Abstract
Scale variation presents a significant challenge in object detection. To address this, multi-level feature fusion techniques have been proposed, exemplified by methods such as the feature pyramid network (FPN) and its extensions. Nonetheless, the input features provided to these methods and the interaction among features across different levels are limited and inflexible. In order to fully leverage the features of multi-scale objects and amplify feature interaction and representation, we introduce a novel and efficient framework known as a multi-resolution and semantic-aware bidirectional adapter (MSBA). Specifically, MSBA comprises three successive components: multi-resolution cascaded fusion (MCF), a semantic-aware refinement transformer (SRT), and bidirectional fine-grained interaction (BFI). MCF adaptively extracts multi-level features to enable cascaded fusion. Subsequently, SRT enriches the long-range semantic information within high-level features. Following this, BFI facilitates ample fine-grained interaction via bidirectional guidance. Benefiting from the coarse-to-fine process, we can acquire robust multi-scale representations for a variety of objects. Each component can be individually integrated into different backbone architectures. Experimental results substantiate the superiority of our approach and validate the efficacy of each proposed module.