Deep Active Semi-supervised Learning For Object Detection

Object detection,which is a classical task in the field of computer vision,has a significant impact on many fields ranging from automatic driving,video surveillance,human-computer interaction to face detection.But the complexity and diversity of environment make it a challenging task.Recently it has made a significant breakthrough due to the development of deep learning technology.However,the performance improvement benefits from large amounts of annotated data.However,data annotation is quite expensive,and meanwhile large collections of unlabeled data leaves unexplored.Active learning is a machine learning strategy that aims to promote the efficiency of data annotation by selecting the most informative data to be annotated.Meanwhile,semi-supervised learning focuses on exploiting unlabeled data to boost the performance.Combining the advantages of both techniques,the task named deep active semi-supervised learning for object detection attracts researchers' attention nowadays.The key issue is to determine the mechanism of measuring the informationcss of unlabeled data.In addition,how to integrate both active and semi-supervised learning in an efficient way should also be considered.First,this thesis explores the mechanism of measuring the informationess of images in the active learning part.We introduce several uncertainty sampling strategies into the object detection task to define the uncertainty of candidate objects and then merge the uncertainties of all candidates to generate the uncertainty of the corresponding image.Second,this thesis further proposes an active learning method for object detection based on the extraction of object instance.It introduces object instance to bridge candidate objects:and images to define the informationess of image more accurately.It extracts object instances from an image via a cluster-based method,and then uses the corresponding linked candidate objects to construct a committee for every object instance.Finally,it uses a query-by-committee method to measure the uncertainty of classification and localization for every object instance.In addition,we ensure the diversity of batch-mode sampling to handle the images,in which no object instance has been extracted.Third,this thesis further integrates both active learning and semi-supervised learning techniques.We propose an active semi-supervised learning method for object detection based on the extraction of object instance.It queries the high-uncertainty data to be human-annotated using the above-mentioned active learning method but generates the pseudo-label for those low-uncertainty data,which is inspired by the idea of self-training.We conduct several experiments on two publicly available datasets.Our experiments demonstrate that the methods based on uncertainty sampling are better than that based on passive learn-ing(random sampling).It shows the power of active learning methods in object detection task.Furthermore,our experiments demonstrate that the proposed methods based on the extraction of object instance boost the performance significantly compared to the baseline method(uncertainty sampling),and also achieve state-of-the-art performance.