Graph Matching Algorithms in Computer Vision

Graph matching constitutes a fundamental class of techniques for establishing correspondences between structured entities in images, where keypoints or regions are represented as nodes and their spatial or semantic relationships as edges. The problem can be formulated as a quadratic assignment problem, seeking a bijection between node sets that maximises overall affinity while respecting structural constraints. Exact solutions are typically infeasible for large graphs due to NP-hard complexity, prompting the development of approximate solvers, spectral relaxations and combinatorial heuristics. More recently, deep learning has been harnessed to learn node and edge embeddings via graph neural networks, enabling end-to-end matching that balances accuracy and efficiency. Applications span object recognition, shape correspondence, scene understanding, medical imaging and animation, and robustness to deformation, occlusion and noise remains a central challenge. Contemporary research emphasises hybrid schemes that integrate learned feature representations with classic optimisation frameworks, thereby improving generalisation across varying domains and sensor modalities.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

Combinatorial learning frameworks have advanced end-to-end graph matching by jointly training convolutional and graph neural network modules supervised by a permutation loss. Such approaches learn both visual descriptors and affinity kernels, converting the matching task into a linear assignment problem that adapts to variable graph sizes and exhibits strong generalisation across object categories and datasets. A spectral matching method builds a shape association graph in which node attributes encode geometric distances and edge attributes capture topological relations. By mapping topological dissimilarities into a geometric metric via Kendall shape space and exploiting the spectral decomposition of the resulting affinity matrix, it achieves global optimum correspondences even under severe non-rigid deformation, demonstrated on cel animation keyframes. A graph representation based on attributed relational SIFT-based regions graph integrates local appearance descriptors with spatial relationships. This structured graph allows efficient matching in diverse image retrieval and scene-analysis tasks by capturing both pointwise feature similarity and higher-order relational context.

Graph Matching Algorithms in Computer Vision publication trend

The graph below shows the total number of articles in graph matching algorithms in computer vision across all publications each year (not limited to Nature Index journals).

Technical terms

Affinity matrix: A matrix encoding pairwise similarity scores between nodes or edges in two graphs, guiding matching algorithms.

Quadratic assignment problem: A formulation of graph matching where node correspondences are optimised under pairwise interaction constraints, known to be NP-hard.

Graph neural network (GNN): A neural architecture that learns to embed nodes and edges of a graph into a feature space, preserving structural information for downstream tasks such as matching.

Spectral matching: A class of methods leveraging eigenvalue decomposition of affinity matrices to approximate global graph correspondences.

Permutation loss: A training objective measuring discrepancy between predicted and ground-truth node correspondences, often used in supervised graph matching.

References

1. Shape correspondence for cel animation based on a shape association graph and spectral matching. Computational Visual Media (2023).
2. Combinatorial Learning of Robust Deep Graph Matching: An Embedding Based Approach. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023).
3. Attributed Relational SIFT-Based Regions Graph: Concepts and Applications. Machine Learning and Knowledge Extraction (2020).