DMSANet: Dual Multi Scale Attention Network

AbstractAttention mechanism of late has been quite popular in the computer vision community. A lot work done to improve performance network, although almost always it results increased computational complexity. In this paper, we propose a new attention module that not only achieves best but also lesser parameters compared most existing models. Our can easily be integrated with other convolutional neural networks because its lightweight nature. The proposed network named Dual Multi Scale Attention Network (DMSANet) is comprised two parts: first part used extract features at various scales and aggregate them, second uses spatial channel modules parallel adaptively integrate local their global dependencies. We benchmark our for Image Classification on ImageNet dataset, Object Detection Instance Segmentation both MS COCO dataset.KeywordsAttention moduleImage classificationObject detectionInstance segmentation