Organizations: School of Automation and Intelligent Sensing, Institute of Medical Robotics, Shanghai Jiao Tong University, Shanghai 200240, China. · School of Mathematical Sciences Shanghai Jiao Tong University.
Robust stereo visual-inertial odometry (VIO) remains challenging in low-texture scenes and under abrupt illumination changes, where point features become sparse and unstable, leading to ambiguous association and under-constrained estimation. Line structures offer complementary geometric cues, yet many efficient point-line systems still rely on point-guided line association, which can break down when point support is weak and may lead to biased constraints. We present a stereo point-line VIO system in which line segments are equipped with dedicated deep descriptors and matched using an entropy-regularized optimal transport formulation, which performs global mass-transport assignment and supports unmatched observations under ambiguity, outliers, and partial observations. The proposed descriptor is training-free and is computed by sampling and pooling network feature maps. To improve estimation stability, we analyze the impact of line measurement noise and introduce reliability-adaptive weighting to regulate the influence of line constraints during optimization. Experiments on EuRoC and UMA-VI, together with real-world deployments in low-texture and illumination-challenging environments, demonstrate improved accuracy and robustness over representative baselines while maintaining real-time performance.
Figures & tables
Fig. 1: Overview and key modules of OTPL-VIO. The upper-left panel shows the stereo point–line VIO pipeline; the remaining panels detail the deep line descriptor, OT-based line association, and reliability-adaptive line weighting.
Fig. 2: Our data-collection platform.
Method
Sequence
E
E
M
H
H
E
M
H
E
M
H
MH01
MH02
MH03
MH04
MH05
V101
V102
V103
V201
V202
V203
Avg.
VINS-Fusion [ 10 ]
16.30
17.80
31.60
33.10
17.50
10.20
9.90
11.20
11.00
12.40
25.20
17.84
DROID-SLAM [ 14 ]
16.30
12.10
24.20
39.90
27.00
10.30
16.50
15.80
10.20
11.50
20.40
18.56
Struct-VIO [ 25 ]
11.90
10.00
28.30
27.50
25.60
7.50
19.70
16.10
8.10
15.20
17.70
17.05
PLF-VINS [ 20 ]
14.30
17.80
22.10
24.00
26.00
6.90
9.90
16.60
8.30
12.50
18.30
16.06
TABLE I: ATE RMSE [cm] on EuRoC dataset. E, M, and H denote Easy, Medium, and Hard sequences, respectively.
Fig. 3: Qualitative evaluation on EuRoC V103 and V203. Estimated trajectories are colored by ATE. Red boxes indicate challenging views: abrupt illumination changes in V103 and low-texture regions in V203. Red crosses and green dots denote failed and successful triangulations, respectively.
Sequence
PL-
ORB-
Basalt
OKVIS
DROID-
AirSLAM
Ours
SLAM [ 1 ]
SLAM3 [ 12 ]
[ 29 ]
[ 11 ]
SLAM [ 14 ]
[ 23 ]
conf-csc1
269.7
✘
127.0
111.8
71.1
49.0
37.3
conf-csc2
159.6
✘
68.2
47.0
13.5
9.1
11.6
conf-csc3
✘
42.6
46.9
8.8
72.4
8.8
3.7
lab-rev
✘
6.3
48.6
86.1
36.4
50.4
38.7
lab-csc
✘
✘
40.3
57.9
31.9
97.9
25.9
TABLE II: ATE RMSE [cm] on UMA-VI illumination-change sequences. ✘ denotes tracking failure or large drift error.
Fig. 4: Qualitative results on UMA-VI. Estimated trajectories are colored by ATE where reference poses are available; dashed lines indicate segments without reference poses. Insets show accumulated end-point drift. Red boxes mark low-texture and illumination-changing views.
Fig. 5: Qualitative results on real-world sequences. Estimated trajectories are overlaid on the floor plans for Static-1, Static-2, and Dynamic-1. Red boxes show representative low-texture, illumination-changing, dynamic, and occlusion scenes; orange boxes enlarge trajectory endpoints.
Sequence
VINS-Fusion [ 10 ]
PL-VINS [ 21 ]
AirSLAM [ 23 ]
Ours
corridor-eng
39.92
34.52
6.20
5.12
parking-csc1
33.14
55.58
48.99
20.19
parking-csc2
26.03
50.24
38.82
18.05
class-csc1
19.23
22.03
18.70
10.29
class-csc2
23.23
29.81
17.50
4.37
Avg.
28.31
38.43
26.04
11.60
TABLE III: ATE RMSE [cm] on UMA-VI Low-Texture sequences.
Sequence
Scenario
VINS-Fusion
PL-VINS
AirSLAM
Ours
Static-1
Low texture + illum. change
56.65
299.82
60.96
39.99
Static-2
Low texture + illum. change
273.12
90.08
37.77
25.71
Dynamic-1
Low texture + dynamic occl.
43.50
116.92
52.61
37.37
TABLE IV: ATE RMSE [cm] on real-world sequences. Illum. and occl. denote illumination and occlusion, respectively.
Method
DLD
OTA
RAW
Low-Tex.
Illum.
Avg.
Baseline
AirSLAM (point-guided line association)
26.04
44.06
38.05
Ablation variants
LBD + NN
18.57
31.36
27.09
LBD + NN + RAW
✓
16.77
30.73
26.07
DLD + NN
✓
17.87
30.82
26.50
TABLE V: Ablation study on UMA-VI subsets [cm]. DLD: deep line descriptor; OTA: OT-based association; RAW: reliability-adaptive weighting.
Method
Illumination Changes
Viewpoint Changes
Overall
P ↑
R ↑
F ↑
P ↑
R ↑
F ↑
P ↑
R ↑
F ↑
LBD+NN
38.44
78.02
51.51
11.08
54.89
18.44
25.93
72.08
38.14
DLD+NN
44.41
90.13
59.50
17.15
84.96
28.55
31.94
88.81
46.98
DLD+OTA
55.18
67.63
60.77
20.99
46.71
28.97
42.00
62.26
50.16
TABLE VI: Line association results on HPatches.
Fig. 6: Qualitative line association examples on HPatches. The two rows show viewpoint- and illumination-change examples, respectively. Blue segments denote detected lines, while green and red connections indicate true positives and false positives, respectively.
Fig. 7: Cross-platform runtime on EuRoC V1 sequences. Boxplots show per-frame runtimes of AirSLAM and our method on V101, V102, and V103.
Seq.
Maplab
PL-
ORB-
Basalt
DVI-
DROID-
AirSLAM
Ours
[ 15 ]
SLAM [ 1 ]
SLAM3 [ 12 ]
[ 29 ]
SLAM [ 16 ]
SLAM [ 14 ]
[ 23 ]
Avg.
7.1
6.1
3.5
5.2
5.1
2.4
3.0
2.8
TABLE VII: Average ATE RMSE [cm] on EuRoC with loop closure.
Method
EuRoC
UMA-VI
AirSLAM [ 23 ]
23.57
47.28
Ours
22.95
38.24
TABLE VIII: Average ATE RMSE [cm] on EuRoC and UMA-VI low-texture subset without IMU.