Efficient QTMT Partitioning for VVC Intra-Frame Coding via Multi-Model Fusion

This is the source code for the paper Efficient QTMT Partitioning for VVC Intra-Frame Coding via Multi-Model Fusion

Dataset Download

• CNN dataset: https://pan.baidu.com/s/1K52HIhc6FwPvKmPzLRJQRA (code: cd5c)
• LGBM dataset: https://pan.baidu.com/s/1RQ1xkmy1t93i8i7U_LVZdQ (code: gfag)

Folder Instruction

• codec: Include the source files and exe of VTM-13.0 implemented with the proposed fast algorithm.
• encoding_test_v13: Include test scripts for the performance of the proposed fast algorithm.
• src: Include files for CNN and LGBM model training.
• test_sequences: For placing test sequences.

How to train CNN models

1. Indicate the following three directories that hold CNN datasets:

   # cnn_engin.py
   self.data_args.root_dir = './'  # root directory for storing CNN dataset
   
   # load_data_cnn.py
   img_dir = os.path.join(root_dir, 'images')  # directory for storing image data
   pickle_dir = os.path.join(root_dir, 'pickles')  # directory for storing pickles 
   

2. Setting the GPU, log path, and training parameters:

   # cnn_engine.py
   self.basic_args.date = "1210" 
   self.basic_args.log_index = '0'  
   # used to determine the path to store training data, models
   # for example: log/1210/MyNet/0/
   

3. Run cnn_engine.py

How to train LGBM models

1. Indicate the path where the LGBM dataset and models are stored

   # train_lgbm.ipynb
   pkl_dir = "parquets_lgbm"  # directory for storing parquets
   save_dir = "scripts/lgbm_scripts"  # directory for storing LGBM models
   

2. Run train_lgbm.ipynb

How to test encoding performance

You can use the script test_script.bat under encoding_test_v13 folder.