Weakly Supervised Object Detection For Specific Scene Images

As one of the most concerned issues in computer vision,object detection has achieved a lot of great achievements in recent years,and its performance and detection accuracy have been greatly improved.Most of the existing methods for object detection are under supervised conditions,that is,a large number of manual annotations of the images of the object proposal bounding box are needed to learn the appearance features of the object and the accurate location information.In this paper,under little annotation cost,with the help of spatial temporal information and deep convolutional neural network,an effective method of weakly supervised object detection is proposed.In this paper,considering the problem of time-consuming manual labeling of video,we use the motion information in video to replace the strong artificial supervising information,and propose a weakly supervised object detection model based on video learning.In some special scenes where the foreground objects are sparse,we can even train an object detection model with the same detection accuracy as supervised method.In view of the fact that the original weakly supervised object detection model based on video learning has a poor foreground extraction result on some videos,leads to the low final detection accuracy.Based on the original model,this paper mainly makes two improvements: 1)the foreground extraction part uses a more robust optical flow method to extract the foreground.This method combines more video action information to make the initial pseudo-label extraction more accurate;2)use improved KCF algorithm for multi-target tracking to optimize the initial pseudo-label,which makes the new weakly supervised model detect more correct objects than before.The above weakly supervised object detection model mainly uses video information.Now we attempt to use the knowledge of the supervised object detection model to transfer them to other videos from similar scenes for object detection.The main idea is to translate object detection into the classification of foreground and background region proposals,and use the residual transfer network to solve the unsupervised domain adaptation for transfer learning.In the final test,our improved residual transfer network can outperform the fine-tuning network and the original weakly supervised object detection model by transfering a supervised model to improve performance accuracy.