[ICRA2022] AutoPlace: Robust Place Recognition with Single-Chip Automotive Radar

@article{cai2021autoplace,
  title={AutoPlace: Robust Place Recognition with Low-cost Single-chip Automotive Radar},
  author={Cai, Kaiwen and Wang, Bing and Lu, Chris Xiaoxuan},
  booktitle={2022 IEEE International Conference on Robotics and Automation (ICRA)},
  pages={3475--3481},
  year={2022},
  organization={IEEE}  
}

0. Environment Setup ⚙️

• Ubuntu 18.04, python 3.8, A100
• PyTorch 1.8.1 + CUDA 11.1
• nuscenes-devkit 1.1.1

pip install --upgrade pip
pip install -r requirements.txt

Go to the nuscenes-devkit package directory (depends on which python you are using), and override the function from_file_multisweep(...) in site-packages/nuscenes/utils/data_classes.py with the function provided in autoplace/nuscenes-devkit_override.py.

1. Dataset preprocessing 📥

You may need to download nuScenes dataset (radar) from nutonomy/nuscenes-devkit.

cd autoplace/preprocess
./gene_woDTR.sh
./gene_wDTR.sh

the generated processed dataset folder should be like:

dataset
├── 7n5s_xy11
│   ├── pcl_parameter.json
│   ├── img
│   ├── pcl
│   ├── rcs
│   ├── nuscenes_test.mat
│   ├── nuscenes_train.mat
│   ├── nuscenes_val.mat
│   ├── database.csv
│   ├── train.csv
│   └── test.csv
└── 7n5s_xy11_remove
    ├── ...

to save you time on downloading/preprocessing the nuScenes dataset, you may as well download my processed dataset from Dropbox and then arrange it in the above way.

2. AutoPlace 🚗

1. train SpatialEncoder (se)

   cd autoplace
   
   python train.py  --nEpochs=50 --output_dim=9216 --seqLen=1 --encoder_dim=256 --net=autoplace --logsPath=logs_autoplace --cGPU=0 --split=val --imgDir='dataset/7n5s_xy11/img' --structDir='dataset/7n5s_xy11'

2. train SpatialEncoder+DPR (se_dpr)

   cd autoplace
   
   python train.py  --nEpochs=50 --output_dim=9216 --seqLen=1 --encoder_dim=256 --net=autoplace --logsPath=logs_autoplace --cGPU=0 --split=val --imgDir='dataset/7n5s_xy11_removal/img' --structDir='dataset/7n5s_xy11'

3. train SpatialEncoder+TemporalEncoder (se_te)

   cd autoplace
   
   python train.py  --nEpochs=50 --output_dim=4096 --seqLen=3 --encoder_dim=256 --net=autoplace --logsPath=logs_autoplace --cGPU=0 --split=val --imgDir='dataset/7n5s_xy11/img' --structDir='dataset/7n5s_xy11'

4. train SpatialEncoder+TemporalEncoder+DPR (se_te_dpr)

   cd autoplace
   
   python train.py  --nEpochs=50 --output_dim=4096 --seqLen=3 --encoder_dim=256 --net=autoplace --logsPath=logs_autoplace --cGPU=0 --split=val --imgDir='dataset/7n5s_xy11_removal/img' --structDir='dataset/7n5s_xy11'
   

5. evaluate a model

   cd autoplace
   
   python train.py --mode='evaluate'  --cGPU=0  --split=test --resume=[logs_folder]

6. apply RCSHR on SpatialEncoder+TemporalEncoder+DPR model (You may need to evaluate SpatialEncoder+TemporalEncoder+DPR model first): modify the path se_te_dpr in autoplace/postprocess/parse/resume_path.json to [logs_folder], then

   cd autoplace/postprocess/parse 
   
   python parse.py  --rcshr --model=se_te_dpr

7. To generate (1) Reall@N curve, (2) PR curve, (3) F1 Score and (4) Average Precision

   cd autoplace/postprocess/vis
   
   python ablation_figure.py 
   python ablation_score.py 

3. SOTA methods ⚔