Learning to Detect Local Features Using Information Change

In this paper we propose that features extracted from deep convolutional neural networks have the structure and information necessary to detect location and scale of the local keypoints. Unlike the previous supervised and unsupervised methods, we define a local feature as an outcome of information change across different receptive fields around an image region. Exploiting the existent representation hierarchy in the deep convectional neural network, we propose a trainable information accumulation pyramid that allows us to relate the change in the receptive field with information change. The network is trained in an unsupervised fashion by applying random set of transformations over the images and minimizing the covariant loss. We demonstrate the efficacy of our proposed keypoint extractor by evaluating its performance for repeatability and matching scores. Our approach results in 3.7% and 5.2% higher over the state-of-the-art algorithms in repeatability and matching score, respectively.