The paper introduces the Pyramid Match Kernel, a novel kernel function designed for discriminative classification tasks involving unordered sets of image features. The kernel maps these sets to multi-resolution histograms and computes a weighted histogram intersection, which implicitly finds correspondences between feature sets based on the finest resolution histogram cell where a matched pair first appears. This approach is efficient, linear in the number of features, and robust to clutter due to its non-penalizing nature for extra features. The kernel is also positive-definite, making it suitable for use in learning algorithms that require Mercer kernels. The authors demonstrate the effectiveness of the kernel through object recognition experiments, showing that it achieves comparable accuracy to current methods but with significantly lower computational costs.The paper introduces the Pyramid Match Kernel, a novel kernel function designed for discriminative classification tasks involving unordered sets of image features. The kernel maps these sets to multi-resolution histograms and computes a weighted histogram intersection, which implicitly finds correspondences between feature sets based on the finest resolution histogram cell where a matched pair first appears. This approach is efficient, linear in the number of features, and robust to clutter due to its non-penalizing nature for extra features. The kernel is also positive-definite, making it suitable for use in learning algorithms that require Mercer kernels. The authors demonstrate the effectiveness of the kernel through object recognition experiments, showing that it achieves comparable accuracy to current methods but with significantly lower computational costs.