DSNet: Efficient Lightweight Model for Video Salient Object Detection for IoT and WoT Applications

Abstract

The most challenging aspects of deploying deep models in IoT and embedded systems are extensive computational complexity and large training and inference time. Although various lightweight versions of state-of-the-art models are also being designed, maintaining the performance of such models is difficult. To overcome these problems, an efficient, lightweight, Deformable Separable Network (DSNet) is proposed for video salient object detection tasks, mainly for mobile and embedded vision applications. DSNet is equipped with a Deformable Convolution Network (DeCNet), Separable Convolution Network (SCNet), and Depth-wise Attention Response Propagation (DARP) module, which makes it maintain the trade-off between accuracy and latency. The proposed model generates saliency maps considering both the background and foreground simultaneously, making it perform better in unconstrained scenarios (such as partial occlusion, deformable background/objects, and illumination effect). The extensive experiments conducted on six benchmark datasets demonstrate that the proposed model outperforms state-of-art approaches in terms of computational complexity, number of parameters, and latency measures.