We propose an approach to reduce the runtime of Exemplar-SVM when multiple detectors are to be applied to the same image. Our underlying assumption is that image windows of similar appearance will yield similar detection responses. Therefore, rather than applying all the detectors to all the candidate image windows, we compute the detector scores only once, for a finite pool of image windows. The detection score for each candidate is constructed from these values, and serves as an approximation to the actual score. Avoiding the computation of numerous SVMs reduces the number of calculations and therefore significantly reduces the overall runtime. Experiments on the PASCAL VOC detection task show that this approximation maintains similar detection accuracy as the original Exemplar-SVM.
MSc Thesis work under supervision of Prof. Lihi Zelnik-Manor.
Note: the talk will be given in Hebrew.