Heterogeneous domain adaptation using previously learned classifier for object detection problem

When a trained classifier on specific domain (source domain) is applied in a different domain (target domain) the accuracy is degraded significantly. The main reason for this degradation is the distribution difference between the source and target domains. Domain adaptation aims to lessen this accuracy degradation. In this paper, we focus on adaptation for heterogeneous domains (where the source and target domain may have different feature spaces) and propose a novel algorithm which uses the pre-learned source classifier to adapt a trained target classifier. In this method, a max-margin classifier is trained on the target data and is adapted using the offset of the source classifier. The main strength of this adaptation is its low complexity and high speed which makes it a proper adaptation choice for problems with large-size datasets such as object detection. We test our method on human detection datasets and the experimental results show the significant improvement in accuracy, in comparison to several baselines.