Recognition of Object Categories and Scenes

该研究项目旨在推进最先进的图像识别技术，使其能够识别真实和不受约束的室内和室外环境中的大量场景和对象类别，即交通场景（汽车，自行车，行人，人脸，街道标志等），城市和自然场景（建筑物，景观等），具有各种刚性和铰接物体以及纹理。如今，几乎每个人都带着数码相机，拍摄照片或短视频从未如此容易。每次重大事件发生后，广播公司都会从公众那里收到数千张图片，这些文件的注释都是手工完成的。犯罪调查人员收集了大量的视觉证据，其分类也是手工完成的。英国拥有欧洲最多的安全摄像头，但这些摄像头提供的数据很少被挖掘。此外，视觉信息的识别和解释是自主智能机器人的主要要求之一。因此，迫切需要一种可靠的识别系统，能够对大量的视觉文件进行自动分类和注释。实现这一目标的任何成功，即为专家自动确定文档浏览的优先级，都将被视为提高工作效率的明显好处。为了实现该项目的目标，必须在特征提取、类别表示和高效搜索领域取得重大进展。最近基于兴趣点的方法证明了在视觉识别的背景下处理大量类别的能力。这些方法显示了成功的场景和对象识别的有希望的方向。基于这些结果，我们提出了一种新的图像特征提取技术，该技术对背景杂波和视点变化具有鲁棒性，这是当前图像识别领域的一大挑战。这些特征将适用于在各种外观和结构水平上同时表示场景和对象，以及对象的分割。中级图像分割方法有可能提供这样的特征，并且可以在类别识别的背景下弥合兴趣点检测器和语义分割之间的差距。虽然目标是同时解决这两个问题，但在识别和分割领域之间几乎没有重叠。我们还建议引入新的层次表示，它将利用新特征的属性，并允许有效地处理大量图像类别。该模型将对不同图像属性（即强度、颜色和纹理）以及不同对象部分和视图之间的关系的多个层次中的类别进行建模。多重层次结构将允许基于与查询相关的图像线索进行从粗到精的分类。在这一领域的工作很少，提出的研究可以为图像表示问题提供新的思路。最后，采用高效的树结构和最近邻搜索技术来处理多类别学习中的大量数据。开发新颖，高效和强大的技术，可以为基本识别问题提供成功的解决方案，并推进最先进的特征提取，类别表示和数据探索，使这个项目非常具有挑战性和冒险性。该项目预计将在36个月内实现目标，它将涉及一名研究生，一名研究助理和一名首席研究员。

This research project proposes to advance state-of-the-art image recognition techniques to be able to recognize a large number ofscenes and object categories in real and unconstrained indoor andoutdoor environments i.e. traffic scenes (cars, bicycle vehicles,pedestrians, human faces, street signs etc.), urban and naturalscenes (buildings, landscapes etc.) with various rigid andarticulated objects as well as textures. Nowadaysalmost everybody carries a digital camera and taking a photo or ashort video has never been easier. Broadcasting companies receivethousands of pictures from the general public after every majorevent and the annotation of those documents is done manually. Crime investigators collect large amounts ofvisual evidence and its classification is also done manually. The UKhas the largest number of security cameras in Europe but the dataprovided by the cameras is very little explored. Furthermore,recognition and interpretation of visual information is one of themajor requirements for autonomous intelligent robots. There is therefore a dire need for a reliable recognition system capable of automatic classification and annotation of large amounts of visual documents. Any success towards achieving that goal i.e., automatic prioritizing of document browsing for experts, will be seen as a clear benefit in improvingthe efficiency of work.To fulfil the objectives of this project major progress has to bemade in the domain of features extraction, category representationand efficient search. Recent interest point based approachesdemonstrate the capability of dealing with large numbers ofcategories in the context of visual recognition. These methods showpromising directions towards successful scene and objectrecognition. Based on these results we propose to develop noveltechniques for extracting image features robust to backgroundclutter and viewpoint change, which are currently great challengesin image recognition domain. Those features will be suitable forsimultaneous representation of scenes and objects at variousappearance and structure levels as well as for segmentation ofobjects. Mid-level image segmentation methods have a potential toprovide such features and can bridge the gap between interest pointdetectors and semantic segmentation in the context of categoryrecognition. There has been little overlap between recognition andsegmentation domains although the goal is to solve both problemssimultaneously.We also propose to introduce novel hierarchical representationswhich will exploit the properties of new features and allow to dealefficiently with large number of image categories. Therepresentation will model the categories in multiple hierarchies ofvarious image attributes i.e., intensity, color and texture as wellas relations between different object parts and views. The multiplehierarchies will allow for coarse-to-fine classification based onimage cues relevant to the query. Very little work has been done inthis area and the proposed research can shed new light on imagerepresentation problems. Finally, efficient tree structures andnearest neighbor search techniques will be employed to handle largeamounts of data in multi-category learning.Developing novel, efficient and robust techniques which may providesuccessful solutions to fundamental recognition problems and advancethe state-of-the-art in feature extraction, categoryrepresentation and data exploration, make this project verychallenging and adventurous. The project is expected to achieve theobjectives within 36 months and it will involve a research student,a research assistant and the principal investigator.