This repository holds experiment code for Visual & Acoustic Discriminative Relevance (VADR) demonstrated on the UCF-101 action recognition dataset.

The model architecture used is a concatenated fusion of VGGish and C3D at the layer before the bottleneck, with a simple MLP for classification of actions.

https://www.crcv.ucf.edu/papers/UCF101_CRCV-TR-12-01.pdf

Unfortunately the original release of UCF-50 included videos without corresponding audio tracks, therefore for the visual-audio aspect of this work, we are using a subset of UCF-101 that only contains classes with audio tracks.

We use a C3D model pretrained on UCF-101 as the video subnetwork, and VGGish, trained on AudioSet for the audio subnetwork. Both architectures are based off VGG. They are concatenated at the feature embedding layer.

We use discriminative relevance (https://arxiv.org/abs/1908.01536), a modification of layerwise relevance backpropagation, to show relevant components in a four tuple consisting of .