search

WangYueFt / detr3d

MIT License
745 stars 140 forks source link

Question about with_box_refine in Detr3DHead #8

Open BoomSky0416 opened 2 years ago

BoomSky0416 commented 2 years ago

According to my experiment result, the prediction result of with_box_refine=False is better than that of with_box_refine=True. mAP of with_box_refine=False is 0.365, and mAP of with_box_refine=True is 0.35. I didn't see any related discussion on this comparison in the paper. Is there any explanation for this? Thank you!

a1600012888 commented 2 years ago

Hi, You are right, we didn't run lots of experiments for with_box_refine=False, and there is no discussion about this.

The results of your experiments are supervising, at least for me. We always used the default iterative refine, and we have one experiment on the iterative refine in Table-5(still, this table of experiments cannot explain your findings.).

BoomSky0416 commented 2 years ago

Thanks for reply! I think it is necessary to do experiments to verify the role of the box_refine module. I would appreciate it if you could run experiments to verify it. Because from my experiments, it seems unnecessary to add this module for improving prediction results. Thanks!

a1600012888 commented 2 years ago

Thanks

xuchen-dev commented 6 months ago

According to my experiment result, the prediction result of with_box_refine=False is better than that of with_box_refine=True. mAP of with_box_refine=False is 0.365, and mAP of with_box_refine=True is 0.35. I didn't see any related discussion on this comparison in the paper. Is there any explanation for this? Thank you!

hi, also set with_box_refine=False and change input size to 512*960 , use Resnet50 backbone . But after 24 epoch the map is just 13, could you give some advise