Abstract:We propose a one-step Local Feature Extraction Network framework to solve the sparse feature matching problem. In our network, we use raw camera data and the Structure from Motion (SfM) algorithm to restore the corresponding relationships of the different feature map. Our network combines the detector and descriptor as one step to build an end-to-end Local Feature Extraction network. At the same time, the whole process is differentiable and we train our network by the loss of feature map. Finally, we train our network on indoor datasets and prove its accuracy and rapidity advantage over other methods.

One-step Local Feature Extraction Using CNN