📃 A Unified Generative Framework for Realistic Lidar Simulation in Autonomous Driving Systems

This repository mainly contains the Semantic-CARLA dataset and the PyTorch implementation of the CoLiGen model explained in our paper.

📑 Table of Contents

• Installation
• Training
• Testing & Visualisation
• Rangenet++ Experiments
• Acknowledgements

⚙️ Installation

Dependencies

• Create a conda environment and activate it.

  conda create --name coligen python=3.6
  

  conda activate coligen
  

• Install PyTorch following official instructions

The code has been tested on Ubuntu 18.04 with CUDA 12.1 and Pytorch 1.8.1.

• Install other Python packages listed in requirements.txt

  pip install -r requirements.txt
  

Dataset

• Download our Semantic-CARLA Dataset and unzip it.

🚆 Training

• Set the dataset parameters for the real dataset of your choice, e.g. kitti, in the configs/dataset_cfg/[DATASET_NAME]_cfg.yaml file. You also need to set the parameters of the simulated Semantic-CARLA dataset in configs/dataset_cfg/carla_cfg.yaml.

• Set the training parameters for the image-to-image translation model of your choice, e.g. cut, in the configs/train_cfg/[MODEL_NAME].yaml file.

• Run the training code using:

python train.py --cfg configs/train_cfg/[MODEL_NAME].yaml 

The log of the training, including tensorboard plots and the model weights are saved in checkpoints/[EXP_NAME].

🧪 Testing & Visualisation

• To test the model for calculating evaluation metrics including 'pixelAcc' on the test set, run:

python train.py --cfg checkpoints/[EXP_NAME]/[MODEL_NAME].yaml --test

• To visualise the outputs of the trained model in the image-based and point cloud representations, run:

python log_output.py --cfg checkpoints/[EXP_NAME]/[MODEL_NAME].yaml --ref_dataset_name [DATASET_NAME]

🔨 Rangenet++ Experiments

• To reproduce the experiments regarding the training of the rangenet++ model, first, create the dataset synthesised by the model:

python infer_dataset.py --cfg checkpoints/[EXP_NAME]/[MODEL_NAME].yaml --data_folder [DATASET_DIR]

• Change your current directory to rangenet/tasks/semantic/.

• Then, set the rangnet++ backbone parameters in config/arch/[BACKBONE].yaml.

• set the training dataset's label parameters in config/[DATASET_NAME].yaml.

• Finally, run:

python train.py --dataset [DATASET_DIR] -a config/arch/[BACKBONE].yaml -dc config/labels/[DATASET_NAME].yaml

📚 Acknowledgements

• CUT, Unet, and CycleGAN Implementation from taesungp

• GcGAN implementation from hufu6371

• Raydrop Synthesis implementation from kazuto1011

• Rangenet++ implementation from PRBonn