Disparity Map Neural Operator

• 1、School of Computer,Beijing University of Posts and Telecommunications,Beijing 100876

Abstract：Training deep neural network is a common approach in stereo vision tasks. However, when applying the trained models to real-world scenarios, two prevalent issues often arise. Firstly, fixed image scale cannot meet the demands of high-resolution image prediction tasks. Secondly, due to computational constraints, existing depth estimation neural networks can only construct the Cost Volume with a fixed step size within a predefined disparity range, limiting their adaptability to scenarios requiring diverse depth perception. These challenges constrain the applicability and generalizability of the models. To address these issues, this paper proposes the Disparity Map Neural Operator (DispNO). DispNO learns a mapping from the space of stereo image functions to the space of disparity functions, enabling the generation of disparicy maps at arbitrary scales. Importantly, DispNO adapts to varying depth perception requirements in stereo vision tasks. Experimental results demonstrate that DispNO exhibits stronger applicability and performance compared to state-of-the-art methods. It provides an effective solution to the challenges of spatial resolution and depth perception adaptability in stereo vision tasks.

Keywords： computer vision disparity estimation neural operator multi-scale learning stereo matching