Axial Constraints for Global Matching-based Optical Flow Estimation

Optical flow estimation is a fundamental task that aims to find the 2-dimensional motion field by identifying correspondences between two input images. For quite long time, correlation volume followed convolutional neural networks (CNN) directly estimates optical was predominant pipeline. However, several pioneering methods proposed global matching recently, pointing out limitation CNN-based are struggling handle large displacements due their locality. Global step identifies at pixel-level using entire volumes once with simple operations like softmax. when softmax combined commonly used regression loss in estimation, there will be vast number of possible can minimize and correctly estimate correspondences. In other words, training objective induces one-to-many solution problem resulting presence noisy gradients. this paper, necessity for more constraints on mitigate aforementioned ill-posed discussed. To acquire such constraints, axial cross-entropy (i.e. constraints) restrict have low variance designed pseudo ground truth proposed. Experimental results show axial-constraints applicable off-the-shelves matching-based frameworks easily lead both quantitative qualitative performance improvement without any architectural changes.