Exact parallel prediction heads used in training?

Dear authors,

Thank you very much for sharing us the excellent code for the 'KM3D' paper .

I would like to reproduce your results on the KITTI dataset and then apply your method to a customed dataset. During the reproduction, I can match the results reported in the tables 'KM3D Baseline and Model Zoo' but with all the parallel heads turned on for training.

More specifically, heads with heatmap prediction for projected 3D keypoints ('--hm_hp_weight'), 3D keypoints offset ('--off_weight'), 2D box width/height estimation ('--wh_weight') are used to obtain the exact same result if I am not wrong. However I have noticed in the framework (figure 1) of 'KM3D' paper that only 5 parallel heads are used and the previous mentioned 3 heads are not included. Also it seems you mentioned in the supplementary material of 'KM3D' that regressing keypoints are better than heatmap keypoints.

When I tried to turn off the 3 heads in the code, the performance drops significantly. It seems these 3 heads are still quite important. For these two train settings, I am both using 'faster.py' for inference.

I am a bit of confused about the exact heads used in the paper. Also could you please help illustrate the reason to remove the 3 heads from 'RTM3D' to 'KM3D' because from my experiment they still play a considerate role.

Thank you very much!

Banconxuan / RTM3D

Exact parallel prediction heads used in training? #34