Zhy

This is a incremental imrpovement that brought about 0.2 mAP, we found that a soft maxtrix is a little bit better than a binary one.

They are not equivalent since geom_idx may contains serveral same indexes

Well, this is not expected for me. In that case, the observed improvement might be only fluatuation. But this proved that a binary matrix will work well as stated in...

图一只是一个我们方法和已有方法的比较，我们只使用了Ldpl这个loss 回归更容易实现是因为如果没有区域选择步骤，背景区域的teacher reg是没有意义的，使用这些信息学习会导致性能下降。而区域选择可以有效的筛选出reg有意义的部位

hi, please test the Teacher (the EMA model)

can you offer the full log file?

please offer the full config log and the history evaluation results

Seems like nothing goes wrong. This might be caused by different GPU (2080ti & V100) or random states.

Yes, the weight is right, but we used EMA at 10000 step and BURN_IN_STEP=20000 since VOC is a full dataset. An early burn-in stage can cause unstable training due to...

seems you are using voc2007 val as eval set, but we use trainval for training and voc2007 test as evaluation set. beside, the best performance is obtained at 60K iter...