Efficient multi-scale representation learning via global-local collaborative network for object detection

Multi-scale object detection remains challenging due to feature homogenisation, which degrades feature discriminability across objects of different scales. To address this issue, we propose GLociNet, a Global-Local Collaborative Network for robust object detection. Specifically, a Gated Efficient Layer Aggregation Network (Gated-ELAN) is developed to enhance feature aggregation through adaptive gating, suppressing redundant information while preserving discriminative representations. Building upon this, a Multi-Scale Feature Enhancement (MS-FE) module is introduced to enrich feature diversity across scales. To further improve scale adaptability, an Axis-Alternating Attention (A3) module is designed to capture scale-aware contextual dependencies. In addition, an ultra-lightweight Vision Transformer (ViT) is incorporated as an auxiliary backbone to complement convolutional representations with long-range contextual modelling. Extensive experiments on benchmark datasets demonstrate the effectiveness of GLociNet. In particular, GLociNet at N-scale (nano-scale) achieves 42.3% Average Precision (AP) on the Microsoft Common Objects in Context (MS COCO) dataset with only 2.9M parameters, outperforming the state-of-the-art detectors.