One Simple Stage learning for Feature Detection, Description and Matching

The task of feature matching plays a fundamental role in visual-based applications. However, conventional methods tend to deal with the learning process of feature detection, feature description, and matching in an independent manner, by which the end-to-end fine-tuning in other downstream tasks can be hardly guaranteed. To address this issue, we propose a simple learning approach for feature detection, description, and matching within single stage. This approach reveals the latent key-points distribution from a given feature map and prior probability is sampled by differential method for training feature detection description and matching simultaneously. Compared to other learning-based methods, our method is capable of learning feature without pre-computed positive-negative pairs and well designed feature representation. The developed approach can ensure feasible integration into any model containing feature detection, description module, and differentiable matching layer.