L3-Net

This is the reproduction of Baidu's L3-Net which is used for automaous driving especially in the field of relocalization. L3_Net is proposed in the paper: Towards Learning based LiDAR Localization for Autonomous Driving, which was published on CVPR. The architecture of the relocalization neural network is as follows:

Install

Install the required packages

git clone https://github.com/LZX-0201/L3-Net.git
cd L3-Net
pip install -r requirements

Download Apollo-SouthBay Dataset

https://apollo.auto/southbay.html

Data Pre-processing

Config the data pre-processing process by modifying the configuration file.

vim uitls/config/samples/sample_L3Net_dataprepare/root_config.yaml

Pre-process the training data before training, which selects key-points in the point clouds and uses data from IMU to calculate the ground truth offset of cars pose.

python utils/save_prepared_data.py

Train

Configuration

Config the data pre-processing process by modifying the configuration file.

vim uitls/config/samples/sample_L3Net/root_config.yaml
vim uitls/config/samples/sample_L3Net/dataset/HighWay237.yaml
vim uitls/config/samples/sample_L3Net/model/probability_offset_model.yaml

Train the network

cd main
python main/train.py --cfg_dir="../utils/config/samples/sample_L3Net/"

Ps: It doesn't take too many epoches for the model to converge. The checkpoints will be saved in L3-Net/checkpoints.

Test

Test the network using the saved checkpoint.

cd main
pyton test_L3Net.py --cfg_dir="../utils/config/samples/sample_L3Net/" --check_point_file=../checkpoints/<epoch_num_check_point.pth>

The test indicators is same with the paper, which include:

Horiz. RMS; Horiz. Max; Long. RMS; Lat. RMS; <0.05m Pct. ; <0.6m Pct. ; <0.7m Pct. ; <0.8m Pct. ; Yaw. RMS; Yaw. Max; <0.1° Pct. ; <0.3° Pct. ; <0.6° Pct.

Supplement

This repository use the previous frame point cloud as the map.
This repository doesn't contain the Temporal Smoothness part.
This repository is used only for study, not for commercial use.

GUDKS/L3-Net