{"paper":{"title":"Metric-Guided Feature Fusion of Visual Foundation Models for Segmentation Tasks","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Label-free metrics identify complementary VFM pairs for fusion that boosts dense prediction performance.","cross_cats":["cs.AI"],"primary_cat":"cs.CV","authors_text":"AntonioManuel Lopez Pena, Danna Xue, JoseLuis Gomez Zurita, Yachan Guo, Yi Xiao","submitted_at":"2026-05-16T07:55:41Z","abstract_excerpt":"Although large-scale visual foundation models (VFMs) achieve remarkable performance in semantic understanding, they still underperform in instance-aware dense prediction tasks. They exhibit different biases in representation: for instance, promptable segmentation models (e.g., SAM2) focus on fine-grained region boundaries, while self-supervised models (e.g., DINOv3) emphasize object-level structure. This observation highlights the potential of combining complementary features from different VFMs to enhance downstream dense prediction tasks. However, naive multi-VFM fusion seldom leads to relia"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our model shows consistent performance gains across multiple dense prediction tasks compared with the baselines, with better object-level semantics and more accurately localized boundaries.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The label-free metrics for Structural Coherence and Edge Fidelity in feature space can reliably identify which VFM encoders are complementary and worth fusing, without any task-specific labels or supervision.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"A label-free metric-guided fusion of complementary features from visual foundation models yields consistent gains in dense prediction tasks with improved object semantics and boundary localization.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Label-free metrics identify complementary VFM pairs for fusion that boosts dense prediction performance.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"50830b49320fce8ddced87309dd8d55cfe6bdcf705a47df085c445125dc9a83a"},"source":{"id":"2605.16864","kind":"arxiv","version":1},"verdict":{"id":"868cdf77-1868-46c1-b763-f381f9d99669","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T20:26:34.510690Z","strongest_claim":"Our model shows consistent performance gains across multiple dense prediction tasks compared with the baselines, with better object-level semantics and more accurately localized boundaries.","one_line_summary":"A label-free metric-guided fusion of complementary features from visual foundation models yields consistent gains in dense prediction tasks with improved object semantics and boundary localization.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The label-free metrics for Structural Coherence and Edge Fidelity in feature space can reliably identify which VFM encoders are complementary and worth fusing, without any task-specific labels or supervision.","pith_extraction_headline":"Label-free metrics identify complementary VFM pairs for fusion that boosts dense prediction performance."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.16864/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_compliance","ran_at":"2026-05-19T20:31:44.381971Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T20:31:19.122114Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T18:41:56.303185Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T18:33:26.378943Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"71a5fc4daf7a12f480c31109a281a13b9b2ff539173344fd0e249b6fd2f000d4"},"references":{"count":47,"sample":[{"doi":"","year":2022,"title":"Improving vision transformers by revisiting high-frequency components","work_id":"d21dbe51-1b67-4166-956b-106e4fef0fdd","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Do computer vision foundation models learn the low- level characteristics of the human visual system? InCVPR,","work_id":"12adea76-f63c-49f0-a732-6e8bf6c0b518","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"Emerg- ing properties in self-supervised vision transformers","work_id":"c74dee00-daa4-4cce-b59d-01b7051ae456","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Sam-adapter: Adapting segment any- thing in underperformed scenes","work_id":"fdfb2839-ed13-4443-ba34-b918c3bedc04","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Vision transformer adapter for dense predictions","work_id":"8216f877-3f01-49d8-9d4f-44102469d1bf","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":47,"snapshot_sha256":"f94968d33f2b6a0af3a76c51a92724c15bd9f80bb23bb0ecf9e123f1f7955f3c","internal_anchors":3},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}