Abotu BEVFormer-S

[LuckySunnye]

trafficstars

Thanks for your great job! I want to training BEVFormer without temporal information, can you tell me how to adjust config.py and code to reproduce the results of BEVFormer-S in your paper?

[Bosszhe]

I have the same question and hope to give a tutorial about training BEVFormer-S, thx.

[yeeoon]

I also have the same question. Can I get a tutorial to reproduce the BEVFormer-S?

[czy341181]

Hi! thanks for your great work. @zhiqi-li I also want to reproduce the result of BEVFormer-s. And I get the following result, which is lower than paper.

I modify the detectors/bevformer.py forward_train function :

[zhiqi-li]

I modify the detectors/bevformer.py forward_train function :

For bevformer-S, you should set video_test_mode=False in your config.

[Roserland]

So, the result of BEVFormer-S in original paper is trained under TemporalSelfAttention while tested under degraded SelfAttention(with video_test_mode=False in config) ?

[zhiqi-li]

So, the result of BEVFormer-S in original paper is trained under TemporalSelfAttention while tested under degraded SelfAttention(with video_test_mode=False in config) ?

NO, BEVFormer-S was trained with only one sample without temporal information, and in this case, TemporalSelfAttention is just one self-attention. So that, for BEVFormer-S, you should set video_test_mode=False while testing.

[czy341181]

Thanks for your timely reply. I got the results of paper.

[Roserland]

So, the result of BEVFormer-S in original paper is trained under TemporalSelfAttention while tested under degraded SelfAttention(with video_test_mode=False in config) ?

NO, BEVFormer-S was trained with only one sample without temporal information, and in this case, TemporalSelfAttention is just one self-attention. So that, for BEVFormer-S, you should set video_test_mode=False while testing.

Thank you! And Excellent work!

[JudasDie]

Thanks for your timely reply. I got the results of paper.

Hi, could you kindly explain how to get the results of the paper, by training with video_test_mode=False, len_queue=1 and testing with video_test_mode=False? Or other setting?

[Daniel-xsy]

@czy341181 Do you train BEVFormer-S from scratch and get the results of the paper?

[wenyuqing]

So, the result of BEVFormer-S in original paper is trained under TemporalSelfAttention while tested under degraded SelfAttention(with video_test_mode=False in config) ?

NO, BEVFormer-S was trained with only one sample without temporal information, and in this case, TemporalSelfAttention is just one self-attention. So that, for BEVFormer-S, you should set video_test_mode=False while testing.

Hi， do you mean set the data queue length to 1 to train static version while keep all other settings the same with the temporal version？