KernelBoost: Supervised Learning of Image Features For Classification

We propose a fully-supervised approach to training classifiers that automatically learn features directly from image data. This drops the dependency on hand-designed filters and features, which is generally a trial-and-error process and often yields far-from-optimal results. Our approach relies on the Gradient Boosting framework, learning discriminative features at each stage in the form of convolutional filters. It depends on just few easy-to-tune parameters, it is simple and general, and we show it outperforms state-of-the-art methods in tasks ranging from pixel classification in very different types of images to object detection.