Computer Vision News 20 Congrats, Doctor Prune! Establishing robust and accurate correspondences between a pair of images is a long-standing computer vision problem with numerous applications, such as structure-frommotion, image registration, or image manipulation. While classically dominated by sparse methods which find matches for salient keypoints, emerging dense approaches offer a compelling alternative paradigm. Given a pair of images, the goal of dense matching methods is to predict a match for every pixel within the images. It is commonly achieved by predicting the flow field, encapsulating relative displacements relating one image to the other. Compared to sparse approaches, dense methods avoid the keypoint detection step, which is the main failure point of sparse approaches. Moreover, instead of solely relying on descriptor similarities to establish matches, they additionally learn to leverage, e.g. local motion patterns and smoothness priors. Prune Truong recently obtained her PhD from the Computer Vision Lab of ETH Zurich. Her thesis focused on dense matching from limited supervision and its applications. She is now a research scientist at Google. Congrats, Doctor Prune! How dense matching works
RkJQdWJsaXNoZXIy NTc3NzU=