search

aharley / simple_bev

A Simple Baseline for BEV Perception
MIT License
502 stars 79 forks source link

I got strange feature map when I use simpleBEV on BEVFusion for 3D object detection #60

Open odroidodroid opened 3 months ago

odroidodroid commented 3 months ago

Hi, I tried 3D object detection with your work, but got a strange feature_bev when I trained a model. Below is the strange feature_bev.

simplebev_camera_channel_0

I replaced LSS to simpleBEV in BEVFusion (from MIT) and loaded image, rots, trans and intrinsics from sample directly like you did. (image size and bev size is same with simpleBEV)

When I start training BEVFusion with simpleBEV, I could get similar feature_bev with simpleBEV one which starts training.

However, after one epoch training on BEVFusion, I got the strange feature_bev.

Do you know a reason or can you guess what is a problem?

I think the biggest difference between two codes are losses. Are they affect the feature_bev?