Naver at ActivityNet Challenge 2019 -- Task B Active Speaker Detection (AVA)

TL;DR

This paper presents a novel approach for active speaker detection using a 3D CNN front-end combined with ensemble classifiers, achieving significant improvements on the AVA-ActiveSpeaker dataset.

Contribution

The paper introduces a new deep learning framework combining 3D CNNs and ensemble classifiers for active speaker detection, outperforming previous baselines.

Findings

Significant accuracy improvements over baseline methods

Effective use of 3D CNNs for temporal feature extraction

Ensemble of temporal convolution and LSTM classifiers enhances performance

Abstract

This report describes our submission to the ActivityNet Challenge at CVPR 2019. We use a 3D convolutional neural network (CNN) based front-end and an ensemble of temporal convolution and LSTM classifiers to predict whether a visible person is speaking or not. Our results show significant improvements over the baseline on the AVA-ActiveSpeaker dataset.