Egocentric Audio Visual Speaker Localization

The simulated dataset Ego-AVST is available on request

Experiments on Ego-AVST

1. Split the dataset into chunks so that every chunk in a subfolder containing image, audio, gt, bbox, depth image

   cd data
   Using the function split in split_avst.py

2. Calculate GCCPHAT for each audio clip in each chunk

   Using the function test in gccphat_avst.py

3. Train

   python main_avst.py

4. Evaluation

   Put the path of pretrained model in args.load_path in main_avst.py

   Comment the training code
   trainer.fit(model, data_module)

   Uncomment the code block
   checkpoint = torch.load(args.load_path)
model.load_state_dict(checkpoint['state_dict'])
model.eval()

Experiments on EasyCom

1. Split the data into training set and test set. Session 1, 2, 3 will be in the test set and the remaining sessions will be in the training set. The whole video list can be found in video.txt under the EasyCom folder.

2. Split the dataset into chunks so that every chunk in a subfolder containing image, audio, gt

   cd data
   Using the function split in split.py

3. Calculate GCCPHAT for each audio clip in each chunk

   Using the function test in gccphat.py

4. Train

   python main.py

5. Evaluation

   Put the path of pretrained model in args.load_path in main.py

   Comment the training code
   trainer.fit(model, data_module)

   Uncomment the code block
   checkpoint = torch.load(args.load_path)
model.load_state_dict(checkpoint['state_dict'])
model.eval()

Reference

EasyCom Dataset https://github.com/facebookresearch/EasyComDataset
Transformer Block: https://github.com/jadore801120/attention-is-all-you-need-pytorch
GCCPHAT Calculation: https://github.com/yinkalario/Two-Stage-Polyphonic-Sound-Event-Detection-and-Localization
Lightning: https://github.com/miracleyoo/pytorch-lightning-template