Recurrent Affine Transform Encoder for Image Representation

This paper proposes a Recurrent Affine Transform Encoder (RATE) that can be used for image representation learning. We propose learning architecture enables CNN encoder to learn the affine transform parameter of images. The proposed decomposes an matrix into two matrices and learns them jointly in self-supervised manner. RATE is trained by unlabeled data without any ground truth infers input images recurrently. inferred represent canonical form greatly reduce variations transforms such as rotation, scaling, translation. Different from spatial transformer network, does not need embedded other networks training with aid objectives. show achieves impressive results terms invariance translation, rotation. also classification performance enhanced more robust against distortion incorporating existing model.