Awesome

RandLA-Net-pytorch

This repository contains the implementation of RandLA-Net (CVPR 2020 Oral) in PyTorch.

We only support SemanticKITTI dataset now. (Welcome everyone to develop together and raise PR)
Our model is almost as good as the original implementation. (Validation set : Our 52.9% mIoU vs original 53.1%)
We place our pretrain-model in pretrain_model/checkpoint.tar directory.

Performance

Results on Validation Set (seq 08)

Compare with original implementation

Model	mIoU
Original Tensorflow	0.531
Our Pytorch Implementation	0.529

Per class mIoU

mIoU	car	bicycle	motorcycle	truck	other-vehicle	person	bicyclist	motorcyclist	road	parking	sidewalk	other-ground	building	fence	vegetation	trunk	terrain	pole	traffic-sign
52.9	0.919	0.122	0.290	0.660	0.444	0.515	0.676	0.000	0.912	0.421	0.759	0.001	0.878	0.354	0.844	0.595	0.741	0.517	0.414

A. Environment Setup

Click this webpage and use conda to install pytorch>=1.4 (Be aware of the cuda version when installation)
Install python packages

pip install -r requirements.txt

Compile C++ Wrappers

bash compile_op.sh

B. Prepare Data

Download the Semantic KITTI dataset, and preprocess the data:

python data_prepare_semantickitti.py

Note:

Please change the dataset path in the data_prepare_semantickitti.py with your own path.
Data preprocessing code will convert the label to 0-19 index

C. Training & Testing

Training

python3 train_SemanticKITTI.py <args>

Testing

python3 test_SemanticKITTI.py <args>

Note: if the flag --index_to_label is set, output predictions will be ".label" files (label figure) which can be visualized; Otherwise, they will be ".npy" (0-19 index) files which is used to evaluated afterward.

D. Visualization & Evaluation

Visualization

python3 visualize_SemanticKITTI.py <args>

Evaluation

Example Evaluation code

python3 evaluate_SemanticKITTI.py --dataset /tmp2/tsunghan/PCL_Seg_data/sequences_0.06/ \
    --predictions runs/supervised/predictions/ --sequences 8

Acknowledgement

Original Tensorflow implementation link
Our network & config codes are modified from RandLA-Net PyTorch
Our evaluation & visualization codes are modified from SemanticKITTI API