Year: 2026
Type: article
Abstract: Bijective correspondence is essential for texture and signal transfer, serving as a core benchmark for shape matching quality. Although the recently proposed Neural Adjoint Map (NAM) advances 3D shape correspondence by generalizing linear spectral alignment to nonlinear, network‐parameterized functions, it is inherently unidirectional, failing to guarantee map bijectivity; on the other hand, existing spectral embedding alignment methods that promote bijectivity rely on the reciprocal property of low‐dimensional functional maps, which we point out potentially introduces an isometric mapping prior and degrades accuracy under non‐isometric scenarios. To address these issues, we propose the Bijective Neural Adjoint Map (BNAM), a novel framework that bridges bidirectional NAMs under a soft invertibility constraint to encourage mutual bijectivity. Additionally, we introduce an iterative upsampling pipeline to efficiently estimate BNAM. Our method promotes map bijectivity while preserving mapping accuracy across challenging scenarios, including near‐/non‐isometric meshes and unstructured point clouds.
Source: Computer Graphics Forum
DOI: 10.1111/cgf.70620
Authors: Jianhui Liu, Zhongke Wu, Xingce Wang
Institution: Beijing Normal University
Language: English
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
Awards: G5173118459, G9001369272
OpenAlex ID: W7233176363
