search

SamsungLabs / imvoxelnet

[WACV2022] ImVoxelNet: Image to Voxels Projection for Monocular and Multi-View General-Purpose 3D Object Detection
MIT License
280 stars 29 forks source link

How to make imVoxelNet support multi-classes in nuScenes dataset? #37

Closed XinchaoGou closed 2 years ago

XinchaoGou commented 2 years ago

Hi. @filaPro Thanks for sharing the code. I noticed that your original paper only mentioned results of "car" in nuScenes. I want to see how it performs under multi-classes. I modified this line to make the network output support 10 classes. https://github.com/saic-vul/imvoxelnet/blob/3512e89ca98e48aebb21a4c9e9fbe5037220b3a4/configs/imvoxelnet/imvoxelnet_nuscenes.py#L26

I modified it to num_classes=10, But still I only get results for single class "car". The other classes are all 0 for mAP. Did you tired this before? Can you help me?

filaPro commented 2 years ago

Hi @XinchaoGou ,

We also removed other classes from NuScenesMultiViewDataset. To revert them back you also need to comment these lines.

Also the performance will be rather poor if we tried only single-scale head for car detection on outdoor KITTI and nuScenes datasets. However for 10 classes it will be much better with 3 scales as in PointPillars.

XinchaoGou commented 2 years ago

Hi @filaPro Thanks a lot for your help! I modified the code as you suggested. The experiment is running. By the way, have you ever ran this experiment? Could you share your result for multi-classes as reference?

filaPro commented 2 years ago

Probably I tried something like this, but for sure don't have any logs now.

XinchaoGou commented 2 years ago

Hi @filaPro , I have the result. but the scores for other classes are very low. What could be the cause? Is it due to the size of anchor? Do you have any suggestions to improve the performance for multi-classes? image

filaPro commented 2 years ago

Yes, current anchors are well enogh only for cat category. You need much more scales and anchors for all classes of nuScenes dataset. For better performance you probably need to redesign neck a bit to return feature maps in 3 levels.

XinchaoGou commented 2 years ago

Thanks a lot for your help!