An Improvement of RetinaNet for Hand Detection in Intelligent Homecare Systems

In this paper, we introduce an AC-Retina approach, which uses RetinaNet as a base architecture and integrates an atrous convolution (AC) module to extract multiscale context information for hand detection in intelligent homecare (IH) systems. Given that the AC module is adopted in the feature pyramid network (FPN), rich semantic information of higher pyramid layers is added to the lower pyramid layers to improve the detection performance. The experimental results show that our AC-Retina method obtains 82.99% AP and outperforms the original RetinaNet on the Oxford hand dataset.