xinge008

It is better to increase your GPU memory. 2GB is too small for the modern networks, especially for the large-scale point clouds.

I have provided the data processing of mulit-scan point cloud segmentation (for SemanticKITTI); You can check this commit https://github.com/xinge008/Cylinder3D/commit/7df59464c9b1828f9f97084997b3d47beda612cd

They share same hyper-parameters, except the num_class;

you can check this issue https://github.com/xinge008/Cylinder3D/issues/28

1. Yes 2. In the SemanticKITTI setting, the iou of car is about 97.2%; You can infer the pre-trained model with semanticKITTI validation set to check it. In your own...

We do not provide the script for the challenge submission; For validation set, you can find it in demo_folder.py

`how to judge the best precision with minium error of coordinates transformation?` In my opinion, these hyper-parameters are closely related to your deploy environments, and usually need a lot of...

Different version of spconv makes the difference. It should be spconv=1.2.1 for Cylinder3D in this repo.

In this line, the voxel prediction has been transformed to point-wise prediction, and the point-wise labels are used for evaluation.

trainval的所有Keyframe blobs跟lidar blobs数据都需要，不过还是建议都下载下来；使用方法在https://github.com/nutonomy/nuscenes-devkit 跟detection是类似的