Here is the PyTorch implementation of our paper.

Please feel free to open an issue if you have any questions or suggestions.

Details the sampling algorithm we use.

cutils contains cpp/cuda functions (neighbor edge-max-pooling, grid subsampling, knn search).

util.py has some python functions (state saving/loading, metric calculation).

timm is a tiny timm. Only functions used are kept to reduce dependencies.

pointnet2_ops_lib is from Pointnet2_PyTorch.

Dataset specific implementation.

Basically all contain: train, test, config, model, dataset, readme.md and pretrained (folder containing training log and trained model).

Please be aware that the code used to train our pretrained models has been (and likely will be) slightly amended, so you may notice discrepancies in the training log. (currently only fixed reset of corls)

We list libraries (tested version) and the way we install it for reference.

We believe newer versions are generally fine and there's no need to follow this guide exactly.

Create and activate environment using conda.

Add conda-forge channel to the environment first and comment the original channels.

Switch mirrors (channel sources) if necessary.

conda create -n dela python=3.10

conda activate dela

Install pytorch.

Specify a version with cuda.

conda install pytorch=1.13.1=cuda112py310he33e0d6_200

Other tools should be automatically installed in this step, e.g., numpy (1.24.3), cudatoolkit (11.8.0), ...

An important thing to note is, as of 2023.5, under this setting, llvm-openmp >= 16.0.0 can cause a problem with pytorch's multi threaded dataloading.

We currently do not know the exact cause, and simply use version 15.

conda install llvm-openmp=15

Then check if cuda kernels can be compiled smoothly.

Remember to set CUDA_HOME and TORCH_CUDA_ARCH_LIST properly.

python
Python 3.10.11 | packaged by conda-forge | (main, May 10 2023, 18:58:44) [GCC 11.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from utils.cutils import knn_edge_maxpooling

Compilation verbose is set to True and can be disabled in line 15 of cutils' init file if this step goes fine.

Install h5py.

conda install h5py=3.8.0

Install pointnet2_ops.

cd utils/pointnet2_ops_lib/
pip install .

Install pointnet2_ops.

cd utils/pointnet2_ops_lib/
pip install .

Install plyfile and scipy.

conda install plyfile=0.8.1 scipy=1.10.1

To train & test on a specific dataset, first enter corresponding directory and set up the dataset following readme.md.

python train.py& 
disown

Saved logs and models are under output/.

Check train.py and other scripts for more details.

To resume from a checkpoint, follow train.py (search for "resume").

python test.py

As random seeds are not fixed, output for segmentation is slightly unstable.

The results of pretrained models should be near reported ones with high probability.

Here we list mean $\pm$ std of 3 random runs.

Training time is measured on Ubuntu 22.04 with an RTX 4090 GPU and a 13600k CPU.

Memory consumption is checked from nvidia-smi with pytorch 1.13.1. Newer pytorch versions may slightly save memory.

mIoU: 73.48 $\pm$ 0.46

Time: around 110 min, 90 min (parallelize 2 jobs), 140 min (with gradient checkpoint), 115 min (parallel + checkpoint)

Memory: 9496 MB, 6452 MB (with gradient checkpoint)

mIoU: 75.88 $\pm$ 0.07

Time: around 320 min, 410 min (with gradient checkpoint)

Memory: 21430 MB, 13152 MB (with gradient checkpoint)

Results without voting:

ins mIoU: 86.94 $\pm$ 0.08

cat mIoU: 85.41 $\pm$ 0.39

Time: around 80 min

Memory: 5004 MB

OA: 90.15 $\pm$ 0.25

mAcc: 89.01 $\pm$ 0.39

Time: around 30 min

Memory: 2552 MB

OA: 93.75 $\pm$ 0.18

mAcc: 91.15 $\pm$ 0.61

Time: around 50 min

Memory: 2594 MB