RandLA-Net: Efficient Semantic Segmentation of Large-Scale Point Clouds

This repository contains a PyTorch implementation of RandLA-Net on S3DIS and Semantickitti.

This repository is mainly based on the repository.

Preparation(S3DIS as example)

Clone this repository.
Install some Python dependencies, such as scikit-learn. All packages can be installed with pip.

Environment:

ubuntu 18.04
python 3.7.16
torch 1.12.1
numpy 1.21.5
torchvision 0.13.1
scikit-learn 0.22.2
pandas 1.3.5
tqdm 4.64.1
Cython 0.29.33 (Cython is important!)

Install python functions. the functions and the codes are copied from the official implementation with Tensorflow.
```
sh compile_op.sh
```
Attention: please check out ./utils/nearest_neighbors/lib/python/KNN_NanoFLANN-0.0.0-py3.7-linux-x86_64.egg/ and copy the .so file to the parent folder.

Update in 2023.2.23: We provide a .so file for python3.7, and you don't need to compile the cpp code if you are using python3.7.
Download the Stanford3dDataset_v1.2_Aligned_Version dataset, and preprocess the data:
```
python utils/data_prepare_s3dis.py
```
Note: Please change the dataset path in the 'data_prepare_s3dis.py' with your own path.

Train a model(S3DIS as example)

  python main_S3DIS.py

Test a model(S3DIS as example)

  python test_S3DIS.py

Results

S3DIS

We train this network for 100 epoches, and the eval results(after voting) in the Area 5 are as follows: mIoU = 62.59%

--------------------------------------------------------------------------------------
62.59 | 91.92 96.32 81.43  0.00 20.59 61.54 55.26 75.03 84.95 56.12 72.33 65.93 52.29 
--------------------------------------------------------------------------------------

While SQN shows the result(mIoU) of RandLA-Net of Area5 is 63.59.

Our results are close to the original paper.

SemanticKITTI

We train the network for 100 epoches, and the eval results(after voting) in the Seq 08 are as follows: mIoU = 54.62%

--------------------------------------------------------------------------------------------------------------------------
54.62 | 93.12 18.31 30.68 79.83 45.59 51.81 70.18  0.00 92.15 41.53 78.42  1.09 87.61 46.32 84.30 58.67 72.12 52.28 33.67 
--------------------------------------------------------------------------------------------------------------------------

The checkpoint is in the output folder.

Model

We provode the pretrained models for S3DIS(Area5) and SemanticKITTI.

S3DIS(Area5): link

SemanticKITTI: link

hiyyg / RandLA-Net-Pytorch-New