Published Paper Details:

Moving object detection and tracking field gains a vast interest in it. Most of the moving objects detection methods operate by trying a binary classifier to sub-windows of an image, after that a non-maximum suppression step comes where detections on overlapping sub-windows are eliminated. As the number of feasible sub-windows in even middling sized image datasets is very large, the classifier is usually learned on only a subset of the windows. This circumvent the computational complexity and difficulty of dealing with the entire set of sub-windows , however, this paper address that, it leads to sub-optimal detector performance. Specially , the main offering of this paper is the initiation of a new method, Max-Margin object detection (MMOD), for understanding how to detect objects in images. This method does not execute any sub sampling, but instead optimizes overall sub-windows. MMOD is used to enhance any object detection method which is linear in the learned parameters, such as Histogram of Oriented Gradient (HOG) or bag-of-visual-word models. Using this strategy we manifest substantial performance gains in publicly available data sets. We represent that a single firm HOG filter can overcome a set-of-the-art deformable part model on the object detection data set and benchmark when the HOG filter is assimilated via MMOD.