Henry

When will the BEVFormer release?

1

Seems implemented by mmdet3d: https://github.com/zhiqi-li/BEVFormer

Can this project be used in other fields, like clothing retrieval?

1

Thanks for sharing this great baseline ！

https://github.com/tianzhi0549/FCOS/blob/0eb8ee0b7114a3ca42ad96cd89e0ac63a205461e/fcos_core/modeling/rpn/fcos/loss.py#L252 how about use different num_pos for each gpu? Thanks a lot!

why the default anchor is 10,13, 16,30, 33,23.... and as for my kmeans result on VOC2007, there aren't the anchor size like 10 ? (the first is 23,31)

1

same as above.

RT

I transfer the code to mmdetection v2.0, but the result is much worse than general training pipeline. What might be the reason？

3

I already modified the key of pretrained models, and load it into the model.

Try to build 3d perception model SMOKE, bug got error: no such package '@caffe//

1

I use the apollo 7.0, and mv BUILD_ to BUILD, and use the command: bazel test --config=unit_test //modules/perception/camera/test:camera_app_obstacle_camera_perception_test but I got error below: (07:16:40) ERROR: /apollo/modules/perception/camera/test/BUILD:6:8: no such package '@caffe//':...

Is it possible to adapt BEVDepth-pure to KITTI dataset

2

Can big olds give me some hints? Thanks

RT

I try to set loss_weight of loss_bbox2d and loss_consistency to 1, and set input resolution to origin 1920 x 1280, but the loss_bbox2d is much higher than KITTI Dataset, why...