GENERATING OPEN WORLD DESCRIPTIONS OF VIDEO USING COMMON SENSE KNOWLEDGE IN A PATTERN THEORY FRAMEWORK

摘要

The task of interpretation of activities as captured in video extends beyond just the recognition of observed actions and objects. It involves open world reasoning and constructing deep semantic connections that go beyond what is directly observed in the video and annotated in the training data. Prior knowledge plays a big role. Grenander's canonical pattern theory representation offers an elegant mechanism to capture these semantic connections between what is observed directly in the image and past knowledge in large-scale common sense knowledge bases, such as Concept-Net. We represent interpretations using a connected structure of basic detected (grounded) concepts, such as objects and actions, that are bound by semantics with other background concepts not directly observed, i.e., contextualization cues. Concepts are basic generators and the bonds are defined by the semantic relationships between concepts. Local and global regularity constraints govern these bonds and the overall connection structure. We use an inference engine based on energy minimization using an efficient Markov Chain Monte Carlo that uses the Concept-Net in its move proposals to find these structures that describe the image content. Using four different publicly available large data-sets, Charades, Microsoft Visual Description Corpus (MSVD), Breakfast Actions, and CMU Kitchen, we show that the proposed model can generate video interpretations whose quality is comparable or better than those reported by state-of-the-art approaches, such as different forms of deep learning models, graphical models, and context-free grammars. Apart from the increased performance, the use of encoded common sense knowledge sources alleviate the need for large annotated training data-sets and help tackle any imbalance in the data through prior knowledge, which is the bane of current machine learning approaches.