Preparation

Please see dataset/README.md to get more details about our datasets-VIL100

Please see INSTALL.md to install environment and evaluation tools

Before training, we should download datasets-VIL100 and models

datasets :  https://pan.baidu.com/s/1NkP_5LMLTn6qsu9pSbyi0g - iy16

models : https://pan.baidu.com/s/1_o13TBbTf258-j7iACDS2Q - sgh2

  • The first training stage loads the model: initial_STM
  • The second training stage loads the model: resume STM and resume ATT

Put them under this structure

  MMA-Net
       |----INSTALL.md
       |----README.md
       |----dataset
       |------|-----VIL100
       |----models
       |----evaluation
       |----options.py
       |----libs
       |----requirements.txt
       |----train.py
       |----test.py

Training and Testing

To train the MMA network, run following command

python3 train.py --gpu ${GPU-IDS}

To test the MMA network, run following command

python3 test.py

The test results will be saved as indexed png file at ${root}/${output}/${valset}.

Additionally, you can modify some setting parameters in options.py to change training configuration.

Evaluation

generate accuracy, fp, fp

python evaluate_acc.py      # Please modify `pre_dir_name` and `json_dir_name` in evaluate_acc.py

Install CULane evaluation tools, please see INSTALL.md

generate F, mIoUevaluate_acc  after the CULane evaluation tools are installed

all pred txt files will be generated under MMA-Net/evaluation/txt/pred_txt after this step

python generate_iou_pred_txt.py      # Please modify `pre_dir_name` and `json_path` in  `generate_iou_pred_txt.py`

results_MMA and temp_MMA will be generated under MMA-Net/evaluation/txt/results_txt after this step.

results_MMA: evaluation results of each sequence

temp_MMA: temporary files generated during evaluation, you can ignore them

python evaluate_iou.py      # `data_root` should be set as your VIL-100 dataset path in `evaluate_iou.py`

Attention!! if you want to evaluation results one more time, please delete all folders/files under MMA-Net/evaluation/txt/results_txt .

GitHub

GitHub - yujun0-0/MMA-Net: VIL-100: A New Dataset and A Baseline Model for Video Instance Lane Detection (ICCV 2021)
VIL-100: A New Dataset and A Baseline Model for Video Instance Lane Detection (ICCV 2021) - GitHub - yujun0-0/MMA-Net: VIL-100: A New Dataset and A Baseline Model for Video Instance Lane Detection ...