COMMONPLUS

COMMON+

Official implementation of the following paper:

• Yijie Lin, Mouxing Yang, Peng Hu, Jiancheng Lv, Hao Chen, Xi Peng. "Learning with Partial and Noisy Correspondence in Graph Matching". TPAMI, 2026. [paper]

COMMON+ proposes a unified deep graph matching framework for learning under partial and noisy correspondence, where graph matching models must simultaneously handle:

• Partial correspondence: some nodes have no valid matches due to outliers, occlusion, truncation, or structural mismatch;
• Noisy correspondence: the supervision itself is imperfect, including both inaccurate node-level matches and corrupted edge-level relations.

As illustrated above, in real-world graph matching scenarios, images often exhibit large appearance variations, viewpoint changes, partial overlaps, and ambiguous structures. These factors make it difficult to establish clean one-to-one correspondences: some keypoints may be incorrectly annotated or spatially shifted, resulting in noisy node-to-node matches, while some nodes may have no valid counterparts at all. Such imperfect node correspondence further contaminates higher-order structural relations, leading to unreliable edge-level correspondence.

To address these challenges, COMMON+ introduces an Align–Fuse–Refine pipeline that combines complementary graph matching experts:

• Supporting VGG16 CNN or VIT backbone to extract visual features;
• A quadratic contrastive learning objective to enhance correspondence-aware structural alignment;
• A dual-expert collaborative framework that integrates KB-QAP-based alignment and L-QAP-based fusion/matching;
• A momentum cooperation and refinement mechanism to progressively identify outliers and correct noisy supervision during training.

This design enables COMMON+ to robustly handle both missing correspondences and corrupted annotations within a unified graph matching framework.

Benchmark Results

COMMON+ supports multiple graph matching benchmarks under different settings, including standard 2GM, partial matching, and noisy-correspondence evaluation.

VGG16-based configs

PascalVOC

• Task: 2GM
• Config: experiments/vgg16_common_plus_voc.yaml

PascalVOC (with outliers)

• Task: 2GM
• Config: experiments/vgg16_common_plus_voc-all.yaml

Willow ObjectClass

• Task: 2GM
• Config: experiments/vgg16_common_plus_willow.yaml

Willow ObjectClass (noisy correspondence)

• Task: 2GM with noisy correspondence
• Config: experiments/vgg16_common_plus_willow_nc.yaml

Willow ObjectClass (partial matching with outliers)

• Task: partial 2GM setting
• Config: experiments/vgg16_common_plus_willow_outlier.yaml

SPair-71k

• Task: 2GM
• Config: experiments/vgg16_common_plus_spair71k.yaml

IMC-PT-50

• Task: partial graph matching
• Config: experiments/vgg16_common_plus_imcpt_50.yaml

IMC-PT-100

• Task: partial graph matching
• Config: experiments/vgg16_common_plus_imcpt_100.yaml

---

ViT-based configs

PascalVOC

• Task: 2GM
• Config: experiments/vit_common_plus_voc.yaml

Willow ObjectClass

• Task: 2GM
• Config: experiments/vit_common_plus_willow.yaml

SPair-71k

• Task: 2GM
• Config: experiments/vit_common_plus_spair71k.yaml

File Organization

├── README.md
│   project overview and usage instructions
├── model.py
│   implementation of the training and evaluation procedures of COMMON+
├── model_config.py
│   declaration of model hyperparameters and configurations
├── sconv_archs.py
│   implementation of spline convolution (SplineCNN) operations
└── transformconv.py
    implementation of transformer-based graph fusion modules

Credits and Citation

Please cite the following paper if you use this model in your research:

@article{lin2022graph,
  title={Graph Matching with Bi-level Noisy Correspondence},
  author={Lin, Yijie and Yang, Mouxing and Yu, Jun and Hu, Peng and Zhang, Changqing and Peng, Xi},
  journal={IEEE International Conference on Computer Vision},
  year={2022}
}
@article{lin2026learning,
  title={Learning With Partial and Noisy Correspondence in Graph Matching},
  author={Lin, Yijie and Yang, Mouxing and Hu, Peng and Lv, Jiancheng and Chen, Hao and Peng, Xi},
  journal={IEEE transactions on pattern analysis and machine intelligence},
  year={2026}
}