Active sampling exploiting detector response pattern for efficient target detection

Abstract

Detecting targets in an image is a fundamental task in computer vision and robotic system. When only a trained detector (binary classifier) is at hand, the target detection problem becomes localizing the correct windows containing targets in the image and can be considered as a sampling problem. Exhaustive sliding window method is a common approach, but it is usually computationally expensive especially when the detection algorithm is time-consuming. In this work, we observe that detector's response scores of sampling windows fade gradually from the peak response window in the detection area and we approximate this scoring pattern with an exponential decay function. By exploiting this property, we propose an active sampling method for efficient target detection to avoid exhaustively searching all the window space. The method estimates the probability of windows containing the target by fusing information from sampled windows and their detector's scores and then decides the next window to be observed. Experiments have shown that our proposed method improves efficiency in human detection applications as it requires fewer windows to achieve similar performance compared to sliding windows and multi-stage particle window (MS-PW) method.