Rotation-Equivariant Keypoint Detection
We show how to train a rotation-equivariant representation to extract local keypoints for image matching. Existing learning-based methods focused on extracting translation-equivariant keypoints using conventional convolutional neural networks (CNNs), but rotation-equivariant keypoint detectors have not been studied extensively. Therefore, we propose a rotation-invariant keypoint detection method using rotation-equivariant CNNs. Our rotation-equivariant representation enables us to estimate local orientations to image keypoints accurately. We propose a dense histogram alignment loss to assign an orientation to keypoints more consistently. We validate the effectiveness compared to existing keypoint detection methods. Furthermore, we check the transferability of our method on public image matching benchmarks.
PDF Abstract