Hilti–Trimble SLAM Challenge 2026 · ICRA 2026

√VINS-Based 360 Visual-Inertial SLAM Systems

2nd Place (SLAM Category)
ICRA 2026 Workshop on Open Challenges in Robotics for Asset Inspection and Management (OCRAIM)
Jungwoo Lee1, Hogyun Kim1, Jiwon Choi1, Jinyong Jeong2, and Younggun Cho1
1Spatial AI and Robotics (SPARO) Lab, Inha University 2Riibotics
2ndSLAM category
2393.9final score, 25 sequences
0.109 mmean RMSE
360° + IMUdual-fisheye images and IMU only
Abstract

A 360-camera visual-inertial SLAM system

This report describes our submission to the 2026 Hilti–Trimble SLAM Challenge. The system is a 360-camera visual-inertial pipeline that uses only the provided dual-fisheye image streams and IMU measurements. First, raw images are preprocessed with low-light enhancement and static/dynamic masking, and then fed into a √VINS frontend as the local odometry, with deep learning-based SuperPoint features, ELSED line segments, and LightGlue matching. The resulting trajectory is refined using a GTSAM-based backend that performs tightly-coupled factor graph optimization with SALAD global descriptors and visual bundle adjustment (BA).

The backend performs local-to-global structureless BA, tightly regularized intrinsic and extrinsic refinement, and global scale adjustment supported by the known inter-camera baseline and cross-view constraints. The design targets long indoor sequences with dark regions, repetitive corridors, dynamic objects, and limited geometric parallax, while keeping the pipeline reproducible in the provided ROS2 workspace.

Pipeline

From two fisheye views to one optimized map

Overview of the 360-camera visual-inertial SLAM pipeline
Overview. Raw dual-fisheye frames are conditioned by low-light enhancement and dynamic masking, then tracked by a √VINS-based frontend with SuperPoint–LightGlue point features and ELSED line segments. Loop closures and short-term cross-camera matches are geometrically verified and inserted as smart projection factors into a GTSAM backend that jointly optimizes poses, calibration and a global scale.
(a)

Input

The 360 camera as two calibrated fisheye views (EUCM model) with the IMU; all poses live in the cam0 frame.

(b)

Pre-processing

Static masks, low-light enhancement (Zero-DCE++) and YOLO26 masks that suppress features on moving people.

(c)

Point & line VINS

√VINS frontend with SuperPoint + LightGlue points and ELSED lines under a vertical-world assumption.

(d)

Loop detection

SALAD global descriptors propose loops; SuperPoint–LightGlue verification across both fisheye views.

(e)

Factor graph

Odometry, BA, loop and cross-camera factors, plus regularized extrinsic, intrinsic and scale variables.

Results

Maps on the floorplan

Bundle-adjusted landmarks and optimized trajectories on several indoor sequences of the challenge, overlaid on the building floorplan. The alignment to the floorplan was done manually, for visualization only.

Estimated maps on the floorplans of six sequences
Official evaluation

Score

2393.90final score (sum over 25 evaluated sequences)
0.109 mmean RMSE
0.095 mmean position error
99.85%mean trajectory coverage

From the official evaluator report. The mean position error stays below 0.1 m across the evaluated sequences, with an average of 95.8 points per sequence.