search

Cc-Hy / CMKD

Cross-Modality Knowledge Distillation Network for Monocular 3D Object Detection (ECCV 2022 Oral)
Apache License 2.0
107 stars 9 forks source link

Training Problem #48

Closed TJRadarLab closed 1 year ago

TJRadarLab commented 1 year ago

Sorry to disturb you, I would like to ask you a few questions: During the calculation of MSE loss for the supervised middle layer BEV features, there are many empty values. How did you consider this? Did you add a mask to the 0 values? When I trained with other data, I found that the overall loss is affected by the BEV layer. The loss of the later anchors is very low, and the recall rate of the test results is very poor. Have you encountered this problem? Thank you, looking forward to your reply~

Cc-Hy commented 1 year ago

Hi, A straightforward method is to add a none-zero mask to the BEV loss. I remember trying it once, but the results seem to be similar. You can try it to see the effect. As for the results on other data, have you checked the teacher model's performance, if yes, is it reasonable? And can you please describe the meaning of 'later anchors', and I can offer some helps.