Hybrid Skip: A Biologically Inspired Skip Connection for the UNet Architecture

In this work we introduce a biologically inspired long-range skip connection for the UNet architecture that relies on perceptual illusion of hybrid images, being images simultaneously encode two images. The fusion early encoder features with deeper decoder ones allows models to produce finer-grained dense predictions. While proven in segmentation tasks, network’s benefits are down-weighted regression tasks as these connections additionally result texture transfer artifacts. Specifically depth estimation, hurts smoothness and introduces false positive edges which detrimental task due maps’ piece-wise smooth nature. proposed HybridSkip show improved performance balancing trade-off between edge preservation, minimization artifacts hurt smoothness. This is achieved by proper balanced exchange information offer high low frequency, features, respectively. code will be made available https://vcl3d.github.io/HybridSkip/.