Web-Scale Semantic Image and Video Understanding

点击翻译按钮获取中文摘要

Visual recognition is a sub-field of computer vision which centers on building algorithms that can automatically and intelligently recognize, catalog and understand image/video content. Significant progress in accuracy and scope has been made on recognition problems in recent years, driven by powerful machine learning algorithms (e.g., deep learning), large labeled datasets and novel problem definitions. However, current approaches lack capabilities that permit accurate fine-grained detailed scene understanding. Most algorithms are limited to coarse image- or video-level interpretations (e.g., choosing among 1,000 noun categories, 200 action classes, or 18 candidate answers for a visual question). Such understanding is useful, but at the same time is still very limiting in enabling the breadth of potential applications ranging from media curation to autonomous navigation. It is difficult to quantify what level of performance or scale is necessary for visual recognition to be broadly successful. I believe the next transformative milestone to be - detailed scene understanding at the accuracy of the current coarse models. This requires developing capabilities of recognizing fine-grained object categories, numbering in 100,000, and spatio-temporal (predicate) relations among the objects and elements of the scene, counting in 1,000. The former would give ability to recognize nearly every object/noun; the latter would enable situated contextual reasoning critically important for AI. My long term research objective is to develop such accurate detailed models for visual understanding at scale; models that can describe and localize objects and people, reason about their spatial and functional relationships, their actions and interactions. This proposal tackles three fundamental sub-challenges to achieving this objective in the corresponding research threads: 1. The ever growing fine-grained set of classes requires development of novel data efficient learning algorithms. As the categories to recognize become more specific, the amount of data per category decreases (e.g., there are millions of car images, but few of 1957 Jaguar XKSS). We will build on our recent work where we developed the only method to date capable of recognizing up to 310,000 categories. 2. Moving beyond recognition of isolated objects, requires reasoning about structures relating objects, people, and scene elements in space and time. Rich flexible structured models will be developed to enable such reasoning. 3. To alleviate the black-box nature of existing architectures, not suitable for decision-critical tasks, we will develop algorithms that enable interpretability and more human-like introspective reasoning. Importantly, the program will also focus on applying the developed algorithms to specific recognition problems relevant for media search/retrieval, augmented reality and medical imaging.