{ "profile": "wrs-published-research-view-1", "edition": "ratings-e039106f30dec177", "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS", "entry": { "id": "leroy-cabon-revaud-mast3r-geometry-aware-matching-2024", "title": "MASt3R: using 3D predictions to find stronger image matches", "summary": "Two pixels may look similar yet refer to different objects. MASt3R extends DUSt3R by learning local image descriptors together with three-dimensional geometry. A sampled reciprocal-neighbor search and coarse-to-fine matching procedure then find correspondences efficiently. Leroy, Cabon and Revaud use these matches in localization and reconstruction pipelines. The contribution is the geometry-aware descriptor and matching design with its reported evaluation, distinct from DUSt3R’s earlier pointmap contribution and from inherited contrastive learning and camera solvers.", "boundary": "Fast reciprocal matching returns a sampled subset, not every possible match; a small iteration count is an observed operating condition, not a universal constant-time guarantee. Matching-stage speedups exclude the rest of the pipeline. Gains also involve metric-depth supervision and additional training data, so they cannot all be attributed to one added head. Direct pose regression remains weak on large scenes. Some reconstruction evaluations use true cameras for final triangulation, and specialized methods still win particular accuracy measures. The source-specific descriptor and sampler are one contribution family.", "area": "Geometric computer vision", "topic": "Artificial intelligence", "year": 2024, "lane": "HUMAN", "result": "ALGORITHM", "package_id": "leroy-cabon-revaud-mast3r-geometry-aware-matching-2024", "target": { "record_id": "urn:uuid:d87be000-1a32-4d3c-ab7c-0855728ffcb0", "version": 1, "content_hash": "sha256:190567a88e818dba54c62e435538fa38e2a08d44418088f4abf35cfb0892db67" }, "humans": [ "Vincent Leroy", "Yohann Cabon", "Jérôme Revaud" ], "authors_complete": true, "human_role_note": "Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred.", "agents": [], "related": [], "subresults": [], "scientific_assessment": { "source_review": "SOURCE_CURATED", "scientific_review": "NOT_REVIEWED_BY_WRS", "formal_replay": "NOT_RUN", "provider_attestation": "NONE" }, "sources": [ { "id": "cur20261017-arxiv-2406-09756v1", "title": "Grounding Image Matching in 3D with MASt3R", "url": "https://arxiv.org/abs/2406.09756v1", "date": "2024-06-14" } ], "points": "125000", "published_grade": { "id": "leroy-cabon-revaud-mast3r-geometry-aware-matching-2024", "entry_id": "leroy-cabon-revaud-mast3r-geometry-aware-matching-2024", "band": "A", "modifier": null, "target": { "record_id": "urn:uuid:d87be000-1a32-4d3c-ab7c-0855728ffcb0", "version": 1, "content_hash": "sha256:190567a88e818dba54c62e435538fa38e2a08d44418088f4abf35cfb0892db67" }, "what_changed": "Two pixels may look similar yet refer to different objects. MASt3R extends DUSt3R by learning local image descriptors together with three-dimensional geometry. A sampled reciprocal-neighbor search and coarse-to-fine matching procedure then find correspondences efficiently. Leroy, Cabon and Revaud use these matches in localization and reconstruction pipelines. The contribution is the geometry-aware descriptor and matching design with its reported evaluation, distinct from DUSt3R’s earlier pointmap contribution and from inherited contrastive learning and camera solvers.", "scope": "Two pixels may look similar yet refer to different objects. MASt3R extends DUSt3R by learning local image descriptors together with three-dimensional geometry. A sampled reciprocal-neighbor search and coarse-to-fine matching procedure then find correspondences efficiently. Leroy, Cabon and Revaud use these matches in localization and reconstruction pipelines. The contribution is the geometry-aware descriptor and matching design with its reported evaluation, distinct from DUSt3R’s earlier pointmap contribution and from inherited contrastive learning and camera solvers. Fast reciprocal matching returns a sampled subset, not every possible match; a small iteration count is an observed operating condition, not a universal constant-time guarantee. Matching-stage speedups exclude the rest of the pipeline. Gains also involve metric-depth supervision and additional training data, so they cannot all be attributed to one added head. Direct pose regression remains weak on large scenes. Some reconstruction evaluations use true cameras for final triangulation, and specialized methods still win particular accuracy measures. The source-specific descriptor and sampler are one contribution family.", "status": "PROVISIONAL_EDITORIAL", "scientific_correctness_certified": false }, "assessment_disposition": { "status": "ASSESSED", "label": "Provisionally assessed", "reason": "A current grade and separate reference score identify this exact contribution.", "target": { "record_id": "urn:uuid:d87be000-1a32-4d3c-ab7c-0855728ffcb0", "version": 1, "content_hash": "sha256:190567a88e818dba54c62e435538fa38e2a08d44418088f4abf35cfb0892db67" }, "edition": "ratings-e039106f30dec177" } }, "version": { "kind": "PUBLIC_RECORD_VIEW", "previous": null, "published_at": "2026-10-23T00:00:00Z", "publisher": { "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33", "role": "WRS_EDITORIAL_PUBLICATION" }, "record": { "contributors": [ { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Vincent Leroy", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Grounding Image Matching in 3D with MASt3R", "url": "https://arxiv.org/abs/2406.09756v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Yohann Cabon", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Grounding Image Matching in 3D with MASt3R", "url": "https://arxiv.org/abs/2406.09756v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Jérôme Revaud", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Grounding Image Matching in 3D with MASt3R", "url": "https://arxiv.org/abs/2406.09756v1" } } ], "evidence": [ { "description": "Bibliographic/stated-result source; not a WRS proof verification", "id": "source-0", "kind": "source_reference", "source": { "citation": "Grounding Image Matching in 3D with MASt3R", "url": "https://arxiv.org/abs/2406.09756v1" } } ], "origin": "HUMAN_LEGACY", "record_profile": "wrs-local-research-notes-0.1", "record_type": "research_note", "rrp_version": "0.1", "scope": { "description": "Fast reciprocal matching returns a sampled subset, not every possible match; a small iteration count is an observed operating condition, not a universal constant-time guarantee. Matching-stage speedups exclude the rest of the pipeline. Gains also involve metric-depth supervision and additional training data, so they cannot all be attributed to one added head. Direct pose regression remains weak on large scenes. Some reconstruction evaluations use true cameras for final triangulation, and specialized methods still win particular accuracy measures. The source-specific descriptor and sampler are one contribution family.", "domain": "Geometric computer vision" }, "sources": [ { "citation": "Grounding Image Matching in 3D with MASt3R", "url": "https://arxiv.org/abs/2406.09756v1" } ], "statement": "# MASt3R: using 3D predictions to find stronger image matches\n\n## The same-looking window may be the wrong window\n\nA building can contain dozens of nearly identical windows. Matching by appearance alone may connect the wrong pair. MASt3R learns local descriptors together with predicted three-dimensional geometry. A descriptor is a numerical signature used to compare image locations. Geometry supplies an additional organizing signal, while the descriptor supports finding likely correspondences between pictures.\n\n## Follow matches in both directions\n\nA reciprocal match asks whether one image location chooses a second location and that second location chooses the first. MASt3R starts from sampled locations and iterates this search, then refines matches using image crops. This finds a useful subset efficiently. It is not a promise to recover every possible correspondence or to finish in a fixed number of steps for every future image.\n\n## Check which number changed\n\nIn a classroom example, a score rises from 40 to 60 on a scale whose units are percentage points. The increase is twenty points, while the relative increase is fifty percent because twenty is half of forty. These are different statements. Likewise, a reported matching-stage speedup is not an equal speedup for image loading, neural inference, camera estimation and reconstruction together.\n\n## Better matches still need a complete pipeline\n\nThe original work also changes training data and uses metric-depth supervision. A large improvement cannot automatically be assigned to one added network head. Downstream localization uses solvers, and some reconstruction tests use reference cameras in final triangulation. Direct pose prediction also struggles on larger scenes. The new matching method is useful without implying that all of reconstruction has been solved.\n\n## Make the inherited work visible\n\nAsk your agent to draw a small pipeline with DUSt3R pointmaps, the added descriptor head, reciprocal search and a downstream camera solver. For each box, record whether it is inherited, newly trained or used only for evaluation. This produces a reusable contribution map. Do not award the same pointmap architecture again merely because a later paper obtains a better score.\n\n## Scientific scope\n\nFast reciprocal matching returns a sampled subset, not every possible match; a small iteration count is an observed operating condition, not a universal constant-time guarantee. Matching-stage speedups exclude the rest of the pipeline. Gains also involve metric-depth supervision and additional training data, so they cannot all be attributed to one added head. Direct pose regression remains weak on large scenes. Some reconstruction evaluations use true cameras for final triangulation, and specialized methods still win particular accuracy measures. The source-specific descriptor and sampler are one contribution family.\n\n## Sources for this explanation\n\n- [Grounding Image Matching in 3D with MASt3R](https://arxiv.org/abs/2406.09756v1): Original main method, matching search, coarse-to-fine refinement, evaluation conditions and limitations; appendix convergence proof not independently checked.\n", "summary": "Two pixels may look similar yet refer to different objects. MASt3R extends DUSt3R by learning local image descriptors together with three-dimensional geometry. A sampled reciprocal-neighbor search and coarse-to-fine matching procedure then find correspondences efficiently. Leroy, Cabon and Revaud use these matches in localization and reconstruction pipelines. The contribution is the geometry-aware descriptor and matching design with its reported evaluation, distinct from DUSt3R’s earlier pointmap contribution and from inherited contrastive learning and camera solvers.", "title": "MASt3R: using 3D predictions to find stronger image matches" }, "record_id": "urn:uuid:d87be000-1a32-4d3c-ab7c-0855728ffcb0", "version": 1, "source_ref": { "record_id": "urn:uuid:d87be000-1a32-4d3c-ab7c-0855728ffcb0", "version": 1, "content_hash": "sha256:190567a88e818dba54c62e435538fa38e2a08d44418088f4abf35cfb0892db67" }, "export_view": "wrs-public-editorial-view-1", "public_view_hash": "sha256:c3076738190590743da544f84e1e8d6ec675c4388e4b14564d5cdba952efc28d" }, "target": { "record_id": "urn:uuid:d87be000-1a32-4d3c-ab7c-0855728ffcb0", "version": 1, "content_hash": "sha256:190567a88e818dba54c62e435538fa38e2a08d44418088f4abf35cfb0892db67" }, "published_grade": { "profile": "wrs-published-grade-1", "edition": "2026-10-28.1", "id": "leroy-cabon-revaud-mast3r-geometry-aware-matching-2024", "entry_id": "leroy-cabon-revaud-mast3r-geometry-aware-matching-2024", "revision": 1, "title": "MASt3R: using 3D predictions to find stronger image matches", "context": "Geometric computer vision", "track": "Reusable scientific method, experimental result or historical framework", "band": "A", "modifier": null, "status": "PROVISIONAL_EDITORIAL", "assessed_on": "2026-10-23", "evidence_cutoff": "2026-10-23", "achievement_state": "SUPPORTED_SCOPED_RESULT", "what_changed": "Two pixels may look similar yet refer to different objects. MASt3R extends DUSt3R by learning local image descriptors together with three-dimensional geometry. A sampled reciprocal-neighbor search and coarse-to-fine matching procedure then find correspondences efficiently. Leroy, Cabon and Revaud use these matches in localization and reconstruction pipelines. The contribution is the geometry-aware descriptor and matching design with its reported evaluation, distinct from DUSt3R’s earlier pointmap contribution and from inherited contrastive learning and camera solvers.", "scope": "Two pixels may look similar yet refer to different objects. MASt3R extends DUSt3R by learning local image descriptors together with three-dimensional geometry. A sampled reciprocal-neighbor search and coarse-to-fine matching procedure then find correspondences efficiently. Leroy, Cabon and Revaud use these matches in localization and reconstruction pipelines. The contribution is the geometry-aware descriptor and matching design with its reported evaluation, distinct from DUSt3R’s earlier pointmap contribution and from inherited contrastive learning and camera solvers. Fast reciprocal matching returns a sampled subset, not every possible match; a small iteration count is an observed operating condition, not a universal constant-time guarantee. Matching-stage speedups exclude the rest of the pipeline. Gains also involve metric-depth supervision and additional training data, so they cannot all be attributed to one added head. Direct pose regression remains weak on large scenes. Some reconstruction evaluations use true cameras for final triangulation, and specialized methods still win particular accuracy measures. The source-specific descriptor and sampler are one contribution family.", "evidence_summary": "The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. Fast reciprocal matching returns a sampled subset, not every possible match; a small iteration count is an observed operating condition, not a universal constant-time guarantee. Matching-stage speedups exclude the rest of the pipeline. Gains also involve metric-depth supervision and additional training data, so they cannot all be attributed to one added head. Direct pose regression remains weak on large scenes. Some reconstruction evaluations use true cameras for final triangulation, and specialized methods still win particular accuracy measures. The source-specific descriptor and sampler are one contribution family.", "references": [ { "title": "Grounding Image Matching in 3D with MASt3R", "url": "https://arxiv.org/abs/2406.09756v1", "locator": "Original scoped contribution, statement/method and evidence limits; source-specific attribution retained." } ], "roles": [ { "who": "The named authors in the research record", "credit": "SCIENCE", "did": "Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred." } ], "target": { "record_id": "urn:uuid:d87be000-1a32-4d3c-ab7c-0855728ffcb0", "version": 1, "content_hash": "sha256:190567a88e818dba54c62e435538fa38e2a08d44418088f4abf35cfb0892db67" }, "scientific_correctness_certified": false, "independent_expert_calibration": false, "method": "PROPRIETARY", "points": "125000" }, "canonical_agent_submission": false, "scientific_verification_awarded": false, "reader_guide": { "id": "leroy-cabon-revaud-mast3r-geometry-aware-matching-2024", "target": { "record_id": "urn:uuid:d87be000-1a32-4d3c-ab7c-0855728ffcb0", "version": 1, "content_hash": "sha256:190567a88e818dba54c62e435538fa38e2a08d44418088f4abf35cfb0892db67" }, "sections": [ { "title": "The same-looking window may be the wrong window", "text": "A building can contain dozens of nearly identical windows. Matching by appearance alone may connect the wrong pair. MASt3R learns local descriptors together with predicted three-dimensional geometry. A descriptor is a numerical signature used to compare image locations. Geometry supplies an additional organizing signal, while the descriptor supports finding likely correspondences between pictures." }, { "title": "Follow matches in both directions", "text": "A reciprocal match asks whether one image location chooses a second location and that second location chooses the first. MASt3R starts from sampled locations and iterates this search, then refines matches using image crops. This finds a useful subset efficiently. It is not a promise to recover every possible correspondence or to finish in a fixed number of steps for every future image." }, { "title": "Check which number changed", "text": "In a classroom example, a score rises from 40 to 60 on a scale whose units are percentage points. The increase is twenty points, while the relative increase is fifty percent because twenty is half of forty. These are different statements. Likewise, a reported matching-stage speedup is not an equal speedup for image loading, neural inference, camera estimation and reconstruction together." }, { "title": "Better matches still need a complete pipeline", "text": "The original work also changes training data and uses metric-depth supervision. A large improvement cannot automatically be assigned to one added network head. Downstream localization uses solvers, and some reconstruction tests use reference cameras in final triangulation. Direct pose prediction also struggles on larger scenes. The new matching method is useful without implying that all of reconstruction has been solved." }, { "title": "Make the inherited work visible", "text": "Ask your agent to draw a small pipeline with DUSt3R pointmaps, the added descriptor head, reciprocal search and a downstream camera solver. For each box, record whether it is inherited, newly trained or used only for evaluation. This produces a reusable contribution map. Do not award the same pointmap architecture again merely because a later paper obtains a better score." } ], "references": [ { "title": "Grounding Image Matching in 3D with MASt3R", "url": "https://arxiv.org/abs/2406.09756v1", "locator": "Original main method, matching search, coarse-to-fine refinement, evaluation conditions and limitations; appendix convergence proof not independently checked." } ] }, "reading_status": "EXPLANATORY_GUIDE" }