GOCor: Bringing Globally Optimized Correspondence Volumes into Your Neural Network
Paper Abstract: Authors propose GOCor, an alternative to the feature correlation layer, an essential building block in matching architectures. Our module’s correspondence volume is the result of an internal optimization procedure, which minimizes a customizable matching-objective during inference. Our robust objective integrates information about similar regions in the scene, allowing our GOCor module to disambiguate repetitive patterns better. We apply effective unrolled optimization, paired with accurate initialization, ensuring efficient end-to-end training and inference. When integrated into state-of-the-art networks, our approach significantly outperforms the feature correlation layer for geometric matching tasks, optical flow, and dense semantic matching.
Prune Truong is a Ph.D. Student at the Computer Vision Lab of ETH Zurich since May 2020. Prune received a master of Science at ETH Zurich with honors in 2019. Main research interests: image matching and alignment, particularly in the dense setting, self-supervised and unsupervised methods, deep probabilistic models.
Twitter Link: https://twitter.com/prunetruong
Martin Danelljan is a Group leader and lecturer at ETH Zurich, Switzerland. Martin received his Ph.D. degree from Linköping University, Sweden, in 2018. His Ph.D. thesis was awarded the biennial Best Nordic Thesis Prize at SCIA 2019. Main research interests: meta and online learning, deep probabilistic models, and conditional generative models. Applications to visual tracking, video object segmentation, dense correspondence estimation, and super-resolution
Twitter Link: https://twitter.com/MDanelljan
Paper Link: https://arxiv.org/pdf/2009.07823.pdf