DMA-Net: Decoupled Multi-Scale Attention for Few-Shot Object Detection

As one of the most important fields in computer vision, object detection has undergone marked development in recent years. Generally, object detection requires many labeled samples for training, but it is not easy to collect and label samples in many specialized fields. In the case of few samples, general detectors typically exhibit overfitting and poor generalizability when recognizing unknown objects, and many FSOD methods also cannot make good use of support information or manage the potential problem of information relationships between the support branch and the query branch. To address this issue, we propose in this paper a novel framework called Decoupled Multi-scale Attention (DMA-Net), the core of which is the Decoupled Multi-scale Attention Module (DMAM), which consists of three primary parts: a multi-scale feature extractor, a multi-scale attention module, and a decoupled gradient module (DGM). DMAM performs multi-scale feature extraction and layer-to-layer information fusion, which can use support information more efficiently, and DGM can reduce the impact of potential optimization information exchange between two branches. DMA-Net can implement incremental FSOD, which is suitable for practical applications. Extensive experimental results demonstrate that DMA-Net has comparable results on generic FSOD benchmarks, particularly in the incremental FSOD setting, where it achieves a state-of-the-art performance.