Learning Multi-Domain Convolutional Neural Networks for Visual Tracking

Abstract

We propose a novel visual tracking algorithm based on a discriminatively trained Convolutional Neural Network (CNN). In order to overcome the limitation of the insufficient training data in visual tracking problems, we pretrains a CNN using a large set of videos with various domains to learn generic target representations. Our network consists of shared layers and multiple branches of domain-specific layers, where each branch is responsible for the binary classification of target and background in each domain. We train the network for each domain iteratively to obtain generic target representations in the shared layers. When tracking a target in a new sequence, we construct a new network by combining the shared layers in the pretrained CNN with a new binary classification layer, which is updated online. Online tracking is then performed by evaluating the candidate windows randomly sampled around the previous target state. The proposed algorithm illustrates outstanding performance in public visual tracking benchmarks.