RandLA-Net-pytorch

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

Updates:

• We extend the model with to train with one synthetic dataset, SynLiDAR
• We replaced the weighted cross entropy in the oroginal implementation with focal loss to alliverate the influence of class imbalance
• Add frequency weighted mIoU as indicator for the training & validation
• Fix some minor bugs in the original implementation
• We improve the mIoU on the validation set from 53.1% on validation set to 55.1%.
• This is a good starting point & backbone choice for those who plan to start their research on point clouds segmentation.

Previous:

• support SemanticKITTI dataset now. (Welcome everyone to develop together and raise PR)
• 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	0.551

• 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
55.1	0.939	0.092	0.347	0.659	0.453	0.548	0.707	0.000	0.920	0.401	0.784	0.006	0.886	0.520	0.855	0.627	0.747	0.568	0.403

A. Environment Setup

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

1. Install python packages

pip install -r requirements.txt

2. Compile C++ Wrappers

sh compile_op.sh

B. Prepare Data

Download the Semantic KITTI dataset, and preprocess the data:

python data_prepare_semantickitti.py --src_path path/to/sequences --dst_path destination/for/preprocessed/sequences

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

1. Training

python3 train_SemanticKITTI.py <args>

Options:
--backbone           select the backbone to be used: choices=['randla', 'baflac', 'baaf']
--checkpoint_path    path to pretrained models(if any), otherwise train from start
--log_dir            Name of the log dir, the file will be in logs/ suffixed with start time
--max_epoch          max epoch for the model to run, default 80
--batch_size         training batch size, default 6 (indicated in oroginal implementation), modify to full utilize the GPU/s
--val_batch_size     batch size for validation, default 30
--num_workers        number of workers for I/O
--focal              whether to use focal loss or not, default True
--focal_gamma        gamma for focal loss, default 2

2. Testing

python3 test_SemanticKITTI.py <args>

Options:
--backbone           select the backbone to be used: choices=['randla', 'baflac', 'baaf']
--infer_type         all: infer all points in specified sequence, sub: subsamples in specified sequence
--checkpoint_path    required. path to the model to test
--test_id            sequence id to test
--index_to_label     whether to convert .npy(label 0-19) back .label(original labels)

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

1. Visualization

python3 visualize_SemanticKITTI.py <args>

options:

--dataset path to dataset for visualization

--config dataset config, default utils/semantic-kitti.yaml

--sequence sequence to visualize

--predictions location for predictions

2. Evaluation

• Example Evaluation code

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

Acknowledgement

• This repo is improved from RandLA-Net PyTorch
• Original Tensorflow implementation link