search

tudelft3d / SUMS-Semantic-Urban-Mesh-Segmentation-public

SUMS: Semantic Urban Mesh Segmentation.
GNU General Public License v3.0
56 stars 14 forks source link

The competing method pointnet training and testing #16

Closed TX2liu closed 1 year ago

TX2liu commented 1 year ago

Hi Doctor Gao: I am interest in your work and very appreciate the source code of your SUMS-Semantic-Urban-Mesh-Segmentation-public. The point clouds (from the mesh by the release exe) which are applied in the competing method pointnet model, but I found the batchsize=1 and the CUDA OUT of MEMORY( NVIDIA RTX 3090).
I use the pointnet++(pointnet2_largemsg) to train the point clouds (from the mesh by the release exe), but I can not found the test code which is to product the semantic_pointcloud for urbanmesh. Thus, would you help me solve these problem. Thanks very much. I appreciate you from the bottom of my heart. Liu.

WeixiaoGao commented 1 year ago

Hi, eval.py is the test code and you need to config the 'eval.yaml'. For CUDA OUT of MEMORY of pointnet, you can try to set cuda: -1 in ../conf/training/default.yaml. There might be some conflicts between the higher versions of Pytorch and torch_point3d.

TX2liu commented 1 year ago

Hi, eval.py is the test code and you need to config the 'eval.yaml'. For CUDA OUT of MEMORY of pointnet, you can try to set cuda: -1 in ../conf/training/default.yaml. There might be some conflicts between the higher versions of Pytorch and torch_point3d.

Thanks, the eval.py can compute the result. how can I output the semantic point clouds? Best wishes! Liu

WeixiaoGao commented 1 year ago

If you have set vis_data: [test] in the visualization config file, you should be able to find the output point cloud in checkpoint_dir.