search

aharley / simple_bev

A Simple Baseline for BEV Perception
MIT License
483 stars 75 forks source link

No performance improvement after using radar data #34

Closed cyn-liu closed 1 year ago

cyn-liu commented 1 year ago

Hi @aharley :

Thanks for your job, which give me some ideas. In my BEV detection model, I Rasterisation the RADAR data and concatenated with the BEV feature output from LSS according to dimensions, but the result was 2-3 points lower.

Do you know why? Or tell me some potential problem.

aharley commented 1 year ago

Did you try rasterizing and concatenating the same way as in this repo? In this repo, it gives a very reliable boost.

cyn-liu commented 1 year ago

Did you try rasterizing and concatenating the same way as in this repo? In this repo, it gives a very reliable boost.

@aharley I referred to the idea of simple_bev and did same radar preprocessing and concatenation method, but my BEV detection model adopts different projection method of image features into BEV features, I don't think this will affect the performance improvement brought by Radar.

I found the simple_bev uses the InstanceNorm2d normalization method, but my model uses BN, Is this the key reason why the model performance cannot be improved?

Best regards Cynthia

aharley commented 1 year ago

I like to avoid BN in general, as it seems tricky to tune. If you are using a small/medium batch size, IN may indeed help stabilize things.