SCNN Tensorflow2实现版本

SCNN implemented by Tensorflow2

Paper Link:"Spatial As Deep: Spatial CNN for Traffic Scene Understanding", AAAI2018

Source Code: "https://github.com/XingangPan/SCNN"

1. Clone the project

   git clone https://github.com/wind754203900/SCNN-TF2
   cd SCNN-TF2
   

2. Create a conda virtual environment and activate it

   conda create -n scnn_tf2 python=3.7 -y
   conda activate scnn_tf2
   

   Then install dependencies

   pip install -r requirements.txt
   

3. Data preparation

   Download Tusimple. Then extract to$TUSIMPLEROOT. The directory arrangement of Tusimple should look like:

   $TUSIMPLEROOT
   |-train_set
     |──clips
     |──label_data_0313.json
     |──label_data_0531.json
     |──label_data_0601.json
     |──readme.md
   |-test_set
     |──clips
     |──test_tasks_0627.json
     |──test_label.json
     |──readme.md
   

   Since the segmentation annotation is not provided for Tusimple, please generate seg segmentation from the json annotation.

   cd data_provider
   python tusimple_processing.py   # modify variable of 'src_dir','dst_dir' and 'test_dir' in python file
   

   After running. You will get

   $TUSIMPLEROOT
   |-train_set
     |──...
     |──training
        |──train_instance.txt
        |──train_binary.txt
        |──validation_instance.txt
        |──validation_binary.txt
        |──gt_image
           |──....png
        |──gt_instance_image
           |──....png
        |──gt_binary_image
           |──....png
   |-test_set
     |──...
     |──test.txt
   

1. Train your model

   First, modify config file in global_config/config.py

   Change TU_DATASETS_TRAIN and TU_DATASETS_VALID to the path where your tusimple train and validation annotaion txt files store in

   # config file in global_config/config.py
   __C.TU_DATASETS_TRAIN = '{your_generated_tusimple_dataset_path}/training/train_instance.txt'
   __C.TU_DATASETS_VALID = '{your_generated_tusimple_dataset_path}/training/validation_instance.txt'
   

   OPTIONAL:

   You can change some training setting about epoches,learning rate and so on in config file

   Start training with: (You also can run the code in Pycharm)

   cd tools
   python train_deeplab_distribute.py
   

   Since I use tf.distribute.MirroredStrategy() in the code, the code will use one gpu or multi-gpus automatical automatically.

   After training, the weights file will be save in weights/{model_name}.h5

   You can also modify the model save path by editing global_config/config.py

   # config file in global_config/config.py
   __C.TRAIN.MODEL_SAVE_PATH = '{weigth_save_path_of_your_model}'.
   

2. Evaluation

   Before evaluation, you also should modify global_config/config.py to your tusimple test_dataset.

   # config file in global_config/config.py
   __C.TEST.DATAROOT = '{your_tusimple_test_dataset_path}/test_set'
   __C.TEST.ANNO_PATH = '{your_tusimple_test_dataset_path}/test_set/test.txt'
   

   If you want to visulize the result, please change the value CFG.TEST.VISUALIZE to True

   # config file in global_config/config.py
   __C.TEST.VISUALIZE = True
   

   Evaluate by(Modify your ground truth json path gt_json_path in evaluate_to_json.py)

   cd tools
   python evaluate_to_json.py
   

   It will generate a josn file in evaluation/evaluate_lane.json.

The support for CULane Dataset will be implemented in the future.