Simultaneous Localization and Mapping (SLAM) is extensively used in Extended Reality(XR) Head Mounted Display(HMD), robots, autonomous driving, and various other fields. This algorithm empowers devices to localize in unfamiliar environments and reconstruct environmental maps using its sensors. SLAM is also crucial in downstream tasks, including environment perception and route planning.

Samsung R&D Institute China-Beijing (SRC-B), having dedicated years to SLAM research, is actively developing XR glasses to cater to consumer needs. To deliver optimal user experience, SRC-B consistently pursues technological innovation and has published several papers in SLAM-related fields at international academic conferences, including the International Conference on Intelligent Robots and Systems (IROS) and the International Conference on Computer Vision (ICCV).

Established in 1987, the International Conference on Computer Vision (ICCV) is a biennial research conference sponsored by the Institute of Electrical and Electronics Engineers (IEEE). It stands among the top conferences in computer vision, alongside the Computer Vision and Pattern Recognition Conference (CVPR) and the European Conference on Computer Vision (ECCV). Beyond papers, ICCV incorporates workshops, with each hosting challenges in various fields. The ICCV 2023 Workshop on Robot Learning and SLAM is a workshop aiming to showcase the latest findings on the theory and practice of both traditional and modern techniques for robot learning, robot perception, and SLAM. The workshop comprises a seminar and a SLAM challenge.

This year’s SLAM challenge features TartanAir and SubT-MRS datasets, aiming to enhance the robustness of SLAM algorithms in challenging environments and advance sim-to-real transfer with a rich set of sensory modalities, including RGB images, LiDAR points, IMU measurements, thermal images and so on. The challenge comprises three tracks: Visual-Inertial SLAM, LiDAR-Inertial SLAM, and Sensor Fusion SLAM. A team of researchers from SRC-B participated in the Visual-Inertial track, navigating through challenging conditions such as lighting changes, darkness, smoke, and self-similar environments and providing a test from simulation to real-world scenarios. The Samsung team secured first place in the Visual-Inertial SLAM Challenge among 29 participating teams. Huazhong University of Science and Technology secured the second place, and Nanyang Technological University clinched the third place.

Our solution is built on our self-developed DVI-SLAM, the first learning-based SLAM framework tightly coupling dual visual inertial optimization. DVI-SLAM integrates photometric and re-projection factors into an end-to-end differentiable structure through a multifactor data association module. Our network dynamically learns and adjusts confidence maps for both visual factors and can be further extended to include the IMU factors. The final submission attains a mean ATE and RPE of 0.547 m and 0.049 m, respectively.

Samsung Team in the ICCV SLAM Challenge 2023