GLAMpoints: Greedily Learned Accurate Match Points

Abstract

We introduce a novel CNN-based feature point detector -Greedily Learned Accurate Match Points (GLAMpoints)learned in a semi-supervised manner. Our detector extracts repeatable, stable interest points with a dense coverage, specifically designed to maximize the correct matching in a specific domain, which is in contrast to conventional techniques that optimize indirect metrics. In this paper, we apply our method on challenging retinal slitlamp images, for which classical detectors yield unsatisfactory results due to low image quality and insufficient amount of low-level features. We show that GLAMpoints significantly outperforms classical detectors as well as state-of-the-art CNN-based methods in matching and registration quality for retinal images.

GLU-Net: Global-Local Universal Network for Dense Flow and Correspondences

“…Finding correspondences between a pair of images is a classical computer vision problem, uniting optical flow, geometric correspondences and semantic matching. This problem dates back several decades [20], with most classical techniques relying on hand crafted [2,3,6,19,37,39,50] or trained [12,44,62] feature detectors/descriptors, or variational formulations [4,20,38]. In recent years, CNNs have revolutionised most areas within vision, includ-ing different aspects of the image correspondence problem.…”

Self-Supervised Keypoint Detection Based on Multi-Layer Random Forest Regressor

“…Within the keypoint learning framework, the inlier-outlier sets of point pair correspondences are sampled for complete differentiation. Finally, GLAMpoints [18] extracts SIFT feature descriptors from the image I and an I warped image using previously known homography H , and finds keypoints that are matched in both directions through a brute force matcher. The keypoints found using this approach are used as ground truths to train the matching network.…”

“…Unsupervised learning-based keypoint detection has been introduced to alleviate the limitations of self-supervised learning because awareness of the predefined ground-truth homography transformations is still required. Quad-networks [19] are used to train neural networks to rank the keypoints in a transformation-invariant manner, and keypoints are detected from the upper/lower quantiles of this ranking using an unsupervised learning method [18]. In addition, with UnsuperPoint [20], keypoint detection is based on unsupervised learning that uses the regression of positions of the points to make the approach end-to-end trainable.…”

“…Unlike [26], we combined a layer-by-layer MRF regressor and SRF classifier, called KeyReg, which detects keypoints with a single forward pass, regardless of the warping of the image. KeyReg was also designed to generate training data through a self-keypoint labeling method similar to that in [18], and to detect keypoints robust to changes in the size of the images through a shared model.…”

“…Because the proposed system learns the keypoints in a self-supervised manner in the same way as in [15] and [18], a self-keypoint labeling process is required for training images. The warped image I is created by applying the i-th homography H to the unlabeled input image I .…”