Study Of Zero-Shot Object Detection Algorithm

Fast-developing deep learning techniques and large-scale labeled datasets have promoted the advancement of computer vision tasks,such as image classification,object detection and image segmentation.The performance of sophisticated CNN models relies heavily on labeled training data.However,the difficulty in collecting and annotating images of all the concepts that beyond daily objects hinders their application in a larger scale.To address this issue,some work has recently been done on zero-shot recognition(ZSR),which can identify completely unseen classes by using the intermediate semantic knowledge of class labels,such as the typical attributes and word vectors.Taking the classification a step further,the closely related object detection problem has remained almost unexploited in zero-shot setting,namely zero-shot detection(ZSD).Its goal is to simultaneously detect and locate each individual instance of unseen object class,even in the absence of any visual examples of those classes during the training phase.At present,several researchers have proposed some methods,trying to tackle the problem by directly using ZSR algorithms in object detection model.It works but the results are not satisfactory.Actually,researchers used to pay their attention on unseen object recognition and ignore the significance of transferability in object proposal.As a detection problem,the first to solve is how to localize the target,that is,to distinguish regions of unseen objects from backgrounds.This is the key difference between ZSD and ZSR and one of the difficulties of ZSD task.Being a challenging realistic problem,ZSD is the topic of increased interest within the community.The main work of this paper is devoted to improving the performance of zero-shot detection.In the early stage,basic research is conducted to get an overview of this field so as to grasp the core of this problem.By analyzing the internal mechanism of generic detector,we promote two novel algorithms to enhance the transferability of object proposal and optimize category prediction,respectively.The contributions of this paper are threefold as follows:i.Summarize the research on zero-shot recognition and object detection,as wellas the recent progress on zero-shot detection,and systematically define ZSDproblem and describe its process.We also highlight the critical technologies inalgorithms,such as visual feature extraction,region proposal and visual-semantic mapping,where our new algorithms are based.ii.Propose the algorithm of confidence distribution to enhance the transferabilityof zero-shot object proposal.Confidence distribution encourages the model toevaluate the likelihood of all the categories by associating the occurrence ofseen and unseen objects,making it transferable in object proposal.The methodincreases the recall of unseen objects during testing,so that zero-shot detectorsachieve better performance.iii.Propose the algorithm to optimize category prediction in zero-shot detectionmodels based on the superclass guidance.Contextual features extracted withdilated convolution layers are leveraged to predict superclass,which is aneffective guidance for final classification.Compared with predicting unseenclasses with merely semantic embeddings,this method gains improvements onthe category prediction step.