A rate-constrained video descriptor based on the information bottleneck principle

大多数现有的视觉分析系统都是基于“压缩-然后分析”范式，其中分析是从已经压缩的图像或视频序列中提取的视觉特征上执行的。为了利用移动设备日益增长的计算能力，黑莓的一个研究团队正在开发一种智能视频处理系统，该系统基于一种被称为“分析-然后压缩”的新范式。在这个新范例中，首先从原始视觉数据中提取特征，然后压缩，为进一步处理、传输或存储做准备。然而，传统的视觉数据压缩技术在压缩从原始视频序列中提取的视觉特征时表现不佳。通过几次技术讨论，我们已经确定了这个问题的根源是面向重建的压缩和面向分析的压缩的不兼容性。本课题旨在通过设计基于信息的特征压缩算法来解决这一问题

Most existing visual analysis systems are based on the "Compress-Then-Analyze" paradigm, where the analysis is performed on visual features extracted from the already-compressed images or video sequences. In order to exploit the increasing computational capabilities of mobile units, a research team at BlackBerry is developing a smart video processing system according to a new paradigm known as "Analyze-Then-Compress". In this new paradigm, features are first extracted from the raw visual data and then compressed in preparation for further processing, transmission or storage. However, it has been observed that the legacy visual data compression techniques have poor performance when used to compress the visual features extracted from the raw video sequences. Through several technical discussions, we have identified that the source of this problem is the incompatibility of reconstruction-oriented compression and analysis-oriented compression. The proposed project aims to tackle this problem by designing a feature compression algorithm based on the information bottleneck principle. In contrast to the conventional compression algorithms that strive to minimize the reconstruction distortion, the new algorithm is designed to preserve the relevant information contained in the compressed features and consequently has the potential to achieve superior performance in a wide range of visual analysis tasks. This project will provide a cutting-edge video processing technology that will enable BlackBerry to become a major player in developing the new video processing and analysis standard, currently being explored by an ad-hoc MPEG group on Compact Descriptors for Video Analysis. In addition to that, it is expected to build the foundation of a long-term collaboration between BlackBerry and McMaster University in the areas of data mining, privacy protection, and health-related information technologies. On the educational front, the project provides an extraordinary opportunity for students and researchers to acquire valuable experience, skills, and knowledge that are in line with the needs of academia and Canadian information technology industries in the big data era.