Robust Structureless Monocular Visual Inertial Initialization Exploiting Line Features and Vanishing Points
Junwan Choi, Woongrae Jo, Dong-Uk Seo, Jinwoo Jeon, Hyun Myung
IROS 2026
cs.RO, cs.CV
2026-09-18
SLIM-init uses vanishing points from 2D lines for gyro bias and line epipolar plus normal-projection residuals for scale alignment. EuRoC mean scale RMSE is 0.120, while structure-based initializers fail outright on custom degenerate sequences.
Monocular visual-inertial odometry has to recover scale, gravity, velocity, and IMU biases before the filter is usable. Weak excitation, tiny parallax, or translation-dominated motion makes point triangulation ill-conditioned, and initialization fails. The same failure shows up after tracking loss, when the estimator has to restart in place.
Both existing families hurt. Structure-based pipelines run visual SfM then IMU alignment, or keep 3D landmarks in a joint solve: slow, and the structure itself collapses at low parallax. Structureless linear alignment is fast but still leans on point parallax, so low texture or near-pure translation wrecks the conditioning. Line-augmented systems usually rebuild 3D line landmarks or line depths, which invites the structure dependence back in.
SLIM-init sits on He et al.’s rotation-translation decoupled structureless frame and never reconstructs 3D points or 3D lines. Points are corners tracked with KLT. Lines come from EDLines, optical-flow prediction, and geometric matching.
Step 1 estimates gyroscope bias only. Beside the usual point residual, a vanishing-point direction term is added. VPs live at infinity, so the residual is translation-invariant and still constrains rotation when parallax is scarce. Each frame keeps one dominant VP from spherical voting plus weighted least squares, gated by track age and angle; unreliable tracks fall back to points only.
Step 2, after bias and rotation are fixed, linearly aligns scale, gravity, and per-keyframe velocity. Two structureless line residuals join the point position term: a line epipolar residual that uses coplanarity of two interpretation planes to pin relative translation, and a line-normal projection residual that projects the point position error onto the interpretation-plane normal, giving the linear system an anisotropic structural direction. Lines are weighted by track persistence and length; not every detection enters the solver.
EuRoC left camera, an initialization attempt every 10 frames with a 10-keyframe window. Success is |scale error| < 0.5; RMSE is computed on successes only. Baselines are VINS-Mono, OpenVINS’s closed form, and structureless DRT-l.
| Method | Scale RMSE | Gravity RMSE (°) | Velocity RMSE (m/s) | Pose RMSE (m) |
| VINS-Mono | 0.195 | 1.696 | 0.126 | 0.125 |
| OpenVINS | 0.131 | 2.745 | 0.192 | 0.132 |
| DRT-l | 0.136 | 1.365 | 0.113 | 0.126 |
| SLIM-init | 0.120 | 1.311 | 0.100 | 0.110 |
Versus DRT-l, mean RMSE drops 11.8% (scale), 4.0% (gravity), 11.5% (velocity), and pose RMSE is 0.110, 12.7% lower. Machine Hall success is 94.01% against 91.77% for DRT-l and 63.84% for VINS-Mono. Vicon rooms sit near 72–73%, with structure-based methods worse.
Custom corridor and lobby sequences at KAIST add weak texture, repetition, low parallax, and translation-dominant motion, with FAST-LIO2 as ground truth. VINS-Mono and OpenVINS never initialize. DRT-l still runs; SLIM-init improves mean scale 0.151 vs 0.166 and pose 0.206 vs 0.216, a small but consistent edge.
The solver averages 4.141 ms, matching DRT-l at 4.150 ms, versus 29.615 ms for VINS-Mono and 284.715 ms for OpenVINS. With line and VP tracking, cost is about 35.86 ms per frame, which the authors treat as real-time for startup and frequent recovery.
The authors call this the first strictly structureless use of lines in monocular VIO initialization: 2D segments enter residuals directly, with no line depth and no 3D line landmarks. For indoor robots, drones, and AR, the value is “it still comes up under degenerate motion,” not another point or two on EuRoC. Code is public, and the paper is at IROS 2026.
The gain is a conditioning patch, not a new estimation paradigm. EuRoC improvements over DRT-l are on the order of ten percent. The gap that matters is the corridors where structure-based methods never start.
The VP ablation is thin: removing the VP residual moves relative gyro-bias error from 1.913% to 1.941%, inside the noise. A single dominant VP assumes Manhattan-like directions; outdoor or curved scenes may have nothing to vote for. The degenerate set is three sequences, so “structure methods all fail” rests on a small sample, and laser-inertial ground truth is a different sensor stack than visual-inertial error.
RMSE is computed only on successful windows, which flatters the averages unless read next to success rates. Line tracking is still an extra front end; feature tracking remains the time bottleneck. Long-term stability needs a VIO backend. This paper only covers the initialization window.