Learning Transformation-Invariant Local Descriptors With Low-Coupling Binary Codes

Despite the great success achieved by prevailing binary local descriptors, they are still suffering from two problems: 1) vulnerable to geometric transformations; 2) lack of an effective treatment highly-correlated bits that generated directly applying scheme image hashing. To tackle both limitations, we propose unsupervised Transformation-invariant Binary Local Descriptor learning method (TBLD). Specifically, transformation invariance descriptors is ensured projecting original patches and their transformed counterparts into identical high-dimensional feature space low-dimensional descriptor simultaneously. Meanwhile, it enforces dissimilar have distinctive descriptors. Moreover, reduce high correlations between bits, a bottom-up strategy, termed Adversarial Constraint Module, where low-coupling codes introduced externally guide With aid Wasserstein loss, framework optimized encourage distribution mimic codes, eventually making former more low-coupling. Experimental results on three benchmark datasets well demonstrate superiority proposed over state-of-the-art methods. The project page available at https://github.com/yoqim/TBLD.