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

This repository contains a PyTorch implementation of RandLA-Net on Semantic KITTI.

Preparation

1. Clone this repository

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

3. Install python functions. the functions and the codes are copied from the official implementation with Tensorflow.

sh compile_op.sh

4. Download the Semantic KITTI dataset, and preprocess the data:

python utils/data_prepare_semantickitti.py

Note: Please change the dataset path in the 'data_prepare_semantickitti.py' with your own path.

Train a model

python main_SemanticKITTI.py

This repository implements the official version as much as possible. There are following differences:

1. We use Conv2D instead of ConvTranspose2D in the decoder part. Since the strides are [1,1] and conv kernels are [1,1], we think the Conv2D and ConvTranspose2D are the same in this condition.

2. We delate the bias in the Conv2D which is followed by a BN layer, since the bias is meaningless in the Conv2D if the Conv2D is followed by a BN layer.

3. We evaluate the network for one epoch after every 10 training epoches.

Results

We train this network for 60 epoches, and the eval results in the 8-th Sequence are as follows:

mean IoU: 52.6

IoU: 
car:           93.07
bicycle:       16.70
motorcycle:    20.64
truck:         67.74
other-vehicle: 41.46
person:        48.79
bicyclist:     65.74
motorcyclist:   0.00
road:          91.44
parking:       42.21
sidewalk:      76.50
other-ground:   6.15
building:      85.55
fence:         38.86
vegetation:    84.19
trunk:         58.49
terrain:       73.63
pole:          51.63
traffic-sign:  36.05

There are some differences with the results in the paper, but we think the overall results is acceptable.

The checkpoint is in the output folder.