STNet: Deep Audio-Visual Fusion Network for Robust Speaker Tracking

• Python 3.6, PyTorch 1.7, opencv-python, numpy, scipy, matplotlib

• AV16.3: the original dataset, available at > http://www.glat.info/ma/av16.3/.
• CAV3D DATASET: Co-located Audio-Visual streams with 3D tracks, available at > https://speechtek.fbk.eu/cav3d-dataset/
• To construct the camera model, you need to download the camera parameters from AV16.3, including the cam.mat and rigid010203.mat.
• For training the STNet, you should prepare the audio-visual samples:
  • tools/prepareAudio.py,prepare_gccphat.py: preprocess the audio signals.
  • tools/prepareSample.py,prepareAuSample: collect image samples and corresponding audio samples.

You can download the pre-trained weigt for STNet which is used in the main paper from follow url:

• filename: model_stnet_ep50.pth (64MB), visualnet_pre.pth (8.9MB), GCFnet_pre.pth (13.7MB)
• https://drive.google.com/drive/folders/1aQ2ZiHiQcxGktMLz8omLXyi9khuNFpfM?usp=sharing

• To train STNet, you need to download seq01, 02, 03, and camera parameters from the AV16.3 dataset. Use the preprocessing files provided in the tools for audio and video synchronization, audio preprocessing, and prepare audio-visual sample pairs for training.
• After preparing the dataset and training samples, set the path of the correct image samples and GCF samples path in models/my_dataset.py and run train.py.

• SOT sequences are from seq08,11,12 in AV16.3 dataset. MOT sequences are from seq24,25,30.
• Run tracking/test_tracking_SOT/MOT.py to track.

• The stGCF and improved vgGCF method are provided in GCF/GCF_extract_stGCF,vg.py.
• Run GCF/stGCF,vgGCF.py to evluate the AO methods.

• A pre-trained Siamese network is employed to extract the visual features. The implementation of SiamFC tracker is described in the paper: Fully-Convolutional Siamese Networks for Object Tracking.
• Run visualnet/VO_MOT.py to evluate the VO methods.