| """New evaluation data, independent of unavailable original RefSeg generator. |
| Controlled RGB generation and official gRefCOCO subsampling with original masks. |
| The inference runner receives RGB + query, never scene objects or target labels. |
| """ |
| from pathlib import Path |
| import json,random,hashlib,collections,concurrent.futures,sys |
| import numpy as np |
| from PIL import Image,ImageDraw,ImageFilter,ImageEnhance |
| from prepare_assets import download |
| ROOT=Path(__file__).resolve().parents[1];DATA=ROOT/'data' |
| COLORS={'red':(214,47,47),'blue':(38,104,210),'green':(43,156,83),'orange':(233,139,34),'purple':(152,73,183),'yellow':(235,204,44),'pink':(222,105,151),'cyan':(38,177,196)} |
| SHAPES=['circle','square','triangle'] |
| def raster(shape,x,y,r,size=320): |
| m=Image.new('L',(size,size));d=ImageDraw.Draw(m);x*=size;y*=size;r*=size |
| if shape=='circle':d.ellipse((x-r,y-r,x+r,y+r),fill=255) |
| elif shape=='square':d.rectangle((x-r,y-r,x+r,y+r),fill=255) |
| else:d.polygon([(x,y-r),(x-r,y+r),(x+r,y+r)],fill=255) |
| return m |
| def save_mask(path,masks): |
| out=np.zeros((320,320),bool) |
| for m in masks:out|=np.asarray(m)>0 |
| Image.fromarray(out.astype('uint8')*255).save(path) |
| def controlled(): |
| records=[];scenes=[];rng_names=list(COLORS) |
| for seed in [11,23,37,53,71]: |
| rng=random.Random(seed) |
| for idx in range(240): |
| split='fit' if idx<40 else 'cal' if idx<80 else 'test';sid=f's{seed}_{idx:04d}' |
| color,ac,other=rng.sample(rng_names,3);shape=rng.choice(SHAPES);ash=rng.choice([s for s in SHAPES if s!=shape]);y=rng.uniform(.38,.54) |
| ax=rng.choice([.07,.335,.665,.93]) |
| objects=[(color,shape,x+rng.uniform(-.012,.012),y+rng.uniform(-.015,.015),rng.uniform(.055,.075)) for x in [.17,.5,.83]] |
| objects += [(ac,ash,ax,.71,rng.uniform(.045,.055)),(other,shape,.21+rng.uniform(-.02,.02),.15,rng.uniform(.05,.065))] |
| positions=[(.5,.15),(.8,.15),(.2,.86),(.5,.86),(.8,.86)] |
| used={(color,shape),(ac,ash),(other,shape)} |
| for x,yy in positions[:rng.randint(2,5)]: |
| c,s=rng.choice([(c,s) for c in rng_names for s in SHAPES if (c,s) not in used]);used.add((c,s));objects.append((c,s,x+rng.uniform(-.025,.025),yy+rng.uniform(-.025,.025),rng.uniform(.045,.068))) |
| image=Image.new('RGB',(320,320),(246,246,243));masks=[] |
| for c,s,x,yy,r in objects: |
| mask=raster(s,x,yy,r);image.paste(COLORS[c],mask=mask);masks.append(mask) |
| folder=DATA/'controlled';(folder/'images').mkdir(parents=True,exist_ok=True);(folder/'masks').mkdir(exist_ok=True) |
| image_path=folder/'images'/f'{sid}.png';image.save(image_path) |
| anchor_path=folder/'masks'/f'{sid}_anchor.png';masks[3].save(anchor_path) |
| absent=rng.choice([(c,s) for c in rng_names for s in SHAPES if (c,s) not in used]);phrase=f'{color} {shape}';anchor=f'{ac} {ash}';otherphrase=f'{other} {shape}' |
| held=split=='test' and idx%2==1;verb=rng.choice(['highlight','select','find']) if held else rng.choice(['segment','show']) |
| plural=f'all {phrase}s';left=f'all {phrase}s to the left of the {anchor}';right=f'all {phrase}s to the right of the {anchor}' |
| if held:left=f'all {phrase}s on the left of the {anchor}';right=f'all {phrase}s on the right of the {anchor}' |
| pairs=[('attribute',[(f'{verb} {plural}',[0,1,2]),(f'{verb} all {otherphrase}s',[4])]),('relation',[(f'{verb} {left}',[i for i in range(3) if objects[i][2]<ax]),(f'{verb} {right}',[i for i in range(3) if objects[i][2]>ax])]),('quantifier',[(f'{verb} the leftmost {phrase}',[0]),(f'{verb} {plural}',[0,1,2])]),('action',[(f'{verb} {plural}',[0,1,2]),(f'do not {verb} {plural}',[])]),('absence',[(f'{verb} {plural}',[0,1,2]),(f'{verb} all {absent[0]} {absent[1]}s',[])]),('paraphrase',[(f'show all {phrase}s',[0,1,2]),(f'select every {phrase}',[0,1,2])])] |
| these=[] |
| for family,endpoints in pairs: |
| for ep,(query,ids) in enumerate(endpoints): |
| rid=f'{sid}_{family}_{ep}';gt=folder/'masks'/f'{rid}.png';save_mask(gt,[masks[i] for i in ids]) |
| r={'id':rid,'scene_id':sid,'domain':'controlled','split':split,'official_split':None,'mode':family,'pair_id':sid+'_'+family,'endpoint':ep,'query':query,'image_path':str(image_path.relative_to(ROOT)),'gt_path':str(gt.relative_to(ROOT)),'anchor_gt_path':str(anchor_path.relative_to(ROOT)),'target_count':len(ids),'seed':seed,'template':'held' if held else 'seen','corruption':'clean'};records.append(r);these.append(r) |
| scenes.append({'scene_id':sid,'split':split,'objects':[{'color':c,'shape':s,'x':x,'y':yy,'radius':r} for c,s,x,yy,r in objects],'rgb_sha256':hashlib.sha256(image.tobytes()).hexdigest()}) |
| if split=='test' and 80<=idx<120: |
| rs=np.random.RandomState(seed*1000+idx) |
| noisy=Image.fromarray(np.clip(np.asarray(image).astype(float)+rs.normal(0,18,(320,320,3)),0,255).astype('uint8')) |
| for name,im in [('blur',image.filter(ImageFilter.GaussianBlur(2))),('noise',noisy),('dim',ImageEnhance.Brightness(image).enhance(.6))]: |
| path=folder/'images'/f'{sid}_{name}.png';im.save(path) |
| for base in these: |
| r=dict(base);r.update(id=base['id']+'_'+name,image_path=str(path.relative_to(ROOT)),pair_id=base['pair_id']+'_'+name,corruption=name);records.append(r) |
| (DATA/'controlled_scenes.jsonl').write_text(''.join(json.dumps(r)+'\n' for r in scenes)) |
| return records |
| def natural(): |
| from pycocotools import mask as maskutils |
| refs=json.loads((ROOT/'assets/grefcoco/grefs(unc).json').read_text());inst=json.loads((ROOT/'assets/grefcoco/instances.json').read_text()) |
| anns={a['id']:a for a in inst['annotations']};images={i['id']:i for i in inst['images']};pools=collections.defaultdict(list) |
| for r in refs: |
| count=0 if r['no_target'] else len(r['ann_id']);group='zero' if count==0 else 'one' if count==1 else 'multi';pools[(r['split'],group)].append(r) |
| selected=[];used=set();rng=random.Random(20260906) |
| |
| for split,source,n in [('fit','train',200),('cal','train',300),('val','val',200),('testA','testA',200),('testB','testB',200)]: |
| for group in (['zero','multi'] if split=='val' else ['zero','one','multi']): |
| count=300 if split=='val' else n |
| pool=list(pools[(source,group)]);rng.shuffle(pool);take=[] |
| for r in pool: |
| if r['image_id'] in used:continue |
| used.add(r['image_id']);take.append((split,group,r)) |
| if len(take)==count:break |
| if len(take)!=count:raise RuntimeError(f'Not enough distinct images for {split}/{group}: {len(take)}') |
| selected.extend(take) |
| out=DATA/'natural';(out/'images').mkdir(parents=True,exist_ok=True);(out/'masks').mkdir(exist_ok=True) |
| failures={} |
| def fetch(item): |
| _,_,r=item;path=out/'images'/r['file_name'] |
| try:download('https://s3.amazonaws.com/images.cocodataset.org/train2014/'+r['file_name'],path);Image.open(path).verify();return r['image_id'],True |
| except Exception as e:return r['image_id'],str(e) |
| with concurrent.futures.ThreadPoolExecutor(max_workers=12) as ex: |
| for iid,result in ex.map(fetch,selected): |
| if result is not True:failures[iid]=result |
| |
| if failures: |
| (DATA/'natural_download_failures.json').write_text(json.dumps(failures,indent=2)) |
| raise RuntimeError(f'{len(failures)} image downloads failed; rerun to complete cached selection') |
| records=[];license_map={x['id']:x for x in inst['licenses']} |
| for split,group,r in selected: |
| if r['image_id'] in failures:continue |
| info=images[r['image_id']];h,w=info['height'],info['width'];gt=np.zeros((h,w),bool) |
| if not r['no_target']: |
| for aid in r['ann_id']: |
| seg=anns[aid]['segmentation'] |
| if isinstance(seg,list):rle=maskutils.merge(maskutils.frPyObjects(seg,h,w)) |
| elif isinstance(seg.get('counts'),list):rle=maskutils.frPyObjects(seg,h,w) |
| else:rle=seg |
| gt|=maskutils.decode(rle).astype(bool) |
| if not r['no_target'] and not gt.any():raise ValueError('Official positive mask is empty') |
| sentence=rng.choice(r['sentences']);rid='g'+str(r['ref_id'])+'_'+str(sentence['sent_id']);path=out/'masks'/f'{rid}.png';Image.fromarray(gt.astype('uint8')*255).save(path) |
| records.append({'id':rid,'scene_id':'coco_'+str(r['image_id']),'domain':'natural','split':split,'official_split':r['split'],'mode':group,'pair_id':None,'endpoint':None,'query':sentence['sent'],'image_path':str((out/'images'/r['file_name']).relative_to(ROOT)),'gt_path':str(path.relative_to(ROOT)),'target_count':0 if r['no_target'] else len(r['ann_id']),'ref_id':r['ref_id'],'sent_id':sentence['sent_id'],'ann_ids':r['ann_id'],'seed':20260906,'template':'natural','corruption':'clean','license':license_map[info['license']],'source_image_url':'https://s3.amazonaws.com/images.cocodataset.org/train2014/'+r['file_name'],'flickr_url':info.get('flickr_url')}) |
| (DATA/'natural_download_failures.json').write_text(json.dumps(failures,indent=2));return records |
| def main(): |
| DATA.mkdir(exist_ok=True);c=controlled();n=natural();records=c+n |
| |
| splits=collections.defaultdict(set) |
| for r in records:splits[r['scene_id']].add(r['split']) |
| assert all(len(s)==1 for s in splits.values()) |
| path=DATA/'manifest.jsonl';path.write_text(''.join(json.dumps(r,ensure_ascii=False)+'\n' for r in records)) |
| counts=collections.Counter((r['domain'],r['split'],r['corruption'],r['mode']) for r in records) |
| summary={'records':len(records),'controlled_records':len(c),'natural_records':len(n),'distinct_scene_ids':len(splits),'manifest_sha256':hashlib.sha256(path.read_bytes()).hexdigest(),'counts':[dict(domain=k[0],split=k[1],corruption=k[2],mode=k[3],n=v) for k,v in sorted(counts.items())],'not_original_dataset':True,'natural_sampling':'One randomly selected expression per distinct image; image-disjoint fit/cal/val/testA/testB; balanced target-count strata; failures recorded without outcome-dependent replacement'} |
| (DATA/'dataset_summary.json').write_text(json.dumps(summary,indent=2));print(json.dumps(summary,indent=2)) |
| if __name__=='__main__':main() |
|
|
