DOMAIN ADAPTIVE OBJECT RECOGNITION AND DETECTION

Abstract

Discriminative learning algorithms rely on the assumption that training and test data are drawn from the same marginal probability distribution. In real world applications, however, this assumption is often violated and results in a significant performance drop. We often have sufficient labeled training data from single or multiple "source" domains but wish to learn a classifier which performs well on a "target" domain with a different distribution and no labeled training data. In visual object detection, for example, where the goal is to locate the objects of interest in a given image, it may be infeasible to collect training data to model the enormous variety of possible combinations of pose, background, resolution, and lighting conditions affecting object appearance. Thus, we generally expect to encounter instances or domains at test time for which we have seen little or no training data.

To this end, we first propose a framework for domain adaptive object recognition and detection using Transfer Component Analysis, an unsupervised domain adaptation and dimensionality reduction technique. The idea is to obtain a transformation in feature space to a latent subspace that reduces the distance between the source and target data distributions. We evaluate the effectiveness of this approach for vehicle detection using video frames from 50 different surveillance cameras.

Next, we explore the problem of extreme class imbalance present when performing fully unsupervised domain adaptation for object detection. The main challenge arises from the fact that images in unconstrained settings are mostly occupied by the background (negative class). Therefore, random sampling will not be effective in obtaining a sufficient number of positive samples from the target domain, which is required by any adaptation method. We propose a variation of co-learning technique that automatically constructs a more balanced set of samples from the target domain. We compare the performance of our technique with other approaches such as unbiased learning from multiple datasets and self-learning.

Finally, we propose a novel approach for unsupervised domain adaptation. Our method learns a set of binary attributes for classification that captures the structural information of the data distribution in the target domain itself. The key insight is finding attributes that are discriminative across categories and predictable across domains. We formulate our optimization problem to learn these attributes and the classifier jointly. We evaluate the performance of our method on a wide range of tasks including cross-domain object recognition and sentiment analysis on textual data both in inductive and transductive settings. We achieve a performance that significantly exceeds the state-of-the-art results on standard benchmarks. In many cases we reach the same-domain performance, the upper bound, in unsupervised domain adaptation scenarios.