search

vacancy / PreciseRoIPooling

Precise RoI Pooling with coordinate gradient support, proposed in the paper "Acquisition of Localization Confidence for Accurate Object Detection" (https://arxiv.org/abs/1807.11590).
MIT License
770 stars 152 forks source link

IoU threshold in IoU Net #23

Open Muran337287 opened 5 years ago

Muran337287 commented 5 years ago

Hi, in your paper, predicted boxes with IoU larger than 0.5 will be fed into IoUNet. I have a question that did you have a try to set the threshold as 0.0? We will regard boxes with IoU>0.5 as positive boxes, but does IoU net have the ability to distinguish boxes with low IoU? If IoU net can only generate IoU score larger than 0.5, will it affect the performance?

vacancy commented 5 years ago

No, definitely if all boxes during training are of IoU>0.5, you can not expect the model to generalize to low-IoU boxes. This threshold is important since using all bounding boxes to train the IoU head (boxes with arbitrary IoU) will affect the performance of the prediction accuracy. (simply because there are much more training examples with larger visual variance)

jbr97 commented 5 years ago

I have some extra words about the thresholds we've tried at the bottom of #36. If you are interested that might be helpful.