EfficientViT-Based Adaptive Feature Normalization for Cross-Domain UAV Object Detection

Abstract

Reliable small-object detection in unmanned aerial vehicle (UAV) imagery is difficult on resource-constrained platforms, and lightweight detectors often degrade when applied to data captured under different sensors, altitudes, and environmental conditions. This study investigates a lightweight framework intended to improve cross-domain robustness, combining an EfficientViT backbone with Adaptive Feature Normalization (AFN), which learns a per-channel blend of Batch and Instance Normalization. The framework also uses MixStyle for domain generalization, BiFPN-inspired multi-scale fusion, and a task-decoupled detection head. The model contains 4.48 million parameters and requires 22.73 GFLOPs at 640×640 input. Training followed a source-only, zero-shot protocol, with VisDrone2019 as the source domain and UAVDT as the target, under a two-class unified vehicle taxonomy. Evaluation covers detection accuracy (mAP), parameter count, and zero-shot cross-domain transfer, with YOLOv8n and RTMDet-tiny as baselines. After a non-finite-loss training defect was traced to DataLoader worker configuration and resolved, single-domain training reached an mAP@0.5 of 0.0385 on VisDrone2019. Ablation results identified the task-decoupled head as the most influential component. In zero-shot VisDrone2019→UAVDT transfer, MixStyle did not improve on the baseline: the two-seed mean mAP@0.5 was 0.0133 for the baseline and 0.0061 with MixStyle, although its effect was small relative to seed-to-seed variance. Under a matched protocol, the proposed model exceeded RTMDet-tiny (mAP@0.5 of 0.0238) but fell well short of YOLOv8n (0.2259). Limitations include the absence of data augmentation, unmeasured inference speed for the proposed model, and an unexecuted DroneVehicle evaluation. The study provides a transparent empirical baseline and identifies priorities for future cross-domain UAV detection research.