Convolutional Neural Network Based End-to-end Object Detection

Deep learning is an important branch of machine learning.Since deep learning achieved a surprising result on image classification,people began to apply it to different domain of computer vision.Object detection is a core problem in computer vision.Recently,object detection using deep learning has achieved great success,which mainly focus on regressing the coordinates of bounding box,e.g.,Faster Region-based Convolutional Network method(Faster R-CNN),Unified,Real-Time Object Detection(YOLO)and Single Shot MultiBox Detector(SSD).Different from these methods that considering bounding box as a whole,a novel object detection method by decomposing bounding box into points based on deep networks was proposed,termed as Point Linking Network(PLN).Specifically,PLN regresses the center and corner points of bounding-box and their links using a fully convolutional network,then map the center point,corners points and their links back to multiple bounding boxes,finally an object detection result is obtained by fusing the multiple bounding boxes.PLN is naturally robust to object occlusion and flexible to object scale variance and aspect ratio variance.In the experiments,PLN with the Inception-v2 model achieves stateof-the-art single-model and single-scale results on the PASCAL VOC 2007,the PASCAL VOC 2012 and the COCO detection benchmarks without bells and whistles.Our method was also verified to be generalized by testing it in People-Art dataset.