Spade-SLAM: Sparse Correspondence Anchors for Dense Self-Supervised Visual SLAM
Authors: Julian Philipp Schmitt, Weirong Chen, Ganlin Zhang, Felix Wimbauer, Daniel Cremers
Organizations: All authors are with Technical University of Munich, Munich, Germany. · All authors are with Munich Center for Machine Learning (MCML), Munich, Germany.
We introduce Spade-SLAM, a self-supervised dense visual SLAM framework that substantially narrows the long-standing performance gap to supervised methods in a zero-shot setting. By demonstrating that high-fidelity tracking can be achieved without explicit ground-truth pose and depth labels, our approach offers a highly effective and label-efficient alternative to fully supervised pretraining. Our core contribution is a complementary sparse-to-dense training regime: While a warping-based photometric loss provides a dense training signal, we stabilize the optimization and suppress drift in ambiguous regions using a sparse correspondence loss derived from image feature matches. Extensive experiments demonstrate that Spade-SLAM significantly improves robustness and accuracy compared to existing self-supervised baselines, closely rivaling SLAM systems trained with explicit pose and depth supervision. The code will be made publicly available at https://github.com/JulianPSchmitt/Spade-SLAM.
Figures & tables
Fig. 2 : Spade-SLAM pipeline. Our method jointly optimizes camera poses, dense depth, and a shared focal length over a frame graph G=(V,E) , where each image pair (i,j) represents an edge for iterative updates. Trained end-to-end via self-supervision, Spade-SLAM leverages photometric warping and sparse correspondence-based reprojection, eliminating the need for ground-truth pose or depth annotations.
Method
MH000
MH001
MH002
MH003
MH004
MH005
MH006
MH007
Avg.
Trad.
ORB-SLAM3 [ 13 ]
15.44
2.92
13.51
8.18
2.59
21.91
11.70
25.88
12.77
DSO [ 14 ]
9.92
0.35
7.96
3.46
-
12.58
8.42
7.50
-
Sup.
DeepV2D [ 15 ]
6.15
2.12
4.54
3.89
2.71
11.55
5.53
3.76
5.03
TartanVO [ 41 ]
4.88
0.26
2.00
0.94
1.07
3.19
1.00
2.04
1.92
DROID-SLAM [ 6 ]
0.07
0.05
0.03
0.02
3.03
0.63
0.38
0.07
0.54
DroidCalib [ 7 ]
0.06
0.03
0.02
0.02
1.89
0.30
0.25
0.13
0.34
TABLE I : TartanAir [ 40 ] benchmark (hard split): ATE RMSE [m]. The † symbol indicates methods using fine-tuning on the evaluation data. - means lack of results. Spade-SLAM performs the best on average across the self-supervised methods and performs close to supervised methods.
Method
MH01
MH02
MH03
MH04
MH05
V101
V102
V103
V201
V202
V203
Avg.
Trad.
ORB-SLAM3 [ 13 ]
0.016
0.027
0.028
0.138
0.072
0.033
0.015
0.033
0.023
0.029
X
-
DSO [ 14 ]
0.046
0.046
0.172
3.810
0.110
0.089
0.107
0.903
0.044
0.132
1.152
0.601
Sup.
DeepV2D [ 15 ]
0.739
1.144
0.752
1.492
1.567
0.981
0.801
1.570
0.290
2.202
2.743
1.298
TartanVO [ 41 ]
0.639
0.325
0.550
1.153
1.021
0.447
0.389
0.622
0.433
0.749
1.152
0.680
MASt3R-SLAM [ 8 ]
0.023
0.017
0.057
0.113
0.067
0.040
0.019
0.027
0.020
0.025
0.043
0.041
DROID-SLAM [ 6 ]
0.011
0.011
0.022
0.043
0.041
0.036
0.012
0.018
0.015
0.010
0.013
0.021
TABLE II : EuRoC MAV [ 42 ] benchmark: ATE RMSE [m]. The † symbol indicates methods using fine-tuning on the evaluation data. X means the system has failed here. - means lack of results. Spade-SLAM outperforms other self-supervised methods on average and matches DROID-SLAM [ 6 ] .
Method
360
desk
desk2
floor
plant
room
rpy
teddy
xyz
Avg.
Trad.
ORB-SLAM2 [ 3 ]
X
0.071
X
0.023
X
X
X
X
0.010
-
ORB-SLAM3 [ 13 ]
X
0.017
0.210
X
0.034
X
X
X
0.009
-
Sup.
DeepV2D [ 15 ]
0.243
0.166
0.379
1.653
0.203
0.246
0.105
0.316
0.064
0.375
TartanVO [ 41 ]
0.178
0.125
0.122
0.349
0.297
0.333
0.049
0.339
0.062
0.206
MASt3R-SLAM [ 8 ]
0.049
0.016
0.024
0.025
0.020
0.061
0.027
0.041
0.009
0.030
DROID-SLAM [ 6 ]
0.069
0.017
0.025
0.022
0.016
0.048
0.022
0.037
0.010
0.030
TABLE III : TUM-RGBD [ 43 ] benchmark: ATE RMSE [m]. X means the system has failed here. - means lack of results. Spade-SLAM achieves the best ATE on average among self-supervised methods and matches supervised DroidCalib [ 7 ] .
Fig. 3 : Trajectory estimation results on TUM-RGBD [ 43 ] ( room ) and 7-Scenes [ 44 ] ( stairs ). Trajectory color encodes the ATE RMSE. Our method demonstrates robust tracking comparable to supervised methods like DroidCalib [ 7 ] , whereas the photometric-only baseline struggles with small parallax and fails significantly in challenging scenes.
Method
Trajectory
Reconstruction
ATE
Acc.
Comp.
Cham.
Sup.
SLAM3R [ 19 ]
0.098
0.053
0.059
0.056
MASt3R-SLAM [ 8 ]
0.066
0.059
0.056
0.057
VGGT-SLAM [ 18 ]
0.070
0.052
0.061
0.056
ViSTA-SLAM [ 20 ]
0.055
0.045
0.056
0.051
DroidCalib [ 7 ]
0.053
0.024
0.068
0.046
TABLE IV : Tracking and Reconstruction Evaluation on 7-Scenes [ 44 ] . Spade-SLAM achieves competitive performance with fully supervised methods.
Fig. 4 : Reconstruction results on 7-Scenes [ 44 ] ( redkitchen ), TUM-RGBD [ 43 ] ( room ) and ScanNet [ 47 ] ( 0054_00 ). Red boxes highlight regions where the photometric-only baseline exhibits misalignment, while Spade-SLAM matches the supervised DroidCalib [ 7 ] .
Supervision
Trajectory
Reconstruction
Flow
ATE [m]
Acc.
Comp.
Cham.
EPE
Loss
photo. only
0.281
0.079
0.135
0.107
2.389
corres. only
0.633
0.245
0.235
0.240
1.920
Ours
0.048
0.031
0.090
0.061
1.580
Sparsity
GT corres. ( 196k )
0.017
0.027
0.077
0.052
0.844
GT corres. ( 20k )
0.020
0.027
0.081
0.056
0.900
TABLE V : Ablation study. Quantitative evaluation on the TartanAir [ 40 ] Hospital environment.
Method
100
101
102
103
104
105
106
107
Avg.
Ours
0.034
0.871
0.096
0.562
0.289
0.151
0.403
0.069
0.309
Ours (adapt.)
0.033
0.733
0.098
0.053
0.098
0.128
0.115
0.068
0.166
TABLE VI : Domain adaptation on ScanNet [ 47 ] , evaluated on held-out scenes 0100--0107 : ATE RMSE [m].
Method
000
001
002
003
004
005
006
007
Avg.
Ours
0.36
0.04
0.03
0.05
2.04
0.88
0.38
0.24
0.50
Ours + DL3DV
0.29
0.07
0.02
0.04
1.54
0.20
0.68
0.12
0.37
TABLE VII : Continued self-supervised training on DL3DV [ 48 ] , evaluated on the TartanAir [ 40 ] hard split: ATE RMSE [m]. Sequence labels omit the MH prefix.
Method
Metric
room0
room1
room2
office0
office1
office2
office3
office4
Avg.
DROID-SLAM [ 6 ]
Accuracy
0.030
0.023
0.019
0.016
0.017
0.023
0.025
0.024
0.022
Completeness
0.040
0.025
0.042
0.022
0.028
0.037
0.049
0.041
0.035
Chamfer
0.035
0.024
0.031
0.020
0.022
0.030
0.037
0.033
0.029
ATE
0.009
0.005
0.004
0.004
0.005
0.007
0.008
0.009
0.006
DroidCalib [ 7 ]
Accuracy
0.034
0.021
0.019
0.017
0.015
0.022
0.027
0.025
0.023
Completeness
0.042
0.026
0.040
0.024
0.027
0.038
0.047
0.043
0.036
TABLE VIII : Tracking and reconstruction results on Replica [ 49 ] . Spade-SLAM performs on par with supervised baselines.
Method
Metric
chess
fire
heads
office
pump.
redkit.
stairs
Avg.
Uncalibrated
ViSTA-SLAM [ 20 ]
Accuracy
0.065
0.015
0.031
0.036
0.061
0.035
0.074
0.045
Completeness
0.063
0.022
0.040
0.048
0.037
0.030
0.154
0.056
Chamfer
0.064
0.019
0.035
0.042
0.049
0.033
0.114
0.051
ATE
0.073
0.035
0.028
0.055
0.129
0.035
0.029
0.055
SLAM3R [ 19 ]
Accuracy
0.043
0.022
0.020
0.035
0.072
0.062
0.116
0.053
Completeness
0.030
0.013
0.015
0.030
0.055
0.061
0.209
0.059
TABLE IX : Per-scene evaluation on the 7-Scenes [ 44 ] dataset.
Fig. 5 : Additional reconstruction results on 7-Scenes [ 44 ] ( office , fire ), and TartanAir [ 40 ] ( MH000 ). Red boxes highlight regions in which our photometric-only baseline produces substantial reconstruction errors. In contrast, Spade-SLAM preserves geometric consistency.
Fig. 6 : Trajectory estimation results on TartanAir [ 40 ] . Trajectory color encodes the ATE RMSE, transitioning from blue (low error) to red (high error). Our method demonstrates stability on challenging scenes while AnyCam and the photometric-only baseline fail.
Method
360
desk
desk2
floor
plant
room
rpy
teddy
xyz
Avg.
E-RayZer [ 27 ]
0.079
0.148
0.206
0.053
0.030
0.086
0.028
0.042
0.012
0.076
Ours
0.118
0.017
0.036
0.018
0.092
0.034
0.033
0.017
0.011
0.042
TABLE X : Quantitative comparison with E-RayZer [ 27 ] . Evaluating ATE RMSE [m] on the first 200 frames of TUM-RGBD (Freiburg 1), Spade-SLAM outperforms the modern Gaussian Splatting-based self-supervised model.
Method
TartanAir
7-Scenes
AnyCam
7.94
0.207
AnyCam (fine-tuned)
5.14
0.167
Ours
0.50
0.061
TABLE XI : Fine-tuned AnyCam on TartanAir [ 40 ] and 7-Scenes [ 44 ] . We report average ATE RMSE [m].
Fig. 7 : Illustration of sparse correspondences used for training on TartanAir [ 40 ] .
We introduce VGGT-SLAM++, a complete visual SLAM system that leverages the geometry-rich outputs of the Visual Geometry Grounded Transformer (VGGT). The system comprises a visual odometry (front-end) fusing the VGGT feed-forward transformer and a Sim(3) solution, a Digital Elevation Map (DEM)-based graph construction module, and a back-end that jointly enable accurate large-scale mapping with bounded memory. While prior transformer-based SLAM pipelines such as VGGT-SLAM rely primarily on sparse loop closures or global Sim(3) manifold constraints - allowing short-horizon pose drift - VGGT-SLAM++ restores high-cadence local bundle adjustment (LBA) through a spatially corrective back-end. For each VGGT submap, we construct a dense planar-canonical DEM, partition it into patches, and compute their DINOv2 embeddings to integrate the submap into a covisibility graph. Spatial neighbors are retrieved using a Visual Place Recognition (VPR) module within the covisibility window, triggering frequent local optimization that stabilizes trajectories. Across standard SLAM benchmarks, VGGT-SLAM++ achieves state-of-the-art accuracy, substantially reducing short-term drift, accelerating graph convergence, and maintaining global consistency with compact DEM tiles and sublinear retrieval.
Geometric foundation models estimate dense geometry and camera parameters from monocular images, but reliable SLAM also requires consistent scale and accurate data association across successive reconstructions. We present DASH-SLAM, a monocular SLAM system that integrates metric depth predictions from pretrained Depth Anything 3 into a shared-keyframe pose graph without task-specific fine-tuning. DASH-SLAM derives scale corrections from depth ratios at shared anchor images and accumulates them across reconstructions, with a damping factor controlling their influence. These corrections rescale odometry translations, loop constraints, and map geometry. Experiments on TUM, Replica, and 7-Scenes show that DASH-SLAM outperforms state-of-the-art methods. We evaluate performance using both rigid and similarity alignment to distinguish metric-scale fidelity from accuracy up to scale. We further evaluate DASH-SLAM on four challenging sequences in perceptually-degraded environments from the Unified Autonomy Stack field dataset, covering tunnels, fog, and a frozen lake. On these sequences, DASH-SLAM outperforms the foundation-model SLAM state-of-the-art and maintains tracking where classical SLAM methods fail.
Otavio Ribas, Jacopo Panerati, Pierre-Yves Lajoie
Department of Computer and Software Engineering, Polytechnique Montreal, Montreal, Canada. · Mila - Quebec AI Institute, Montreal, Canada. · National Research Council Canada, Montreal, Canada.
Visual SLAM is a well-established technology utilized in a wide range of real-world applications. However, its performance still degrades under challenging visual conditions, such as low texture, severe motion blur, and poor illumination. Systems based on deep learning outperform classical geometry-based ones and achieve state-of-the-art results by combining learned 2D data association and uncertainty with differentiable geometric optimization in recurrent architectures. Still, it remains unclear exactly which components are fundamentally responsible for this success. In this paper, we ask: Is the superior performance of deep learning-based systems driven primarily by learned 2D data association, the combination of learned 2D data association and uncertainty, or the recurrent architecture itself? We investigate this question empirically by conducting a controlled study. Our findings reveal that the success of DL-based V-SLAM systems hinges on learned 2D data association and uncertainty rather than their recurrent architecture, underscoring the necessity of learning-based paradigms for the design of these components.
Giovanni Cioffi, Davide Scaramuzza
Robotics and Perception Group, Department of Informatics, University of Zurich, Switzerland