Bevformer temporal and no temporal configs are the same

[jasonkhadka]

Following configs are the same. How did you test non-temporal model?
Temporal config:

zoo/BEVFormer/projects/configs/bevformer/bevformer_base.py

Non-temporal configs:

zoo/BEVFormer/projects/configs/bevformer/bevformer_base_no_temp.py

[Daniel-xsy]

I've corrected it. You only need to change the following argument to switch between temporal and non-temporal mode.

video_test_mode=False

Thanks for pointing out this issue.

[jasonkhadka]

@Daniel-xsy Isn't that only for testing? The training is still done with temporal information with Temporal Attention?

[Daniel-xsy]

Yes, that's correct. In the official BEVFormer paper, the authors trained two versions: BEVFormer and BEVFormer-S. The BEVFormer-S is trained without temporal information, i.e., video_test_mode=False for both training and testing. You can find more details here.

However, since we couldn't find the training configuration and checkpoint of BEVFormer-S, we opted to use the BEVFormer checkpoint and turn off video_test_mode during testing. This might cause a discrepancy between training and testing, as evident from the lower performance of BEVFormer-non-temporal compared to the official results. We might try to reproduce this version and update the checkpoint in the future if possible.