Weakly Supervised Real-time Object Detection Based On Pseudo-Annotation Generator

Object detection has always been an extremely important branch in the field of computer vision,especially in recent years,deep learning technology has made rapid progress,and object detection has also attracted a new upsurge.However,the current object detection method is usually fully-supervised,which means that the data set's annotation not only needs to include the category of the object but also the position and size of the object,but in the application environment,such detailed manual labeling of data sets consumes a lot of manpower and material resources,and the cost is extremely high.For this reason,weakly-supervised object detection methods have begun to attract more and more researchers' attention.Compared with fullysupervised data,weakly supervised data sets only need to include a small amount of annotations.For example,the data set only need image-level annotations.In current detection of weakly labeled objects,the commonly used method is based on the method of region proposals.In the object detection,image candidate regions are generated first,and then these candidate regions are classified.However,it takes a large amount of computing resources to generate these regions proposals.In the real application environment,the timeliness of the object detection is very demanding,and even run in real time.In view of the above reasons,this research is on a real-time object detector in weakly-supervised.To achieve this goal,we present a weakly supervised real-time object detection based on pseudo-annotation generator with saliency guiding.In this paper,a class saliency map with attention mechanism is introduced.Through the learning of the region of interest in the image,the problem of lack of position information in the weakly-supervised is solved..Firstly learning a classifier to get the derivative of the feature maps with respect to the class score by back-propagation,and then the generated class saliency maps with discriminative information,Finaly the class saliency maps are used to generate pseudo-annotations,which are approximations of the ground truth annotations.Use these approximate pseudoannotations to train our object detection network.In order to meet the timeliness requirement of the actual scene,this paper chooses a regression-based object detection method as our object detector.In this paper,we use the PASCAL VOC dataset for training and testing.The experimental results show that the proposed method can be similar to the state of the art,its mean averange precision is 33.9%(m AP),and the detection speed achieves a breakthrough.Frame per second improves to 45.Experiments show that the proposed method is effective and successfully implements a weakly-supervised real-time object detection.