Map-free Relocalization Leaderboard | Niantic Spatial, Inc.

Map-Free BenchmarkHome Dataset Leaderboard Submit

Evaluation Leaderboard

Click a column to sort. Bars indicate relative performance per metric.

Public benchmark

All submissions

Methods are ranked by AUC (VCRE < 45px) by default. Lower-is-better metrics sort ascending first; higher-is-better metrics sort descending first.

Methods evaluated with a single reference image and a single query frame.

Method	AUC (VCRE < 45px)↕	AUC (VCRE < 90px)↕	Precision (VCRE < 45px)↕	Precision (VCRE < 90px)↕	Median Reproj. Error (px)↕	AUC (Err < 25cm, 5°)↕	Precision (Err < 25cm, 5°)↕	Median Trans. Error (m)↕	Median Rot. Error (°)↕	Estimates for % of frames↕
Mert Bülent Sarıyıldız, Philippe Weinzaepfel, Thomas Lucas, Pau de Jorge, Diane Larlus, Yannis Kalantidis · NAVER LABS EuropePaper Code	0.840	0.943	64.4%	81.1%	47.43	0.768	55.9%	0.389	4.564	99.6%
Submitted 2024-11-14 Authors Mert Bülent Sarıyıldız, Philippe Weinzaepfel, Thomas Lucas, Pau de Jorge, Diane Larlus, Yannis Kalantidis · NAVER LABS Europe Description A MASt3R decoder fine-tuned over DUNE, a ViT-Base Distilled UNiversal Encoder [CVPR 2025] Links Paper Code
Mert Bülent Sarıyıldız, Philippe Weinzaepfel, Thomas Lucas, Pau de Jorge, Diane Larlus, Yannis Kalantidis · NAVER LABS EuropePaper Code	0.834	0.939	63.8%	79.5%	49.70	0.754	53.8%	0.411	4.596	99.9%
Submitted 2025-06-12 Authors Mert Bülent Sarıyıldız, Philippe Weinzaepfel, Thomas Lucas, Pau de Jorge, Diane Larlus, Yannis Kalantidis · NAVER LABS Europe Description A MASt3R decoder fine-tuned over DUNE, a ViT-Small Distilled UNiversal Encoder [CVPR 2025] Links Paper Code
David Nordström, David Nordström, Xinyue Zhang, Thibaut Loiseau, Vincent Lepetit, Fredrik Kahl · Chalmers University of TechnologyCode	0.817	0.945	60.2%	79.2%	55.37	0.755	50.2%	0.440	5.041	100.0%
Submitted 2026-06-18 Authors David Nordström, David Nordström, Xinyue Zhang, Thibaut Loiseau, Vincent Lepetit, Fredrik Kahl · Chalmers University of Technology Description RoMa-Omega is a paper where we take the RoMa v2 model and add a VGGT-Omega backbone. Paper to be released shortly. Links Code
Vincent Leroy, Yohann Cabon and Jerome Revaud · Naver Labs EuropePaper Code	0.817	0.933	63.0%	79.3%	48.79	0.740	54.7%	0.368	2.202	99.4%
Submitted 2024-07-23 Authors Vincent Leroy, Yohann Cabon and Jerome Revaud · Naver Labs Europe Description Forward with (ref, query) gives pointmaps and pixel matches. These latter ones are used for Ess. matrix estimation, and the scale is obtained from the predicted pointmaps. (re-submission of previous results) Links Paper Code
David Nordström, Johan Edstedt, Georg Bökman, Jonathan Astermark, Anders Heyden, Viktor Larsson, Mårten Wadenbäck, Michael Felsberg, Fredrik Kahl · Chalmers University of Technology, Linköping University, Lund University, University of AmsterdamPaper Code	0.796	0.925	57.2%	75.7%	56.70	0.754	48.8%	0.500	4.358	99.2%
Submitted 2026-05-06 Authors David Nordström, Johan Edstedt, Georg Bökman, Jonathan Astermark, Anders Heyden, Viktor Larsson, Mårten Wadenbäck, Michael Felsberg, Fredrik Kahl · Chalmers University of Technology, Linköping University, Lund University, University of Amsterdam Description LoMa is a strong local feature matching that scales the LightGlue objective to unprecedented scales. It even beats RoMa and RoMa v2 on select benchmarks. Links Paper Code
Shigeki Kobayashi · GO Drive Inc.	0.736	0.870	48.2%	63.2%	80.84	0.662	39.8%	0.834	11.89	100.0%
Submitted 2026-09-28 Authors Shigeki Kobayashi · GO Drive Inc.
Lili Zhao, Zhili Liu, Lei Yang, Meng Guo · China Mobile Research Institute, N/A, China Mobile Research Institute, China Mobile Research InstitutePaper	0.681	0.796	39.9%	51.4%	125.6	0.593	31.9%	1.754	31.22	100.0%
Submitted 2024-07-25 Authors Lili Zhao, Zhili Liu, Lei Yang, Meng Guo · China Mobile Research Institute, N/A, China Mobile Research Institute, China Mobile Research Institute Description Metric3d for 3d scaled depth estimation and loftr for initial patch matching.Then an geometric pnp algorithm is used for pose estimation. Links Paper
Mingyu Xiao ; Runze Chen ; Haiyong Luo ; Fang Zhao;Juan Wang;XuePeng Ma · Beijing University of Posts and Telecommunications;Research Center for Ubiquitous Computing Systems, Institute of Computing Technology, Chinese Academy of Sciences;Shouguang Cheng Zhi Feng Xing Technology Co., Ltd.Paper	0.656	0.849	45.3%	65.0%	79.96	0.469	28.6%	0.828	11.74	99.0%
Submitted 2024-07-25 Authors Mingyu Xiao ; Runze Chen ; Haiyong Luo ; Fang Zhao;Juan Wang;XuePeng Ma · Beijing University of Posts and Telecommunications;Research Center for Ubiquitous Computing Systems, Institute of Computing Technology, Chinese Academy of Sciences;Shouguang Cheng Zhi Feng Xing Technology Co., Ltd. Links Paper
David Nordström · Chalmers University of Technology	0.641	0.780	36.0%	49.4%	126.0	0.551	27.5%	1.693	26.71	98.5%
Submitted 2026-01-29 Authors David Nordström · Chalmers University of Technology Description Simply take the official example for SG + EMAT and change depths for metric depths from DAv3.
RoMa: Edstedt et al. - MicKey: Barroso-Laguna et al. · RoMa: Linköping University. - MicKey: Niantic.	0.604	0.734	37.8%	52.4%	111.9	0.546	31.4%	1.177	15.60	99.9%
Submitted 2024-06-07 Authors RoMa: Edstedt et al. - MicKey: Barroso-Laguna et al. · RoMa: Linköping University. - MicKey: Niantic. Description [ECCV '24 Workshop Baseline]: RoMa correspondences with MicKey depth maps predictions.
Axel Barroso-Laguna, Sowmya Munukutla, Victor Adrian Prisacariu, Eric Brachmann · NianticPaper Code	0.572	0.748	31.2%	49.3%	129.5	0.325	13.3%	1.657	27.26	100.0%
Submitted 2023-11-07 Authors Axel Barroso-Laguna, Sowmya Munukutla, Victor Adrian Prisacariu, Eric Brachmann · Niantic Description Matching 2D Images in 3D: Metric Relative Pose from Metric Correspondences Links Paper Code
Axel Barroso-Laguna, Sowmya Munukutla, Victor Adrian Prisacariu, Eric Brachmann · NianticPaper Code	0.558	0.741	30.1%	49.2%	126.9	0.283	12.0%	1.591	26.00	100.0%
Submitted 2023-10-31 Authors Axel Barroso-Laguna, Sowmya Munukutla, Victor Adrian Prisacariu, Eric Brachmann · Niantic Description Matching 2D Images in 3D: Metric Relative Pose from Metric Correspondences Links Paper Code
SuperGlue: Sarlie et al. - MicKey: Barroso-Laguna et al. · SuperGlue: ETH Zurich, Magic Leap, Inc. - MicKey: Niantic	0.556	0.711	29.8%	43.0%	139.9	0.490	23.5%	1.696	26.13	98.5%
Submitted 2024-07-02 Authors SuperGlue: Sarlie et al. - MicKey: Barroso-Laguna et al. · SuperGlue: ETH Zurich, Magic Leap, Inc. - MicKey: Niantic Description [ECCV '24 Workshop Baseline]: SuperPoint + SuperGlue correspondences with MicKey depth maps predictions.
LoFTR: Sun et al. - MicKey: Barroso-Laguna et al. · LoFTR: Zhejiang University, SenseTime Research - MicKey: Niantic	0.550	0.715	27.2%	40.7%	155.0	0.467	20.3%	1.922	33.60	100.0%
Submitted 2024-07-02 Authors LoFTR: Sun et al. - MicKey: Barroso-Laguna et al. · LoFTR: Zhejiang University, SenseTime Research - MicKey: Niantic Description [ECCV '24 Workshop Baseline]: LoFTR correspondences with MicKey depth maps predictions.
Hitesh Jain, Sagar Verma · Granular AIPaper	0.548	0.720	28.0%	45.5%	142.0	0.273	10.8%	1.842	30.78	100.0%
Submitted 2024-07-03 Authors Hitesh Jain, Sagar Verma · Granular AI Description Using the depth map from the Depth Anything Model as the ground truth, we trained the Mickey Variant network in a supervised manner. Links Paper
Chris Rockwell, Nilesh Kulkarni, Linyi Jin, JJ Park, Justin Johnson, David F. Fouhey · University of Michigan, New York UniversityPaper Code	0.481	0.680	25.3%	44.2%	137.1	0.392	17.7%	1.482	17.26	100.0%
Submitted 2024-02-16 Authors Chris Rockwell, Nilesh Kulkarni, Linyi Jin, JJ Park, Justin Johnson, David F. Fouhey · University of Michigan, New York University Description Our flexible method produces accurate and robust pose estimates using complementary strengths of Correspondence + Solver and Learning-Based methods. Links Paper Code
Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, Jerome Revaud · NAVER LABS EUROPEPaper Code	0.459	0.698	28.7%	50.4%	115.8	0.394	21.4%	0.989	7.131	99.6%
Submitted 2024-03-03 Authors Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, Jerome Revaud · NAVER LABS EUROPE Description DUSt3R --> pixel correspondences between output 3d points Links Paper Code
Chris Rockwell, Nilesh Kulkarni, Linyi Jin, JJ Park, Justin Johnson, David F. Fouhey · University of Michigan, New York UniversityPaper Code	0.456	0.668	25.8%	44.2%	135.5	0.351	17.1%	1.498	17.28	100.0%
Submitted 2024-02-16 Authors Chris Rockwell, Nilesh Kulkarni, Linyi Jin, JJ Park, Justin Johnson, David F. Fouhey · University of Michigan, New York University Description Our flexible method produces accurate and robust pose estimates using complementary strengths of Correspondence + Solver and Learning-Based methods. Links Paper Code
Johan Edstedt · Linköping University	0.437	0.670	26.4%	45.7%	128.9	0.408	22.9%	1.235	11.19	100.0%
Submitted 2023-06-24 Authors Johan Edstedt · Linköping University Description Same as loftr one, 2px opencv MAGSAC etc.
Jaime Spencer, Chris Russell, Simon Hadfield, Richard Bowden · University of Surrey, Oxford Internet Institute, University of Surrey, University of SurreyPaper Code	0.420	0.635	19.4%	34.4%	165.1	0.295	11.1%	2.238	37.82	100.0%
Submitted 2023-08-08 Authors Jaime Spencer, Chris Russell, Simon Hadfield, Richard Bowden · University of Surrey, Oxford Internet Institute, University of Surrey, University of Surrey Description We replace the baseline DPT-KITTI with KBR++ depth predictions. KBR++ is a self-supervised monocular depth estimation system trained on a combination of SlowTV, CribsTV, MannequinChallenge & Kitti. The model is transferred to MapFreeReloc in a zero-shot fashion, without fine-tuning. Since KBR++ is trained only on monocular data (ie without a stereo baseline), the predictions are scaleless. Metric-scale predictions are obtained by aligning KBR++ to the provided baseline DPT-KITTI using least-squares. Links Paper Code
Jaime Spencer, Chris Russell, Simon Hadfield, Richard Bowden · University of Surrey, Oxford Internet Institute, University of Surrey, University of SurreyPaper Code	0.420	0.631	19.5%	34.0%	167.5	0.278	10.5%	2.326	39.52	100.0%
Submitted 2023-08-08 Authors Jaime Spencer, Chris Russell, Simon Hadfield, Richard Bowden · University of Surrey, Oxford Internet Institute, University of Surrey, University of Surrey Description We replace the baseline DPT-KITTI with KBR depth predictions. KBR is a self-supervised monocular depth estimation system trained on a combination of SlowTV, MannequinChallenge & Kitti. The model is transferred to MapFreeReloc in a zero-shot fashion, without fine-tuning. Since KBR is trained only on monocular data (ie without a stereo baseline), the predictions are scaleless. Metric-scale predictions are obtained by aligning KBR to the provided baseline DPT-KITTI using least-squares. Links Paper Code
Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David McKinnon, Yanghai Tsin, Long Quan · Apple Inc, HKUSTPaper	0.414	0.641	20.8%	37.0%	161.8	0.361	16.3%	1.901	29.25	100.0%
Submitted 2024-07-03 Authors Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David McKinnon, Yanghai Tsin, Long Quan · Apple Inc, HKUST Description [ECCV'24 Workshop Baseline]: ASpanFormer: Detector-Free Image Matching with Adaptive Span Transformer Links Paper
	0.397	0.618	18.2%	33.5%	166.8	0.269	9.8%	2.312	39.42	100.0%
Submitted 2022-09-27
	0.387	0.615	18.8%	34.7%	167.6	0.346	15.5%	1.987	30.58	100.0%
Submitted 2022-09-27
	0.385	0.602	20.6%	36.2%	160.3	0.346	16.8%	1.881	25.44	100.0%
Submitted 2022-09-27
	0.378	0.599	19.6%	36.0%	157.0	0.253	10.7%	2.108	32.58	100.0%
Submitted 2022-09-27
Shigeki Kobayashi · GO Drive Inc.	0.373	0.541	37.3%	54.1%	106.2	0.220	22.0%	1.360	18.20	100.0%
Submitted 2026-09-28 Authors Shigeki Kobayashi · GO Drive Inc.
Shigeki Kobayashi · GO Drive Inc.	0.369	0.533	36.9%	53.3%	108.1	0.213	21.3%	1.400	19.30	100.0%
Submitted 2026-09-26 Authors Shigeki Kobayashi · GO Drive Inc.
Shigeki Kobayashi · GO Drive Inc.	0.360	0.526	36.0%	52.6%	113.1	0.187	18.7%	1.470	20.70	100.0%
Submitted 2026-09-25 Authors Shigeki Kobayashi · GO Drive Inc.
Map-free leaderboard baseline · Niantic - Map-free leaderboard baseline	0.355	0.527	19.5%	33.2%	138.8	0.314	15.9%	1.445	18.52	78.0%
Submitted 2024-07-03 Authors Map-free leaderboard baseline · Niantic - Map-free leaderboard baseline Description [ECCV '24 Workshop Baseline]: DISK+LightGlue correspondences with DPT-KITTI depth maps.
Michał J. Tyszkiewicz, Pascal Fua, Eduard Trulls · École Polytechnique Fédérale de Lausanne (EPFL), Google Research, Zurich	0.346	0.536	15.1%	26.9%	208.2	0.264	10.2%	2.593	51.96	100.0%
Submitted 2024-07-02 Authors Michał J. Tyszkiewicz, Pascal Fua, Eduard Trulls · École Polytechnique Fédérale de Lausanne (EPFL), Google Research, Zurich Description [ECCV '24 Workshop Baseline]: DISK correspondences with DPT-KITTI depth maps.
Shigeki Kobayashi · GO Drive Inc.	0.327	0.502	32.7%	50.2%	118.9	0.153	15.3%	1.630	23.00	100.0%
Submitted 2026-09-24 Authors Shigeki Kobayashi · GO Drive Inc.
Johan Edstedt, Georg Bökman, Mårten Wadenbäck, Michael Felsberg · Linköping University, Chalmers University of TechnologyPaper Code	0.325	0.531	16.9%	31.2%	167.4	0.265	12.5%	2.021	30.30	88.0%
Submitted 2024-07-03 Authors Johan Edstedt, Georg Bökman, Mårten Wadenbäck, Michael Felsberg · Linköping University, Chalmers University of Technology Description [ECCV'24 Workshop Baseline]: DeDoDe: Detect, Don't Describe -- Describe, Don't Detect for Local Feature Matching Links Paper Code
	0.313	0.505	13.7%	25.0%	222.9	0.253	10.3%	2.930	61.43	100.0%
Submitted 2022-09-27
	0.299	0.468	14.3%	25.1%	192.0	0.191	7.8%	3.354	63.74	100.0%
Submitted 2022-09-27
Pierre Gleize, Weiyao Wang, Matt Feiszli · Meta AIPaper Code	0.192	0.312	9.8%	18.1%	176.4	0.157	7.3%	2.208	33.81	52.2%
Submitted 2024-07-03 Authors Pierre Gleize, Weiyao Wang, Matt Feiszli · Meta AI Description [ECCV'24 Workshop Baseline]: SiLK: Simple Learned Keypoints Links Paper Code
	0.181	0.355	18.1%	35.5%	166.4	0.105	10.5%	1.831	23.27	100.0%
Submitted 2022-09-27
	0.181	0.387	18.1%	38.7%	148.7	0.061	6.1%	1.698	22.94	100.0%
Submitted 2022-09-27
	0.156	0.402	15.6%	40.2%	147.1	0.061	6.1%	1.682	22.52	100.0%
Submitted 2022-09-27

Methods evaluated using sequences of multiple query frames (9-frame window).

Method	AUC (VCRE < 45px)↕	AUC (VCRE < 90px)↕	Precision (VCRE < 45px)↕	Precision (VCRE < 90px)↕	Median Reproj. Error (px)↕	AUC (Err < 25cm, 5°)↕	Precision (Err < 25cm, 5°)↕	Median Trans. Error (m)↕	Median Rot. Error (°)↕	Estimates for % of frames↕
Baseline Axel Barroso-Laguna, Sowmya Munukutla, Victor Adrian Prisacariu, Eric Brachmann · NianticPaper Code	0.572	0.752	31.3%	49.6%	128.9	0.326	13.6%	1.656	26.92	100.0%
Submitted 2024-05-20 Authors Axel Barroso-Laguna, Sowmya Munukutla, Victor Adrian Prisacariu, Eric Brachmann · Niantic Description [ECCV '24 Workshop Baseline]: Matching 2D Images in 3D: Metric Relative Pose from Metric Correspondences, without using multiple images or device poses. Only using the last frame in the query sequence, not using device poses. Links Paper Code
Mingyu Xiao ; Runze Chen ; Haiyong Luo ; Fang Zhao · Beijing University of Posts and TelecommunicationsPaper	0.387	0.721	25.2%	53.1%	105.8	0.190	10.8%	0.951	15.10	99.1%
Submitted 2024-08-22 Authors Mingyu Xiao ; Runze Chen ; Haiyong Luo ; Fang Zhao · Beijing University of Posts and Telecommunications Description Map-Free Visual Relocalization Enhanced by Instance Knowledge and Depth Knowledge use only single frame Links Paper

Click a column header to sort. Bars indicate relative performance per metric. Entries update automatically from public benchmark submissions.

Reconstruct

Localize

Understand

Capture