microsoft / RegionCLIP

[CVPR 2022] Official code for "RegionCLIP: Region-based Language-Image Pretraining"
Apache License 2.0
696 stars 52 forks source link

AP on ground truth region #59

Closed erjui closed 1 year ago

erjui commented 1 year ago

Hello, thank you for the nice work first of all! Now I am trying to retrieve the result from table 4 using ground truth boxes. But the thing is that I thought precisions should be always the same if I use ground truth boxes, but the result seems to be different over different thresholds. So could you give a little explanation about this?!

image
YiwuZhong commented 1 year ago

The results seem to be quite close across different thresholds (61.0 vs. 61.4 vs. 61.0). One possibility is that the difference is caused by the precision during floating-point computation.