Paper Title
Exploration of Deterministic Sampling for Convolutional Neural Networks Tracker for Single Object Tracking

Abstract
This paper presents a deterministic approach to train and update object appearance model for a convolutional neural networks (CNNs)tracker. A network of three convolutional and three fully connected layers is used to model the object of interest appearance based on solely the initialization frame information. The proposed sampling method is only given the ground truth location of the object on the first frame. Positive training samples are extracted by translating the ground truth with varying step size in eight directions. Meanwhile, negative training samples are extracted from the eight neighborhood regions of equal size that surround the foreground object. Then, the output sample is selected by combining the highest samples based on combine score of its appearance similarity. The experimental test shown that the proposed approach can provide an alternative solution to train and update the object appearance model using CNNs tracker. Index Terms - Convolutional neural network, Deep learning, Deterministic sampling, Visual object tracking.