Working hard to know your neighbor's margins:Local descriptor learning loss

摘要: We introduce a novel loss for learning local feature descriptors which is inspired by the Lowe's matching criterion SIFT. show that proposed maximizes distance between closest positive and negative patch in batch better than complex regularization methods; it works well both shallow deep convolution network architectures. Applying to L2Net CNN architecture results compact descriptor -- has same dimensionality as SIFT (128) shows state-of-art performance wide baseline stereo, verification instance retrieval benchmarks. It fast, computing takes about 1 millisecond on low-end GPU.