In object detection with SSD Mobilenetv2 + FPN(TF2 Object detection API), what are the possible reasons for hard example mining to not work well?

Question

I am using the Tensorflow Object Detection API to build a detection model. Hard example mining seemed to work really well with SSD+Mobilenetv2 model(used with the TF1 version of the API). However with similar settings in the TF2 version with FPN SSD+Mobilenetv2+FPN model, I achieve similar metrics for mAP on relevant category but see a lot more false positives in evaluation even after adding hard example mining. What could be the possible reasons for that, any other ways to reduce false positives?

user_7282944 · Answer

Solution is to not use negative hard mining, but instead use focal loss. This is in line with the Focal Loss / RetinaNet paper : https://arxiv.org/pdf/1708.02002v2.pdf

In object detection with SSD Mobilenetv2 + FPN(TF2 Object detection API), what are the possible reasons for hard example mining to not work well?

1 Answers