Error in accumulate: math_matches error: Assertion error in line 157 about faster_coco_eval

[MiXaiLL76]

Thanks for the issue!
This error is caused by the fact that I incorrectly specified the extra_calc operation criteria.

I'll fix it within an hour and post a new version!

[MiXaiLL76]

FIX ff6e336

without extra_calc

from faster_coco_eval import COCO, COCOeval_faster
coco_gt = COCO('gt.json')
coco_dt = coco_gt.loadRes('preds.json')

cocoEval = COCOeval_faster(coco_gt, coco_dt, iouType='bbox')
cocoEval.evaluate()
cocoEval.accumulate()
cocoEval.summarize()

print(cocoEval.stats_as_dict)

{'AP_all': 0.844909645287565, 'AP_50': 0.9574501792007959, 'AP_75': 0.9422624028756038, 'AP_small': 0.6037607333485913, 'AP_medium': 0.8790270967302074, 'AP_large': 0.9185868759796408, 'AR_1': 0.17954286408399003, 'AR_10': 0.738941830046937, 'AR_100': 0.9033280920110436, 'AR_small': 0.7773887684368035, 'AR_medium': 0.9255266295014362, 'AR_large': 0.9589629629629629, 'AR_50': 0.9933915046286044, 'AR_75': 0.9782595186036742}

with extra_calc

from faster_coco_eval import COCO, COCOeval_faster
coco_gt = COCO('gt.json')
coco_dt = coco_gt.loadRes('preds.json')

cocoEval = COCOeval_faster(coco_gt, coco_dt, iouType='bbox', extra_calc=True)
cocoEval.evaluate()
cocoEval.accumulate()
cocoEval.summarize()

print(cocoEval.stats_as_dict)

{'AP_all': 0.844909645287565, 'AP_50': 0.9574501792007959, 'AP_75': 0.9422624028756038, 'AP_small': 0.6037607333485913, 'AP_medium': 0.8790270967302074, 'AP_large': 0.9185868759796408, 'AR_1': 0.17954286408399003, 'AR_10': 0.738941830046937, 'AR_100': 0.9033280920110436, 'AR_small': 0.7773887684368035, 'AR_medium': 0.9255266295014362, 'AR_large': 0.9589629629629629, 'AR_50': 0.9933915046286044, 'AR_75': 0.9782595186036742, 'mIoU': 0.9211815696510985, 'mAUC_50': 0.99020986735309}

Please write after checking the results and I will close the issue

[MiXaiLL76]

The new version of the library is already on pypi

[patrontheo]

Thanks for the fix!
It also works on my side, and the computed metrics seem to match the ones from pycocotools :).