SDC-Radar-Detection

Flow

The radar object detection competetion of NYCU 2024 SDC

Installation

We use MMRotate as our detection frameworkm, which requires Python 3.7+, CUDA 9.2+ and PyTorch 1.6+.

git clone https://github.com/jimmylin0979/SDC-Radar-Detection.git
cd SDC-Radar-Detection

# pytorch
# please install following https://pytorch.org/get-started/previous-versions/
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch

# mmrotate
# please install mmrotate following https://mmrotate.readthedocs.io/en/latest/install.html#installation
cd mmroate 
pip install -U openmim
mim install mmcv-full
mim install mmdet\<3.0.0
pip install -v -e .

# install other dependencies
cd ..
pip install -r requirements.txt

PreProcess

Download SDC dataset zip file, unzip and place them under data folder with below format:

# Dataset structure 
data
├── Competition_Image
├── mini_test
├── mini_train
├── whole_train
├── test2dota.py
├── train2dota.py
├── ...

# cd into data folder
cd data

# the following command will generate 'mini_train_dota' folder for transfered dataset
python3 train2dota.py   # transfer mini_train to DOTA format
python3 test2dota.py    # transfer mini_test to DOTA format
python3 preprocess.py   # filter out too small radar value 

# the following command will generate 'Competetion_Image_preprocessed'
python3 preprocess_Competition.py   # filter out too small radar value for competition image

Training

In this competetion, I use ReDet as our detection model, to reproduce, please download pretrained weight of ReResNet backbone from ReDet, and put it on mmrotate/work_dirs/pretrain.

# cd in mmrotate folder
cd mmrotate

# the training command in mmrotate
# train model declared in CONFIG_FILE, and store everything in WORK_DIR
CONFIG_FILE=configs/sdc/???.yaml
WORK_DIR=../results/???
python tools/train.py ${CONFIG_FILE} --work-dir ${WORK_DIR}

# Example
# train with redet with config redet_re50_refpn_1x_dota_ms_rr_le90, and store the checkpoint into work dir
python3 tools/train.py configs/sdc/redet_re50_refpn_1x_dota_ms_rr_le90.py --work-dir ../results/redet_re50_refpn_1x_dota_ms_rr_le90

Inference

# inference the rotated object detection model with specific config files, checkpoint path and image root
# it will generate a viz folder for visualization and a json file for predictions
CONFIG_FILE=../results/???/???.yaml
CKPT_DIR=../results/???
IMG_DIR=???
python inference.py --config ${CONFIG_FILE} --ckpt ${CKPT_DIR} --root ${IMG_DIR}

# Example
# inference the redet with config, and checkpoint on images under `./data/Competetion_Image_preprocessed` folder
python inference.py --config results/redet_re50_refpn_1x_dota_ms_rr_le90/redet_re50_refpn_1x_dota_ms_rr_le90.py --ckpt results/redet_re50_refpn_1x_dota_ms_rr_le90/latest.pth --root ./data/Competetion_Image_preprocessed

PostProcess

# postprocess with eraseGhost function to improve model's FP
JSON_PATH=???
SAVE_PATH=???
IMG_DIR=???
pytho postprocess.py --json-path ${JSON_PATH} --save-path ${SAVE_PATH} --root-image ${IMG_DIR}

# Example 
# postprocess with prediction files (.json) and save new one as predictions.json, with provided root image dir for visualization 
python postprocess.py --json-path ./results/redet_re50_refpn_1x_dota_ms_rr_le90/rotated_redet_re50_refpn_1x_dota_ms_rr_le90.json --save-path predictions.json --root-image ./data/Competition_Image_preprocessed

jimmylin0979/SDC-Radar-Detection