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

Xijun Xie,Feifei Lee,Qiu Chen

DOI: https://doi.org/10.2139/ssrn.4180097

2022-01-01

SSRN Electronic Journal

Abstract:As one of the most important fields in computer vision, object detection has made significant development in recent years. Generally, object detection requires a large number of labeled samples for training, but it is not so easy to collect and label samples in many special fields. Therefore, general detectors will face the problem of overfitting and poor generalization ability when recognizing unknown objects if there are few samples. With the continuous maturity of few-shot classification (FSC) methods, it is also applied to few-shot object detection (FSOD) tasks. However, many FSOD methods cannot make good use of support information and deal with the potential problem of information relationship 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 Decoupled Multi-scale Attention Module (DMAM), consisting of three main parts: multi-scale feature extractor, multi-scale attention module, and decoupled gradient module (DGM). DMAM carries out multi-scale feature extraction and layer-to-layer information fusion, which can utilize 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 DMA-Net has comparable results on generic FSOD benchmarks, especially in the incremental FSOD setting where it achieves state-of-the-art performance.