Perception Imitation: Towards Data-free Simulator for Autonomous Vehicles

Getting Start

conda create --name imitator python=3.9
conda activate imitator
pip install -r requirements.txt

Data Preparation

1. Download the source data of nuScenes and CARLA, and extract them into ~/Datasets/.

2. Download the target models' results (ImitatorData）and put them in ~/Perception_Imitation/data/. The relationship of different target model trained with different dataset and its results are illustrated in the table:

   Dataset	Results
   nuScenes	nuscenes_ %TargetModel _match_gt.json
   Kitti	kitti_%TargetModel _match_gt.pkl
   CARLA	carla_ %TargetModel _match_gt.pkl

   where TargetModel = {‘pp’, ‘cp’, ‘pvrcnn’}.

3. The data is structured as:

   |— Datasets 
   |  |— carla 
   |  |  |— carla_new 
   |  |  |— Maps 
   |  |— nuScenes 
   |  |  |— maps 
   |  |  |— v1.0-mini 
   |  |  |— v1.0-trainval 
   |  |  |— ... 
   |— Perception_Imitation 
   |  |— data 
   |  |  |— nuscenes_pp_match_gt.json 
   |  |  |— kitti_cp_match_gt.pkl 
   |  |  |— ... 
   |  |— dataset 
   |  |  |— ... 
   |  |— ... 
   

Training

Baseline

1. Running code :

   python main/train.py --cfg_dir utils/config/samples/sample_carla 

2. Change datasets

   Replace "sample_carla" with "sample_nuscenes" or "sample_kitti".

3. Change target model

   Modify different value of the key "target_model" in file sample_carla/dataset/scene_occ_xxx.yaml .

Our imitator

1. Running code:

   python main/train.py --cfg_dir utils/config/samples/sample_carla_improve 
   

2. The rest are the same as the baseline.

Evaluation

Gaussian and Multimodal

Running code:

python main/test_baseline_gaussian.py --cfg_dir utils/config/samples/sample_carla_improve/ 

Baseline or Imitator

Running code:

python main/test_baseline_evaluate.py --cfg_dir utils/config/samples/sample_carla_improve/ 

Visualization

Running code:

python main/plot_qualitative_results.py 

and the images will be saved in the format of eps in ~/ADModel_Pro/ output/pic/.