本文为中国香港中文大学（作者：XIAO, Tong）的博士论文，共98页。

人物识别是计算机视觉中的一个基本问题，其目的是从数码照片库中找到感兴趣的目标人物。它是许多实际应用程序中的关键组件，包括智能手机应用程序、自动驾驶汽车、家庭安全系统和智能监控摄像头。由于深度学习研究的发展和大规模标注数据集的出现，深度神经网络现在能够识别成千上万的对象类别。然而，由于缺少足够大的数据集来监督模型训练，人物识别仍然具有挑战性。此外，现有的小型数据集通常都有自己的图像偏差，这使得学习一个可以概括所有这些领域的模型变得很困难。同时，现有的研究大多简化了问题的设置，使得研究方法与实际应用之间存在一定的差距。

本文从三个方面来解决这些问题，使人物身份识别能够扩展到现实世界的数据和应用。首先，我们提出了一个半监督的深度学习框架，该框架使用带噪声的标记数据，而不是标注良好的数据。我们收集了一个带有错误标注的大规模衣物数据集，从中我们可以学习到有助于识别人物衣物的良好表示。其次，我们发展了一个联合的单任务学习算法和一个域引导的dropout技术，从多个具有领域偏差的人物识别数据集中学习单一模型。它使我们能够集体使用社区中不同人员提供的数据。最后，我们将重点放在更真实的问题设置上，即在整个场景图像中找到目标人物。我们开发了一个统一的框架，结合了人的检测、识别以及损失函数，有效地实现了识别模型的训练。

Human identification, which aims at findinga target person of interest from a gallery of digital photos, is one of thefundamental problems in computer vision. It is a key component in manyreal-world applications including smart phone apps, self-driving cars, homesecurity systems, and intelligent surveillance cameras. Thanks to thedevelopment of deep learning research and large-scale well annotated datasets,deep neural networks are now capable of recognizing thousands of objectcategories. However, human identification is still challenging because it lacksa dataset large enough to supervise the model training. Moreover, existingsmall datasets usually have their own image biases, which makes it hard tolearn a single model that generalizes over all these domains. Meanwhile, mostof the existing research simplified the problem setting, which leaves a gapbetween research approaches and practical applications. In this dissertation weaddress these challenges from three aspects to make human identificationscalable to real-world data and applications. First, we propose asemisupervised deep learning framework that uses noisy-labeled rather than wellannotated data. We collect a large-scale clothing dataset with noisyannotations, from which we can learn good representations for clothes that helprecognize human. Second, we develop a joint single task learning algorithm anda domain guided dropout technique to learn a single model from multiple humanidentification datasets with domain biases. It enables us to collectively usethe data contributed by different people in the community. At last, we focus onthe more realistic problem setting that finds a target person in whole sceneimages. We develop a unified framework that combines person detection andidentification, as well as a loss function that trains the identification modeleffectively.

1. 引言

2. 深度学习基础

3. 人物识别背景

4. 从噪声标注中学习特征

5. 多域人物的重新识别

6. 从人物重新识别到人物搜索

7. 结论

更多精彩文章请关注公众号：