Multi-scale representation based on convolutional neural networks for tracking

Abstract

Visual Tracking technology is one of the major branches in computer vision. Although it has been studied for many years, there are still a number of challenges need to be overcome. In this paper, we propose a tracking algorithm based on multi-scale convolutional neural networks trained on large amounts of tracking sequence data with ground-truth bounding targets. Instead using the raw pixels to feed to the models, we use the image gradient to learn the object representation. We implement this by generating multiple scale version images from Laplacian pyramid, and we maintain a pool of networks corresponding to each kind of video for each scale and utilize the VGG-net to pre-train our models. From the models, we can extract multi-scale feature representations to encode the appearance. In addition, we improved the multiple instance learning tracking algorithm by introduce a penalty factor in the sigmoid function to solve the saturation problem. Using the multi-scale feature representations, we train a classifier combined with the improved MIL algorithm. The results comparing with several state-of-the-art methods on challenging sequences have proved the effectiveness of our proposed algorithm.