Answers are grounded in this paper and cite their sources.
Ground-robot visual descriptors may predict unmanned aerial vehicle (UAV) relative-altitude telemetry, but their transferability across recordings remains uncertain. We studied two ground robots and two UAVs in an agricultural field in Rengo, Chile, evaluating altitude-telemetry prediction across recordings from one UAV and separately assessing RGB–thermal detection, cross-UAV detector transfer, and complementary viewpoints. The regression target was DJI takeoff-referenced relative altitude; no independent calibrated altitude measurement was available. Class-specific regressors using visual descriptors with optional gimbal metadata were selected through source-only temporal validation and tested on the opposite recording in a retrospective evaluation. Against the DJI telemetry reference, ground-truth boxes yielded macro-averaged mean absolute error of 0.907 m and root mean squared error (RMSE) of 1.065 m; these are agreement errors, not independently measured physical-altitude errors. Clear improvement over a source-mean baseline occurred only for A1 → A2; reverse transfer showed no clear improvement, and all four test R2 values were negative. On matched detections, YOLO26n boxes changed RMSE by −0.016 to +0.071 m, while coverage ranged from 41.3% to 95.1%. Detection results varied across domains, architectures, and seeds. With manual time alignment and known target correspondence, counting a target as detected when either UAV detected it increased recall. These results show that the relationship between visual descriptors and altitude telemetry varies across recordings and that prediction error must be considered alongside detection coverage.
No takes yet. Share an insight, caveat, or question.
Galaz et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: