search

SxJyJay / MSMDFusion

[CVPR 2023] MSMDFusion: Fusing LiDAR and Camera at Multiple Scales with Multi-Depth Seeds for 3D Object Detection
Apache License 2.0
167 stars 10 forks source link

the training log files #39

Closed Wang-pengfei closed 3 months ago

Wang-pengfei commented 3 months ago

Hi, can you provide me with the training log files for the second stage of transfusion_nusc_voxel_LC and MSMDFusion_nusc_voxel_LC for my reference? Thank you!

SxJyJay commented 3 months ago

I can not find the relevant training logs. You can provide intermediate results and I can help you to check them.

Wang-pengfei commented 3 months ago

Thank you for your assistance!

I'm currently working on reproducing TransFusion. After training for 20 epochs on LiDAR, I obtained results with mAP: 0.6463 and NDS: 0.6979, which are close to the original TransFusion-L.

However, when I continued training for 6 epochs for fusion, the first epoch yielded very low results with mAP: 0.3156 and NDS: 0.2860. Is this normal?

Wang-pengfei commented 3 months ago

I would like to know where the initialization weights for the 2D model come from. Have you retrained the 2D model?

SxJyJay commented 3 months ago

It is not normal in my memory. The performance will not decrease dramatically in the second training stage. Please refer to here for the pretrained 2D backbone.