ADLE: adaptive distribution likelihood estimation for stereo matching
Shichao Wang, Mingxing Jia, Shijie Chang, Zongying Yu · Measurement Science and Technology · 2025
Abstract Stereo matching technology estimates the depth information of two images by their disparity, thereby obtaining the 3D information of the object. Stereo matching technology based on deep learning is also constantly developing. However, researchers have primarily focused on designing specific, highly optimized and customized structures to enhance performance. This approach often involves incorporating numerous priors, which can potentially limit flexibility of the model. Consequently, the same model may require adjustment of numerous parameters across different tasks and scenarios to achieve optimal outcomes. In this paper, we propose a novel approach to re-establish the original regression process from the perspective of uncertainty analysis. Our strategy involves generating and adjusting an implicit error distribution, which helps mitigate the adverse effects of inappropriate prior error distributions on model optimization, and does not require conducting some experiments in advance to obtain the necessary prior knowledge to guide the design of the loss function. Simultaneously, only two lightweight fully connected layers are used to achieve reversible mapping of implicit distributions, achieving efficient end-to-end learning. Our experimental results indicate that this method improves performance of stereo matching network, outperforming many other advanced works. The endpoint error on the SceneFlow test dataset was reduced by 0.1 pixels, while that on the Middlebury 2014 test dataset was diminished by 0.2 pixels. Moreover, the method achieved excellent performance on both the KITTI 2012 and KITTI 2015 benchmarks.