search

fundamentalvision / BEVFormer

[ECCV 2022] This is the official implementation of BEVFormer, a camera-only framework for autonomous driving perception, e.g., 3D object detection and semantic map segmentation.
https://arxiv.org/abs/2203.17270
Apache License 2.0
3.23k stars 524 forks source link

rotate_center #96

Open minrui-hust opened 2 years ago

minrui-hust commented 2 years ago

rotate_center is default to be [100, 100], that is ok for base config, which has bev resolution of 200x200, but for small and tiny config, which has 150x150 and 50x50, it seems rotate_center is not the center of the bev image for these configs. is this on purpose or a mistake?

zhiqi-li commented 2 years ago

Yeah, it's a mistake. Fixing it will obtain higher scoresđŸ˜‚.

minrui-hust commented 2 years ago

Thanks for your reply. Another question: In temporal_self_attention.py, you concate the ‘query’ and ‘value[:bs]’ in line 195, which i guess you mean to concatenating previous bev_query and current bev_query in feature dimension, howevery 'value' is reshaped from shape [B,2,query_len,F]( in encoder.py line 203), which has batch in first dimension, after reshape, ’value[:bs]‘ will not be the first batch(which is according to pre_query), but half pre_quey and half cur_query. Maybe reshape from [2, B, query_len, F] to [2*B, query_len, F] is what you intended?

zhiqi-li commented 2 years ago

Thanks for pointing out our mistakes, would you like to raise a pull requests and become a contributor of BEVFormer?

minrui-hust commented 2 years ago

very glad to, i will make a PR after train and eval.

Zhutianyi7230 commented 1 year ago

@zhiqi-li @minrui-hust hello,I have a question. if I change the bev coords to [200,100](means bev_h=200,bev_w=100),and I change the rotate center to be [50,100], will it work? Thanks.

usergxx commented 1 month ago

ge the bev coords to [200,100](means bev_h=200,bev_w=100),and I change the rotate center to b

I think it will work, but we don't need to set the center. The rotation function by default rotates around the center of the tensor.