Robust Descriptor Learning Using Variational Auto-Encoders
Abstract: Image matching is the task of finding points in one image corresponding to the same points in the other image. Classical feature descriptors fail to match points when the images are under extreme viewpoint or seasonal changes. This thesis tackles the problem of image matching when two images are under severe changes. We propose two methods that use Variational Auto-Encoders (VAE). Variational Auto-Encoders are unsupervised generative models that encode images into a low dimensional space, called Latent Space. To improve the robustness of our methods, we train theVAE with a loss function that learns to discriminate between similar and dissimilar pairs of patches, called triplet loss. The first method, called FT-VAE, is a VAE trained with the triplet loss that creates more robust features towards rotation or seasonal changes. The second architecture, called VAE2Enc, is a novel architecture, trained in two steps, that encourages encoding rotation or seasonal changes in a small part of the latent space while creating more robust features. Empirical evaluation of FT-VAE demonstrates competitive results compared to the state of the art methods in patch pair classification.
AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)