CCoL

Pytorch implementation for CVPR21 "Cyclic Co-Learning of Sounding Object Visual Grounding and Sound Separation".

Unlike most of previous works, we assume that there are potential silent objects in videos during training. We perform sounding object-aware separation during training to alleviate the silent object issue. With a cyclic co-learning framework, our model can simultaneously address sounding object visual grounding and audio-visual separation tasks.

Environment

The code is developed under the following configurations.

• Hardware: 1-4 GPUs (change [--num_gpus NUM_GPUS] accordingly)
• Software: Ubuntu 16.04.3 LTS, CUDA>=8.0, Python>=3.5, PyTorch>=1.2

Training

The frames and detection results used in the paper can be downloaded from this link. If you want to use a new dataset, please follow the following steps.

1. Prepare video dataset.

   a. Download MUSIC dataset from: https://github.com/roudimit/MUSIC_dataset/blob/master/MUSIC_solo_videos.json

   b. Download videos.

2. Preprocess videos. You can do it in your own way as long as the index files are similar.

   a. Extract frames at 1fps and waveforms at 11025Hz from videos. We have following directory structure (please first ignore the detection results):

   data
   ├── audio
   |   ├── acoustic_guitar
   │   |   ├── M3dekVSwNjY.wav
   │   |   ├── ...
   │   ├── trumpet
   │   |   ├── STKXyBGSGyE.wav
   │   |   ├── ...
   │   ├── ...
   |
   └── frames
   |   ├── acoustic_guitar
   │   |   ├── M3dekVSwNjY.mp4
   │   |   |   ├── 000001.jpg
   │   |   |   ├── ...
   │   |   ├── ...
   │   ├── trumpet
   │   |   ├── STKXyBGSGyE.mp4
   │   |   |   ├── 000001.jpg
   │   |   |   ├── ...
   │   |   ├── ...
   │   ├── ...
   |
   └── detection_results
   |   ├── acoustic_guitar
   │   |   ├── M3dekVSwNjY.mp4.npy
   │   |   ├── ...
   │   ├── trumpet
   │   |   ├── STKXyBGSGyE.mp4.npy
   │   |   ├── ...
   │   ├── ...
   

   b. We created the index files train.csv/val.csv/test_sep.csv. And the first two files are with the following format:

   /acoustic_guitar/M3dekVSwNjY.mp3, /acoustic_guitar/M3dekVSwNjY.mp4,1580
   /trumpet/STKXyBGSGyE.mp3, /trumpet/STKXyBGSGyE.mp4,493
   

   For each row, it stores the information: AUDIO_PATH,FRAMES_PATH,NUMBER_FRAMES

   c. Detect objects in video frames. We used object detector trained by Ruohan used in his Cosep project (see CoSep repo). The detected objects for each video are stored in a .npy file.

3. Train the grounding-only model for warming up

./scripts/train_grd.sh

4. Train the co-learning model

./scripts/train_col.sh

3. Train the cyclic co-learning model. Note that the CCoL model performance can be first improved with finetuning (setting grounding lr as 1e-5 and separation lr as 1e-4).

./scripts/train_ccol.sh

4. During training, visualizations are saved in HTML format under data/ckpt/MODEL_ID/visualization/.

Evaluation

1. Evaluate the trained joint grounding and separation model. Our pre-trained model can be downloaded from here. Please put it into data/ckpt.

./scripts/eval_ccol.sh

2. Evaluate the sounding object-aware separation model on videos with silent objects.

./scripts/eval_ccol_silent.sh

Reference

If you use the code from the project, please cite:

    @inproceedings{tian2021cyclic,
     title={Cyclic Co-Learning of Sounding Object Visual Grounding and Sound Separation},
     author={Tian, Yapeng and Hu, Di and Xu, Chenliang},
     booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
     pages={2745--2754},
     year={2021}
    }

Acknowledgement

We borrowed a lot of code from SoP and used detector from Ruohan' CoSep. We thank the authors for sharing their code. If you use our codes, please also cite their nice works.