MDND: Unsupervised Learning Guided by Non-Differentiable Refinement for Shape Correspondence

2026-07-17 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

MDND is a novel Deep Functional Map (DFM) paradigm designed to overcome the limitations of traditional DFM frameworks, which are constrained by their reliance on end-to-end differentiability. This new approach integrates highly accurate, non-differentiable refinement techniques, previously incompatible with DFM's differentiable nature, to improve performance on challenging non-isometric shapes. MDND employs a dual-branch architecture: one non-differentiable branch utilizes a multiscale iterative solver to generate robust correspondences, serving as a refined target. Concurrently, a fully differentiable branch learns to predict correspondences directly from features. The entire system is trained unsupervised, end-to-end, by enforcing a consistency loss that guides the differentiable branch to learn from the superior outputs of its non-differentiable counterpart. Experiments show MDND achieves leading results, demonstrating significant robustness against non-isometric deformations and topological noise.

Key takeaway

For Computer Vision Engineers developing shape correspondence systems, MDND offers a critical advancement. If you are struggling with non-isometric deformations or topological noise, consider adopting MDND's dual-branch approach. This method allows you to integrate highly accurate, non-differentiable refinement techniques, previously inaccessible, into your unsupervised learning pipelines. You can achieve superior robustness and leading performance by leveraging a consistency loss to guide your differentiable models.

Key insights

Merging differentiable and non-differentiable components enables unsupervised learning for robust shape correspondence.

Principles

• Differentiability constraints limit DFM performance.
• Non-differentiable refinement improves robustness.
• Consistency loss can guide unsupervised learning.

Method

MDND uses a dual-branch architecture: a non-differentiable branch with a multiscale iterative solver generates refined targets, while a differentiable branch learns from these via a consistency loss in unsupervised training.

In practice

• Integrate non-differentiable solvers into DFM.
• Use consistency loss for unsupervised learning.
• Apply to non-isometric shape matching tasks.

Topics

• Deep Functional Maps
• Shape Correspondence
• Unsupervised Learning
• Non-Differentiable Refinement
• Computer Vision
• MDND

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Coarse-to-Fine: A Hybrid Self-Supervised Method for Non-rigid 3D Shape Matching — A hybrid self-supervised coarse-to-fine method improves non-rigid 3D shape matching by integrating dual functional map learning streams.
• Geometry-Guided Modeling of Foundation Features Enables Generalizable Object Shape Deformation Learning — A geometry-guided deformation framework improves 3D object shape recovery by adapting templates to diverse views and novel categories.
• Inference with non-differentiable surrogate loss in a general high-dimensional classification framework — A kernel-smoothed decorrelated score enables robust inference for high-dimensional linear decision rules using non-differentiable surrogate loss functions.
• PSP: Harnessing Position and Shape Priors for Cross-Domain Few-Shot Medical Image Segmentation — Harnessing invariant anatomical position and shape priors enables robust cross-domain medical image segmentation despite texture discrepancies.
• Learning general conditional independence structures via the neighbourhood lattice — The neighbourhood lattice decomposition enables nonparametric, high-dimensional conditional independence learning without faithfulness assumptions.
• Bayesian 3D Steerable CNNs: Enabling Equivariance and Uncertainty Quantification Simultaneously — Bayesian Steerable-CNNs unify SE(3)-equivariance with uncertainty quantification via stochastic kernels and variational inference.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.