search

sfzhang15 / RefineDet

Single-Shot Refinement Neural Network for Object Detection, CVPR, 2018
Other
1.43k stars 393 forks source link

Why remove easy negative of ARM in training ODM? #172

Closed Edwardmark closed 5 years ago

Edwardmark commented 5 years ago

Why remove easy negative of ARM in training ODM? The easy negatives with neg_conf>=0.99 can be easily filtered in OHEM, so why bothered to remove them before OHEM in ODM training? Confused about why refineDet can deal with class imbalance, is it the job od OHEM?

sfzhang15 commented 5 years ago

@Edwardmark Good question. As you have said, the easy negatives with neg_conf>=0.99 in ARM can be easily filtered by OHEM in ODM. Actually, most of the easy negatives in ARM can be filtered by OHEM in ODM, but still has a few negatives that are easy in ARM but hard in ODM. This is the reason why the two-step classification only has 0.5 mAP improvement. We believe that using the focal loss to replace the OHEM loss can make the two-step classification more useful.

Edwardmark commented 5 years ago

@sfzhang15 Thank you very much, I understand that now.

ujsyehao commented 5 years ago

@Edwardmark You can refer to Selective Refinement network, refindet + focal loss.