Parameter-efficient small object detection via P2 head co-design for UAV remote sensing images

Abstract

Small object detection from UAVs is fundamentally limited by the resolution loss imposed by standard P3–P5 feature pyramids, which leave nearly 47% of VisDrone annotations under-resolved. We address this by co-designing a stride-4 P2 detection head with shared Distribution Focal Loss and channel scaling, reducing parameters by 7.4% while improving mAP50 by 2.95% and AP_S by 3.19 pp on VisDrone-DET2019. On the RTX 3090 (FP16), the model runs at 128.9 FPS. A channel-selective variant (Variant G, 8.46M) recovers large-object AP by up to 9.7 pp; on the RK3588 NPU (INT8), it achieves 8.4 FPS within 9.70 MB NPU memory. The results indicate that resolution and parameter efficiency in UAV detection are not inherently at odds.