TR2018-132

Active Descriptor Learning for Feature Matching



Feature descriptor extraction lies at the core of many computer vision tasks including image retrieval and registration. In this paper, we present an active learning method for extracting efficient features to be used in matching image patches. We train a Siamese deep neural network by optimizing a triplet loss function. We develop a more efficient and faster training procedure compared to the state-of-the-art methods by increasing difficulty during batch training. We achieve this by adjusting the margin in the loss and picking harder samples over time. The experiments are carried out on Photo Tourism dataset. The results show a significant improvement on matching performance and faster convergence in training.