We propose an efficient method to learn deep local descriptors for\ninstance-level recognition. The training only requires examples of positive and\nnegative image pairs and is performed as metric learning of sum-pooled global\nimage descriptors. At inference, the local descriptors are provided by the\nactivations of internal components of the network. We demonstrate why such an\napproach learns local descriptors that work well for image similarity\nestimation with classical efficient match kernel methods. The experimental\nvalidation studies the trade-off between performance and memory requirements of\nthe state-of-the-art image search approach based on match kernels. Compared to\nexisting local descriptors, the proposed ones perform better in two\ninstance-level recognition tasks and keep memory requirements lower. We\nexperimentally show that global descriptors are not effective enough at large\nscale and that local descriptors are essential. We achieve state-of-the-art\nperformance, in some cases even with a backbone network as small as ResNet18.\n