Local-global Fusion for Small-object Detection

Object detectors struggle with small objects, and this challenge is exacerbated in adverse weather such as fog, rain, and snow. We introduce a Local-Global Fusion (LGF) module for one-stage detectors, inserted at the C3 feature level of a RetinaNet backbone to specifically enhance small-object detection. LGF fuses a local-detail branch (depthwise convolution capturing fine spatial detail) and a global-context branch (global pooling capturing scene context) using a lightweight per-pixel, per-channel gating mechanism. Experiments on a COCO-derived dataset (clear versus synthetic adverse weather) and the ACDC real-weather benchmark show that LGF improves small-object average precision (APsmall) by approximately +1.1–1.2 points on both clear and weather conditions (such as foggy, rainy and snowy) +0.7 points over a strong RetinaNet baseline. These improvements come with no significant change in overall mAP and do not harm medium or large object detection. Ablation studies indicate that the majority of the performance gain comes from fusing local and global features, while the learned gating provides a modest refinement. The LGF module incurs only a slight increase in model complexity (+9% parameters, +3.6% FLOPs) and runtime (+6% per image latency at 640px resolution). The results demonstrate that a simple, targeted feature fusion at the C3 level can improve small-object detection robustness in weather conditions without sacrificing accuracy on larger objects.