Research focus
Cross-Modal Geometric Uncertainty for Adaptive RGB–Thermal Visual Odometry
An M.S. research project investigating whether the relationship between RGB and thermal geometric uncertainty can improve visual odometry in degraded environments.
This project investigates whether metrics-aware geometric uncertainty from synchronized RGB and thermal observations can provide a meaningful measure of their relative reliability for visual odometry.
The initial question is not simply whether uncertainty can weight individual measurements. It is whether the relationship between RGB and thermal uncertainty is calibrated, comparable, and predictive of which modality is providing more reliable geometric information as conditions change.
The first phase will characterize that relationship against geometric error across challenging conditions such as low illumination, poor texture, obscurants, and changing thermal contrast. If the evidence supports it, a second phase will evaluate whether relationship-aware feature selection or weighting improves RGB–thermal visual odometry compared with appropriate baselines.
At a high level, the work uses the first semester to establish and evaluate the cross-modal reliability question, then uses the second semester to develop and test an adaptive approach if the findings justify it. The project does not assume a fixed fusion architecture or guaranteed improvement; its scope will be refined as results and lab direction develop.
The research is being carried out in Carnegie Mellon’s AirLab, under the advisement of Sebastian Scherer.