Virtual Adversarial Training in Feature Space to Improve Unsupervised Video Domain Adaptation

TL;DR

This paper introduces a novel approach applying Virtual Adversarial Training directly in feature space to enhance unsupervised video domain adaptation, addressing instability issues and improving performance over existing methods.

Contribution

It proposes applying Virtual Adversarial Training in feature space and introduces substitutes for unstable entropy minimization techniques in domain adaptation.

Findings

Achieves competitive or superior results on multiple video domain adaptation benchmarks.

Demonstrates stability improvements over previous entropy minimization methods.

Enhances the state-of-the-art TA$^3$N model with new techniques.

Abstract

Virtual Adversarial Training has recently seen a lot of success in semi-supervised learning, as well as unsupervised Domain Adaptation. However, so far it has been used on input samples in the pixel space, whereas we propose to apply it directly to feature vectors. We also discuss the unstable behaviour of entropy minimization and Decision-Boundary Iterative Refinement Training With a Teacher in Domain Adaptation, and suggest substitutes that achieve similar behaviour. By adding the aforementioned techniques to the state of the art model TA$^3$N, we either maintain competitive results or outperform prior art in multiple unsupervised video Domain Adaptation tasks