Object Detection Based On Contextual Semantic Information

With increasing demand of running Convolutional Neural Networks(CNNs)on mobile devices,real-time object detection has made great progress in recent years.However,modern approaches usually compromise detection accuracy to achieve realtime inference speed.Some light weight top-down CNN detectors suffer from problems of spatial information loss and lack of multi-level semantic information.In this paper,we introduce an efficient CNN architecture,the Multi-level Semantic Pyramid Network(MSPNet),for real-time object detection on devices with limited resource and computational power.The proposed MSPNet consists of two main modules to enhance spatial details and multi-level semantic information.The multiscale feature fusion module integrates different level features to tackle the problem of spatial information loss.Meanwhile,a light weight multi-level semantic enhancement module is developed which transforms multiple layer features to strengthen semantic information.The proposed light weight object detection framework has been evaluated on CIFAR-100,PASCAL VOC and MS COCO datasets.Experimental results demonstrate that our method achieves state-of-the-art results while maintains a compact structure for real-time object detection.