Research map / Label-efficient and robust learning / Recognition and long-tail
MARCO: Navigating the Unseen Space of Semantic Correspondence
Claudia Cuttano, Gabriele Trivigno, Carlo Masone, Stefan Roth
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add adapters to frozen DINOv2 (ablation) | MARCO architecture · SPair-71k | PCK@0.10: 53.9 -> 76.9 | Improved |
| add the upsampling head after adapters (ablation) | MARCO architecture · SPair-71k | PCK@0.10: 76.9 -> 87.2 | Improved |
| use coarse-to-fine sigma schedule instead of fixed sigma=3 (ablation) | MARCO supervision · SPair-71k | PCK@0.01: 19.5 -> 27.0 | Improved |
| use coarse-to-fine sigma schedule instead of fixed sigma=1 (ablation) | MARCO supervision · SPair-71k | PCK@0.10: 81.0 -> 87.2 | Improved |
| add dense self-distillation loss (ablation) | MARCO · SPair-U | PCK@0.10: 41.8 -> 67.5 | Improved |
| use raw semantic flow as pseudo-labels in dense loss (ablation) | MARCO dense self-distillation · SPair-U | PCK@0.10: 41.8 -> 49.6 | Mixed |
| add Delaunay triangulation and piecewise-affine warp to MNN matches (ablation) | MARCO dense self-distillation · SPair-U | PCK@0.10: 52.5 -> 64.7 | Improved |
| add GT keypoint anchoring of flow clusters (ablation) | MARCO dense self-distillation · SPair-U | PCK@0.10: 64.7 -> 67.5 | Improved |
| fully fine-tune the DINOv2 backbone instead of keeping it frozen (ablation) | DINOv2 · SPair-U | PCK@0.10: 54.9 -> 43.9 | Hurt |
| use AdaptFormer instead of LoRA (ablation) | MARCO adaptation · SPair-U | PCK@0.10: 65.6 -> 67.5 | Improved |
| add flow-anchoring self-distillation loss to Geo-SC | Geo-SC · SPair-U | PCK@0.10: 56.9 -> 63.4 | Improved |
| use coarse-to-fine loss instead of InfoNCE + l2 (ablation) | Adapter + Upsample model · SPair-71k | PCK@0.01: 20.0 -> 26.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- GECO — GECO: Geometrically Consistent Embedding with Lightspeed Inference (ICCV 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Jamais Vu: Exposing the Generalization Gap in Supervised Semantic Correspondence (NeurIPS 2025 · similarity 0.90)
- Unsupervised Semantic Correspondence Using Stable Diffusion (NeurIPS 2023 · similarity 0.89)
- Semantic Correspondence: Unified Benchmarking and a Strong Baseline (TPAMI 2025 · similarity 0.88)
- Mechanisms of Object Localization in Vision-Language Models (CVPR 2026 · similarity 0.87)
- Grounding Everything in Tokens for Multimodal Large Language Models (CVPR 2026 · similarity 0.86)
- From Correspondence to Actions: Human-Like Multi-Image Spatial Reasoning in Multi-modal Large Language Models (ICML 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
