Here is the PyTorch implementation of the paper Very Deep Local Aggregation Networks for Point Cloud Analysis.
- May, 2024: We released the Full Code and Pretrain Weights of S3DIS. And the next step will be released after receiving clear review comments.
- Jan, 2024: We released the model code of DeepLA-Net.
- Dec, 2023: We released our project repo for DeepLA-Net, if you have any questions related to our work, please feel free to open an issue.
To make our polished code and reproduced experiments available as soon as possible, this time we will release what we already finished immediately after a validation instead of releasing them together after all work is done. We list a task list as follows:
- Release model code of DeepLA-Net;
- Release scratched config of semantic segmentation;
- ScanNet
- S3DIS
- Release scratched config of object classification;
- Release scratched config of part segmentation;
- Ubuntu: 18.04 or higher
- PyTorch: 1.11.0
- CUDA: 11.3
- To create conda environment, command as follows:
conda create -n deeplanet python=3.8 conda activate deeplanet conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch conda install llvm-openmp conda install plyfile=0.8.1 scipy=1.10.1 h5py=3.8.0 ninja
- Install pointnet2_ops:
cd utils/pointnet2_ops_lib/ pip install .
- Download S3DIS data by filling this [here].
- Change the data path in the S3DIS/config file
- Run preprocessing code for S3DIS as follows:
cd S3DIS python prepare_s3dis.py
- (Alternative) Our preprocess data can also be downloaded [here], please agree the official license before download it.
- training on S3DIS:
python train.py
- testing on S3DIS:
python test.py
- S3DIS Area5 Segmentation Results.
Year | Model | Val mIoU | Weights |
---|---|---|---|
ICCV 2021 | PT v1 | 70.4% | - |
NIPS 2022 | PointNeXt | 70.8% | - |
CVPR 2023 | PointMeta | 72.0% | - |
CVPR 2024 | PT v3 | 74.8% | - |
2024 | DeepLA-Net (Ours) | 75.8% | [link] |
- ScanNet Testset Segmentation Results. NO.1 among fully supervised methods, NO.5 among all methods. Note that, methods ranked before ours all used additional data for training.
- ResLFE_HDS means the two modules in our paper (Residual Local Feature Extraction "ResLFE" and Hybird Deep Supervision "HDS")