Scalix: Uncertainty-Aware Scale-Consistent Monocular SLAM
Authors: Sebastian Barbas Laina, Tianyi Zhang, Panagiotis Petropoulakis, Simon Schaefer, Simon Boche, Jaehyung Jung, Cedric Le Gentil, Stefan Leutenegger
Cameras are ubiquitous sensors in robotics due to their compact form factor and the perceptual richness captured through visual information. Monocular SLAM enables robots to understand the environment with a minimum setup, however, it inherently suffers from scale ambiguity. A common solution is to provide multi-modal sensor configurations, such as visual-inertial systems, where scale is observable unless the robot navigates under a constant-velocity motion, a common scenario in mobile robotics. With the advent of deep-learning, geometric foundation models have been used to address this problem, but the depths maps are often noisy and scale-inconsistent across frames. In this paper, we propose Scalix, a real-time monocular SLAM framework that achieves metric-scale state estimation by integrating learned depth cues into a probabilistic factor-graph formulation. By augmenting existing monocular depth models with both per-pixel depth uncertainty and per-frame scale uncertainty, Scalix treats scale predictions as independent measurements within its optimization, leading to improved scale consistency through multi-view data associations. Experiments in large-scale outdoor and indoor environments demonstrate state-of-the-art performance on both metric and up-to-scale benchmarks while maintaining real-time operation and generalization.
Figures & tables
Fig. 1: Scalix on a Large-scale Sequence [ 16 ] . Scalix achieves accurate, metric-scale trajectory estimation on challenging scenes. The main figure shows the estimated trajectory (colored by time), the sparse landmarks (upper-right diagonal) and the dense back-projected depth maps scaled using the optimized scale factors (lower-left diagonal). In the lower-left corner, we directly align the trajectory to the map via SE(3) alignment.
Fig. 2: Scalix System Overview. Our method builds upon OKVIS2 [ 2 ] , a classical SLAM architecture comprising a frontend and a backend. Components with pink background indicate our key contributions. Each keyframe is associated with an additional scale state and we use an enhanced monocular depth model to obtain depth and scale measurements. In the frontend, landmark positions are initialized using the predicted depth and scale and the backend adds depth and scale error terms into the factor graph optimization. The pose graph is extended to account for the scale in the marginalization process.
Fig. 3: Monocular Depth Network Augmented with Decoupled Uncertainty Prediction. Modules with dark blue background denote the original depth prediction network architecture in Metric3Dv2 [ 23 ] . The yellow box denotes the scale uncertainty head, which outputs a global per-frame scale as a 1D scalar. The purple box denotes the depth uncertainty head, which outputs independent per-pixel uncertainty. The two uncertainties are combined to form the total uncertainty, and we visualize the projected total uncertainty in 3D space.
Fig. 4: Factor Graph in Scalix. Top: the real-time local window estimator. The current keyframe and non-keyframe states are connected by visual reprojection errors. For every keyframe state in the optimization window, depth and scale errors are incorporated into the cost function. Bottom: the full real-time estimator. Old measurements are marginalized and substituted by pose-scale graph edges, reducing the computational complexity.
Method
00
01
02
03
04
05
06
07
08
09
10
Avg.
Avg.
3.7 km
2.5 km
5.1 km
0.6 km
0.4 km
2.2 km
1.2 km
0.6 km
3.2 km
1.7 km
0.9 km
success
SIM(3)
ORB-SLAM3 † [ 1 ]
8.94
103.76
24.90
0.76
0.79
5.33
12.87
2.55
4.45
8.39
6.57
16.30
7.55
LDSO [ 7 ]
9.32
11.68
31.98
2.85
1.22
5.10
13.55
2.96
129.00
21.64
17.36
22.42
11.77
DROID-SLAM [ 10 ]
92.10
344.60
107.60
2.38
1.00
118.50
62.47
21.78
161.60
72.32
118.70
100.28
50.77
DPV-SLAM++ [ 11 ]
8.30
11.86
39.64
2.50
0.78
5.74
11.60
1.52
110.90
76.70
13.70
25.75
10.63
DROID+M3D † [ 10 , 23 ]
12.84
150.44
20.03
2.08
0.83
7.78
3.19
18.32
9.73
3.76
4.83
21.26
8.34
TABLE I: Trajectory evaluation on the Kitti dataset. RMSE ATE [m] with SIM(3) and SE(3) alignments ( Best , and second best ).
Fig. 5: Qualitative Trajectory Comparison on KITTI [ 16 ] under Sim(3) Alignment. The color gradient from red to green indicates trajectory error from high to low. The colored lines represent the estimated trajectories and the gray lines represent the ground-truth trajectories.
Method
chess
fire
heads
office
pump.
kit.
stairs
Avg.
SIM(3)
ORB-SLAM3 † [ 1 ]
4.2
2.9
16.3
8.6
40.8
37.3
1.2
15.9
LDSO † [ 7 ]
4.0
12.9
51.0
14.9
–
49.9
23.9
26.1
DROID-SLAM [ 10 ]
3.6
2.7
2.5
6.6
12.7
3.5
2.6
5.0
DROID+M3D † [ 10 , 23 ]
4.9
6.4
9.6
14.2
16.3
4.7
14.6
10.1
CUT3R [ 13 ]
5.9
5.3
6.4
13.9
14.7
9.4
6.7
8.9
MAST3R-SLAM [ 12 ]
5.3
2.5
1.5
9.7
8.8
4.1
1.1
4.7
TABLE II: Trajectory evaluation on the 7scenes dataset. RMSE ATE [cm] with SIM(3) and SE(3) alignments ( Best , and second best ).
SIM(3)
SE(3)
Scalix
w/o s-opt
Scalix
w/o s-opt
Seq.
prop.
direct
prop.
direct
00
2.85
2.56
2.17
3.96
3.55
2.18
01
10.72
96.91
66.65
21.47
575.58
618.65
02
15.04
18.23
18.16
16.55
20.11
18.47
03
1.90
1.84
1.67
9.44
10.76
13.47
TABLE III: Trajectory evaluation on the Kitti dataset. RMSE ATE [m] with SIM(3) and SE(3) alignments ( Best , and second best ).