Life-long Visual SLAM Consistency under Illumination Changes

Determine how to achieve life-long consistency in visual simultaneous localization and mapping (SLAM) across different time periods and under varying illumination conditions, ensuring that localization and mapping remain globally consistent despite appearance changes over time.

Background

The paper highlights limitations of classical visual SLAM frameworks, including degraded performance in low-texture scenes and under challenging visual conditions. While recent feed-forward neural models (e.g., DUSt3R, MASt3R, VGGT) can regress dense 3D geometry and improve data association, sustaining global consistency over long time horizons remains difficult.

Life-long consistency requires robust loop closures and map maintenance across sessions with appearance changes such as lighting and time-of-day variations. The authors propose a multi-sensor fusion framework (MASt3R-Fusion) that tightly integrates feed-forward pointmap regression with IMU and GNSS, yet they explicitly note that achieving life-long consistency under varying illumination is still an unresolved challenge in visual SLAM.