Rethinking Audio-visual Synchronization for Active Speaker Detection

Usage

dataset

The preparation of AVA-Activespeaker dataset follows https://github.com/TaoRuijie/TalkNet_ASD

Training

python trainTalkNet.py --dataPath /path/to/data/

Eval

python trainTalkNet.py --dataPath /path/to/data/ --evaluation True --evaluation_pth /path/to/model

Unsynchronization eval

python trainTalkNet.py --dataPath /path/to/data/ --evaluation True --evaluation_pth /path/to/model --syncDestroy audioSwap --audio_change_ratio 0.5

Reference

https://github.com/TaoRuijie/TalkNet_ASD

https://github.com/clovaai/lookwhostalking