Improved weighted multiple instance learning for object tracking

Visual tracking usually requires an object appearance model that is robust to changing illumination, partial occlusion, large pose and other factors encountered in video. Currently, a technique called “tracking by detection” has been developed and studied with promising results. A typical tracking by detection algorithm called MIL (Multiple Instance Learning) has become one of the most popular methods in tracking domain. This technique is designed to alleviate the drift problem by using an MIL based appearance model to represent training data in the form of bags. In this paper, we improved the WMIL tracker in two aspects. First, we propose a new bag model that integrates the importance of the samples in the positive bag naturally with the distribution field of the image patches as a part of the weighting function. Then, in order to solve the potential overfitting problem, we propose a dynamic function to estimate probability of instance by introducing extra parameters instead of the original logistic function. Experiments on some publicly available benchmarks of video sequences demonstrate the effectiveness and robustness of our approach.