Transductive transfer learning for computer vision
10th July, 2015
Speaker: Teofilo de Campos, senior research fellow at University of Surrey, Guildford, U.K
Abstract: One of the ultimate goals of open ended learning systems is to take advantage of previous experience in dealing with future problems. We focus on classification problems where labelled samples are available in a known problem (the source domain), but when the system is deployed in the target dataset, the distribution of samples is different. Although the number of classes and the feature extraction method remain the same, a change of domain happens because there is a difference between the typical distribution of data of source and target samples. This is a very common situation in computer vision applications, e.g., when a synthetic dataset is used for training but the system is applied to images "in the wild". We assume that a set of unlabelled samples is available from the target domain. This constitutes a Transductive Transfer Learning problem, also known as Unsupervised Domain Adaptation. We propose to tackle this problem by adapting the feature space of the source domain samples, so that their distribution becomes more similar to that of the target domain samples. Therefore a classifier re-trained on the updated source space can give better results on the target samples. We propose to use a pipeline which consists of three main components:
- a method for global adaptation of the marginal distribution of the data using Maximum Mean Discrepancy;
- a sample-based adaptation method, which translates each source sample towards the distribution of the target samples;
- a class-based conditional distribution adaptation method.
We conducted experiments on a range of image classification and action recognition datasets and showed that our method gives state-of-the-art results.
I will also give an overview of the work I have been doing recently on image matching and on vision tools for a project on spatial audio.