Asymmetric Modality Domain Randomization for Drone-Based RGB–Thermal Tiny Person Detection Under Controlled Synthetic Sensor Shifts

Abstract

Drone-based red–green–blue (RGB)–thermal tiny person detection must account for the different appearance statistics of visible and thermal sensing. We propose Asymmetric Modality Domain Randomization (AMDR), a training-only strategy that applies modality-conditioned perturbations while preserving image pairing, geometry, annotations, and the deployed detector graph. Using a fixed Quality-aware RGBT Fusion Detector (QFDet*) baseline, we compare AMDR with Operator-Matched Symmetric Domain Randomization (OM-SDR) across eight training seeds. With a fixed epoch-12 checkpoint, AMDR achieves a mean mAP50 of 46.89% with a standard deviation of 0.53 percentage points on RGBTDronePerson, exceeding Base and OM-SDR by 2.40 and 0.96 percentage points, respectively. Across the main 12-condition controlled synthetic-shift suite, both randomized strategies exceed Base in condition-wise mean mAP50, while their overall averages are nearly identical. A separate unseen synthetic thermal column-bias evaluation further probes a spatially structured corruption family absent from training. The weaker VTUAV-det result indicates dataset-dependent behavior rather than universal improvement. AMDR introduces no additional inference-time parameters, floating-point operations (FLOPs), or network branches.