Domain Adaptive Fisher Vector for Visual Recognition
In this paper, we consider Fisher vector in the context of domain adaptation, which has rarely been discussed by the existing domain adaptation methods. Particularly, in many real scenarios, the distributions of Fisher vectors of the training samples (i.e
- PDF / 563,808 Bytes
- 17 Pages / 439.37 x 666.142 pts Page_size
- 34 Downloads / 208 Views
Interdisciplinary Graduate School, Nanyang Technological University, Singapore, Singapore [email protected] 2 School of Computer Engineering, Nanyang Technological University, Singapore, Singapore [email protected] 3 School of Electrical and Information Engineering, University of Sydney, Sydney, Australia [email protected]
Abstract. In this paper, we consider Fisher vector in the context of domain adaptation, which has rarely been discussed by the existing domain adaptation methods. Particularly, in many real scenarios, the distributions of Fisher vectors of the training samples (i.e., source domain) and test samples (i.e., target domain) are considerably different, which may degrade the classification performance on the target domain by using the classifiers/regressors learnt based on the training samples from the source domain. To address the domain shift issue, we propose a Domain Adaptive Fisher Vector (DAFV) method, which learns a transformation matrix to select the domain invariant components of Fisher vectors and simultaneously solves a regression problem for visual recognition tasks based on the transformed features. Specifically, we employ a group lasso based regularizer on the transformation matrix to select the components of Fisher vectors, and use a regularizer based on the Maximum Mean Discrepancy (MMD) criterion to reduce the data distribution mismatch of transformed features between the source domain and the target domain. Comprehensive experiments demonstrate the effectiveness of our DAFV method on two benchmark datasets.
Keywords: Domain adaptation
1
· Fisher vector
Introduction
Constructing global feature representations based on local descriptors of images/videos is a common approach in a multitude of visual recognition tasks. As a commonly used encoding method, Fisher vector [1] encodes both first and second order statistical information of local descriptors w.r.t. the generative model (e.g., Gaussian Mixture Model (GMM)) trained based on them, and one Gaussian model in the GMM corresponds to one component in the extracted c Springer International Publishing AG 2016 B. Leibe et al. (Eds.): ECCV 2016, Part VI, LNCS 9910, pp. 550–566, 2016. DOI: 10.1007/978-3-319-46466-4 33
Domain Adaptive Fisher Vector for Visual Recognition
551
Fisher vector. Recently, Fisher vector achieves excellent performance for object recognition [2–5] or human action recognition [6,7]. To extract Fisher vector, we generally train a GMM based on the local descriptors of training samples and extract Fisher vectors for both training and test samples based on the pre-trained GMM. However, the GMM trained on the training samples does not consider the data distribution of test samples properly and thus lacks the generalization ability [8] on the test samples, leading to unsatisfactory recognition performance on the test datasets. According to the terminology in the field of domain adaptation, the training dataset and the test dataset are referred to as the source domain and the target domain, respectively. When the t
Data Loading...
