LoCoNet

LoCoNet: Long-Short Context Network for Active Speaker Detection (2023 CVPR)

Code Usage

1. Download the audio-visual features and the annotation csv files from Google Drive. The directories should look like as follows:

|-- features
    |-- resnet18-tsm-aug
        |-- train_forward
        |-- val_forward
    |-- resnet50-tsm-aug
        |-- train_forward
        |-- val_forward
|-- csv_files
    |-- ava_activespeaker_train.csv
    |-- ava_activespeaker_val.csv

2. Training / Validation

bash run.sh conf/LoCoNet/ResNet18/large [gpu_id] 1

Note

• Code for training only LoCoNet backend with pre-trained feature provided by SPELL.