Exploring pseudo-labels for domain adaptation in egocentric video

Student thesis: Master's ThesisMaster of Science by Research (MScR)

Abstract

The task of unsupervised domain adaptation offers an effective way to assess the generalisability of models. It incorporates the problems of unlabelled data and domain shift, which are common in real-world settings due data collected from different people, locations and equipment. Pseudo-labelling approaches to this problem attempt to annotate the unlabelled data using the knowledge from another labelled distribution, so it can be used to train a model which adapts well across both distributions. These methodologies show great promise as both complimentary and stand-alone techniques. By exploring this approach in two different ways for egocentric data, this work shows that noise reduction of pseudo-labels does not always lead to better adaptation and the methodology as a whole can struggle on larger and more complex datasets, with numerous classes.

A popular training regime which randomly samples clips from videos is confirmed in this work to cause variations in the feature representations and class predictions across the same video. More sophisticated pseudo-labelling methods are then proposed to avoid harmful segments of videos and to exploit more discriminative ones. Although some of these methods reduce pseudo label noise much better than random sampling, they cause a degradation of the model during adaptation. Heavier focus on particular video regions causes a bias towards the largest class and reduces the diversity of samples seen by the model.

The efficacy of pseudo-labelling is further analysed in comparison to a more complex dataset. Some success is possible, but adaptation difficulty shows that the pseudo-label class distribution matters more when the dataset has a large set of classes. The work shows that when the pseudo-label class distribution does not match that of the test class distribution very well, the methodology can struggle considerably. This points to a difficult problem for pseudo-labelling strategies on the unsupervised domain adaptation task.
Date of Award5 Dec 2023
Original languageEnglish
Awarding Institution
  • University of Bristol
SupervisorNello Cristianini (Supervisor) & Dima Damen (Supervisor)

Cite this

'