,f,u[a],u[r1],p) and s.consistent(v,f,u[a],u[r2],q)]
A={a for a in universe if s.spatial(v,f,u[a],u[r1],p)};B={a for a in universe if s.spatial(v,f,u[a],u[r2],q)};I=A&B
if len(I)!=1 or len(A)<2 or len(B)<2 or len(universe)<4:continue
target=next(iter(I));d1=self.rng.choice(sorted(A-I));d2=self.rng.choice(sorted(B-I));rest=[a for a in universe if a not in [target,d1,d2]]
if not rest:continue
opts=[title(a) for a in [target,d1,d2,self.rng.choice(rest)]]
self.add('C04',v,f,f'Which structure has its bounding-box center both {self.phrase(p)} the {display(r1)} center and {self.phrase(q)} the {display(r2)} center?',opts,title(target),{'op':'spatial_intersection','references':[u[r1],u[r2]],'predicates':[p,q],'candidate_ids':[u[a] for a in universe]},target);done=True;break
if done:break
# Raw IVT categories support multi-instance class queries, without invented instance identity.
for inst,verb in sorted({t[:2] for t in ts}):
targets={t[2] for t in ts if t[:2]==(inst,verb)}
if len(targets)<2 or not targets<=set(anatomies):continue
for ref,p in itertools.product(refs,INV):
universe=[a for a in anatomies if s.consistent(v,f,u[a],u[ref],p)];B={a for a in universe if s.spatial(v,f,u[a],u[ref],p)};I=targets&B
if len(I)!=1 or len(B)<2 or len(universe)<4:continue
target=next(iter(I));neg=list(set(universe)-I);ans=title(target)
opts=[ans]+[title(a) for a in self.rng.sample(neg,3)]
self.add('C05',v,f,f'Which structure is both {PARTICIPLE[verb]} by a {display(inst)} and has its bounding-box center {self.phrase(p)} the {display(ref)} center?',opts,ans,{'op':'action_spatial_intersection','instrument':inst,'verb':verb,'reference':u[ref],'predicate':p,'candidate_ids':[u[a] for a in universe]},target,self.run_episode(v,f,[t for t in ts if t[:2]==(inst,verb)]));break
else:continue
break
# Relation-filtered counts use full instrument inventory. Positive and zero intersections allowed.
if complete_tools and 1<=len(toolids)<=3:
attempts=list(itertools.product(refs,INV,refs,INV));self.rng.shuffle(attempts)
emitted_counts=set()
for r1,p,r2,q in attempts[:90]:
if r1==r2 and p==q or p==INV[q] and r1==r2:continue
if any(not s.consistent(v,f,i,u[r1],p) or not s.consistent(v,f,i,u[r2],q) for i in toolids):continue
A={i for i in toolids if s.spatial(v,f,i,u[r1],p)};B={i for i in toolids if s.spatial(v,f,i,u[r2],q)};n=len(A&B)
if not A or not B or n in emitted_counts:continue
self.add('C06',v,f,f'How many instruments have bounding-box centers both {self.phrase(p)} the {display(r1)} center and {self.phrase(q)} the {display(r2)} center?',['0','1','2','3'],str(n),{'op':'spatial_count','references':[u[r1],u[r2]],'predicates':[p,q],'tool_ids':toolids},n,tags={'sets':[sorted(A),sorted(B)]});emitted_counts.add(n)
if len(emitted_counts)>=2:break
# Spatial target selection -> observed instrument, no planning/next-action claims.
if len({t[0] for t in facts})>=2 and {t[2] for t in ts} <= set(u):
targets={t[2] for t in facts}
for ref,p in itertools.product(refs,INV):
if any(not s.consistent(v,f,u[a],u[ref],p) for a in targets):continue
allowed={a for a in targets if s.spatial(v,f,u[a],u[ref],p)};insts={t[0] for t in facts if t[2] in allowed}
if len(insts)!=1 or not allowed or allowed==targets:continue
inst=next(iter(insts));ans=title(inst)
self.add('C07',v,f,f'Which instrument is acting on a structure whose bounding-box center is {self.phrase(p)} the {display(ref)} center?',self.opts(ans,[title(t) for t in INSTRUMENTS]),ans,{'op':'spatial_target_instrument','reference':u[ref],'predicate':p},inst);break
# CVS: every conflict-free labeled second in domain is eligible, not one-per-interval.
@staticmethod
def phrase(p):return {'left':'to the left of','right':'to the right of','above':'above','below':'below'}[p]
def cvs(self):
for v,good in self.s.cvt.items():
for f,r in good.items():
for k in range(4):
seg=r[f'episode_{k}']
if f<seg['start']+2 or f>seg['end']-2:continue
score=r['scores'][k] if k<3 else r['cv'];answer=CVS_OPTIONS[k][score]
self.add(f'V0{k+1}',v,f,CVS_QUESTIONS[k],CVS_OPTIONS[k],answer,{'op':'cvs','criterion':k,'scores':r['scores'],'rows':r['rows'],'label_source':r['source'],'segment':seg},str(score) if k<3 else answer,f'{v}:V{k+1}:{seg["start"]}:{seg["end"]}',tags={'calot_progress':self.calot_progress(v,f)})
def calot_progress(self,v,f):
x=self.s.phase[v];a=next(i for i,r in enumerate(x) if r['Phase_gt']==1);b=next(i for i,r in enumerate(x) if r['Phase_gt']==2)
return (f-a)/(b-a)
def dynamic(self):
s=self.s
for v in s.videos:
n=len(s.phase[v]);dtchoices=[3,5,8,10]
for f in range(3,n-13,2):
for dt in self.rng.sample(dtchoices,2):
h=f+dt
if (v,f) not in s.graphs or (v,h) not in s.graphs:continue
pres=[s.tools(v,t) for t in [f,h]]
if any(t is None for t in pres):continue
uu=[s.unique(v,t) for t in [f,h]];objs=[s.clean_objects(v,t) for t in [f,h]]
ts=[{x for x in s.ivt[v].get(t,set()) if x[1]!='null_verb' and x[2]!='null_target'} for t in [f,h]]
# Stable presence states, category-level observation of both endpoints.
tool=self.rng.choice(INSTRUMENTS);states=[tool in p for p in pres]
stable=all(s.tools(v,k) is not None and (tool in s.tools(v,k))==state for t,state in zip([f,h],states) for k in range(t-1,t+2))
if stable and all(not state or tool in u for state,u in zip(states,uu)):
opts=[f'{a} → {b}' for a,b in itertools.product(['Absent','Present'],repeat=2)];ans=' → '.join('Present' if b else 'Absent' for b in states)
self.add('D01',v,[f,h],f'Which pair describes {display(tool)} visibility in frame 1 and frame 2, in that order?',opts,ans,{'op':'temporal_presence','tool':tool},ans,tags={'tool':tool,'changed':states[0]!=states[1],'dt':dt})
common=[i for i in INSTRUMENTS if all(i in u and i in p for u,p in zip(uu,pres))]
for inst in common:
targets={t[2] for t in ts[0] if t[0]==inst}&{t[2] for t in ts[1] if t[0]==inst}
for target in sorted(targets):
verbs=[{t[1] for t in a if t[0]==inst and t[2]==target} for a in ts]
if any(len(x)!=1 for x in verbs):continue
aa=[next(iter(x)) for x in verbs]
if not all({x[1] for x in s.ivt[v].get(k,set()) if x[0]==inst and x[2]==target}=={a} for t,a in zip([f,h],aa) for k in range(t-1,t+2)):continue
av=list(dict.fromkeys(aa))
if len(av)==1:
other=[a for a in self.inst_verbs[inst] if a not in av]
if not other:continue
av.append(self.rng.choice(other))
fmt=lambda a,b:f'{GERUND[a]} → {GERUND[b]}'
opts=[fmt(a,b) for a,b in itertools.product(av,repeat=2)];ans=fmt(*aa)
self.add('D02',v,[f,h],f'What is the {display(inst)} doing to the {display(target)} in frame 1 and frame 2, respectively?',opts,ans,{'op':'temporal_verb','instrument':inst,'target':target},ans,tags={'instrument':inst,'target':target,'changed':aa[0]!=aa[1],'dt':dt});break
verbs={t[1] for t in ts[0] if t[0]==inst}&{t[1] for t in ts[1] if t[0]==inst}
for verb in sorted(verbs):
tt=[{t[2] for t in a if t[:2]==(inst,verb)} for a in ts]
if any(len(x)!=1 for x in tt):continue
aa=[next(iter(x)) for x in tt]
if not all({x[2] for x in s.ivt[v].get(k,set()) if x[:2]==(inst,verb)}=={a} for t,a in zip([f,h],aa) for k in range(t-1,t+2)):continue
av=list(dict.fromkeys(aa))
if len(av)==1:av.append(self.rng.choice([a for a in self.all_targets if a not in av]))
fmt=lambda a,b:f'{title(a)} → {title(b)}';opts=[fmt(a,b) for a,b in itertools.product(av,repeat=2)];ans=fmt(*aa)
self.add('D03',v,[f,h],f'Which target is the {display(inst)} {GERUND[verb].lower()} in frame 1 and frame 2, respectively?',opts,ans,{'op':'temporal_target','instrument':inst,'verb':verb},ans,tags={'instrument':inst,'verb':verb,'changed':aa[0]!=aa[1],'dt':dt});break
phases=[int(s.phase[v][t]['Phase_gt']) for t in [f,h]]
if min(phases)>0 and all(int(s.phase[v][k]['Phase_gt'])==p for t,p in zip([f,h],phases) for k in range(t-1,t+2)):
av=list(dict.fromkeys(phases))
if len(av)==1:av.append(self.rng.choice([p for p in range(1,7) if p not in av]))
fmt=lambda a,b:f'{PHASES[a]} → {PHASES[b]}';opts=[fmt(a,b) for a,b in itertools.product(av,repeat=2)];ans=fmt(*phases)
self.add('D04',v,[f,h],'Which surgical phase is shown in frame 1 and frame 2, respectively?',opts,ans,{'op':'temporal_phase'},ans,tags={'changed':phases[0]!=phases[1],'dt':dt})
for inst in common:
refs=[a for a in ANATOMIES if all(a in u for u in uu)];self.rng.shuffle(refs)
for ref in refs[:2]:
for dirs in [('left','right'),('above','below')]:
vals=[]
for t,u in zip([f,h],uu):
vv=[p for p in dirs if s.spatial(v,t,u[inst],u[ref],p)]
if len(vv)!=1:break
vals.append(vv[0])
if len(vals)!=2:continue
fmt=lambda a,b:f'{a.capitalize()} → {b.capitalize()}';opts=[fmt(a,b) for a,b in itertools.product(dirs,repeat=2)];ans=fmt(*vals)
self.add('D05',v,[f,h],f'Relative to the {display(ref)} bounding-box center, is the {display(inst)} center {dirs[0]} or {dirs[1]} in frame 1 and frame 2, respectively?',opts,ans,{'op':'temporal_spatial','instrument':inst,'reference':ref,'axis':list(dirs)},ans,tags={'changed':vals[0]!=vals[1],'dt':dt})
break
else:continue
break
def main():
p=argparse.ArgumentParser();p.add_argument('--ori',default='/home/ach18533cl/workspace/ori_data');p.add_argument('--root',default='/home/ach18533cl/workspace/SceneBench');a=p.parse_args();root=Path(a.root);cfg=json.loads((root/'config.json').read_text());s=Sources(a.ori,cfg['test_videos']);gen=Generator(s,cfg)
for v in s.videos:
for f in range(0,len(s.phase[v]),2):gen.frame(v,f)
print('Generated single frames',v,{k:len(x) for k,x in gen.pool.items()},flush=True)
gen.cvs();gen.dynamic()
for k,rows in gen.pool.items():write_jsonl(root/f'work/candidates/{k}.jsonl',rows)
write_json(root/'work/pool_summary.json',dict(counts={k:len(x) for k,x in gen.pool.items()},answers={k:dict(collections.Counter(x['answer_class'] for x in rr)) for k,rr in gen.pool.items()},source_rejections=dict(s.counts)))
write_json(root/'work/source_cvs_rows.json',s.cvs_rows)
print('DONE', {k:len(x) for k,x in gen.pool.items()},flush=True)
if __name__=='__main__':main()