Sample Distillation for Object Detection and Image Classification

We propose a novel approach to efficiently select informative samples for large-scale learn-ing. Instead of directly feeding a learning algorithm with a very large amount of samples, as it is usually done to reach state-of-the-art performance, we have developed a “distilla-tion ” procedure to recursively reduce the size of an initial training set using a criterion that ensures the maximization of the information content of the selected sub-set. We demonstrate the performance of this procedure for two different computer vision problems. First, we show that distillation can be used to improve the traditional bootstrap-ping approach to object detection. Second, we apply distillation to a classification problem with artificial distortions. We show that in both cases, using the result of a distillation process instead of a random sub-set taken uniformly in the original sample set improves performance significantly