Core Concepts
学会で発表されたGenCorresは、新しい教師なしの共同形状マッチングアプローチであり、一貫した形状対応を実現します。
Abstract
Introduction:
GenCorres introduces a novel unsupervised joint shape matching approach.
Key idea is to learn a mesh generator for consistent shape matching among synthetic shapes.
Overcomes challenges of existing methods by leveraging data-driven power and unifying consistent and pairwise matching.
Problem Statement and Approach Overview:
Input shape collection S used to learn mesh generator mθ.
Three stages: Implicit Shape Generator, Mesh Generator Initialization, Mesh Generator Refinement.
Experimental Evaluation:
Evaluated on Human and Animal deformable shape collections.
Outperforms state-of-the-art JSM techniques and implicit/point cloud shape generators.
Ablation study shows importance of geometric deformation regularization and cycle-consistency.
Related Work:
Discusses relevant work in pairwise shape matching, generative model-based correspondences, matching under implicit surfaces, neural implicit representations.
Stats
GenCorresはICLR 2024で発表されました。
GenCorresは他のJSM技術よりも優れたパフォーマンスを示します。
Quotes
"GenCorres performs JSM among synthetic shapes, whose size is much larger than the number of input shapes."
"Experimental results show that GenCorres considerably outperforms state-of-the-art JSM techniques."