An Active Search Strategy for Efficient Object Class Detection
Place: Large lecture room.
Affiliation: University of Edinburgh.
Modern object class detectors typically apply a window classifier to all the windows in a large set, either in a sliding window manner or using object proposals. In this work, we propose an active search strategy that sequentially chooses the next window to evaluate based on previously observed windows, rather than going through the whole window set in an arbitrary order. This results in a substantial reduction in the number of classifier evaluations and in a more elegant approach in general, avoiding wasteful computation in uninteresting areas and focusing on the promising ones. Our search strategy is guided by two forces. First, we exploit context as the statistical relation between the appearance of a window and its location relative to the object, as observed in a training set. For example, the method can learn that cars tend to be on roads below the sky. Therefore, observing a window in the sky in a test image suggests the car is likely to be far below, whereas a window on the road suggests making a smaller horizontal move. We learn the context force in a Random Forest framework that provides great computational efficiency as well as accurate results. Second, we exploit the score of the classifier to attract the search to promising areas surrounding a highly scored window, and to keep away from areas near low scored ones. In experiments on the challenging SUN 2012 dataset using state-of-the-art detector R-CNN, our method matches the detection accuracy of evaluating all windows independently, while evaluating 9x fewer windows. As our method adds little overhead, this translates into an actual wall-clock speed-up.