[Paper Review] Robust Near-Isometric Matching via Structured Learning of Graphical Models

Julian McAuley, Tibério S. Caetano+1 more|ArXiv.org|Sep 21, 2008

Advanced Image and Video Retrieval Techniques22 references4 citations

TL;DR

This paper proposes a structured learning approach using a tractable graphical model that jointly encodes distance, angle, and appearance features to enable robust near-isometric shape matching. By learning feature weights via large-margin structured prediction, the model achieves significant accuracy improvements over prior methods while maintaining efficient inference, even under scale and appearance variations.

ABSTRACT

Models for near-rigid shape matching are typically based on distance-related features, in order to infer matches that are consistent with the isometric assumption. However, real shapes from image datasets, even when expected to be related by "almost isometric" transformations, are actually subject not only to noise but also, to some limited degree, to variations in appearance and scale. In this paper, we introduce a graphical model that parameterises appearance, distance, and angle features and we learn all of the involved parameters via structured prediction. The outcome is a model for near-rigid shape matching which is robust in the sense that it is able to capture the possibly limited but still important scale and appearance variations. Our experimental results reveal substantial improvements upon recent successful models, while maintaining similar running times.

Motivation & Objective

• Address the limitations of traditional isometry-based shape matching that ignore appearance and scale variations.
• Improve robustness in real-world datasets where shapes deviate from perfect isometry due to noise, scale changes, and appearance differences.
• Develop a structured learning framework that integrates multiple feature types (distance, angle, appearance) into a single tractable graphical model.
• Enable efficient, exact inference in the correspondence space while learning the relative importance of different feature types through structured prediction.
• Overcome the issue of feature scaling in heterogeneous feature spaces by using a two-stage learning approach to stabilize model training.

Proposed method

• Formulate shape matching as a structured prediction problem over correspondences using a graphical model with cliques capturing pairwise and higher-order geometric relationships.
• Integrate three types of features: isometry-invariant distance features, angle features, and appearance-invariant features (e.g., SIFT) into the model’s potential functions.
• Use a two-stage structured learning approach to first identify candidate correspondences and then refine them using higher-order features.
• Apply large-margin structured prediction to jointly learn the weights of all feature types, allowing the model to automatically prioritize relevant features.
• Ensure computational efficiency by leveraging chordal graphical models with small clique size, enabling exact MAP inference in polynomial time.
• Address feature scale sensitivity by normalizing features before learning and using a two-stage training process to stabilize weight estimation.

Experimental results

Research questions

• RQ1 Can a unified graphical model that combines geometric, angular, and appearance features improve robustness in near-isometric shape matching?
• RQ2 How does structured learning of feature weights compare to fixed or heuristic weighting schemes in terms of matching accuracy?
• RQ3 To what extent do higher-order features (e.g., third-order angles) contribute to performance beyond first- and second-order features?
• RQ4 Can the model maintain high efficiency while incorporating richer feature representations and learning their relative importance?
• RQ5 How does the model perform under occlusions and feature scale variations, and can learning mitigate these issues?

Key findings

• The proposed model achieves substantial performance improvements over recent state-of-the-art methods on benchmark shape matching datasets, even with similar running times.
• The second-stage learning process assigns near-zero weight to first-order features, indicating that higher-order features provide more discriminative information once candidate matches are identified.
• The inclusion of scale-invariant third-order features significantly improves matching accuracy, capturing structural patterns not encoded in lower-order features.
• The two-stage learning approach effectively stabilizes training despite heterogeneous feature scales, enabling reliable weight learning.
• The model remains exact and efficient due to the chordal structure of the graphical model, allowing tractable MAP inference.
• The method outperforms baseline models using only distance or appearance features, demonstrating the benefit of joint modeling of multiple invariances.