search

xinge008 / Cylinder3D

Rank 1st in the leaderboard of SemanticKITTI semantic segmentation (both single-scan and multi-scan) (Nov. 2020) (CVPR2021 Oral)
Apache License 2.0
859 stars 180 forks source link

Wrong prediction of released nuscences Model #83

Closed bei181 closed 3 years ago

bei181 commented 3 years ago

Hi, your work is really impressive and creative and thanks for the released codes!

However, I met a problem when doing predictions on nusences lidar data, by using your model weights "model_load_nuscenes.pt". All the predicted labels are the same class, namely 'terrain'. Is there any mistake about your weights file?

The following figure is the visualization of the prediction.

Looking forward to your reply. image

xinge008 commented 3 years ago

How about your results of the pretrained model in nuScenes? If the number is right, I suppose the visualization may make the mistake. For example, the color mapping is not correct.

bei181 commented 3 years ago

Thanks, I have found the mistake. It is due to the failure of loading model weights. The keys in your weights are all beginned with 'polar' while in the built network they are beginned with 'cylinder'. I correct the ‘keys' and it works now, so it is not the problem on colormap.