Understanding video through the lens of language

Thesis

The increasing abundance of video data online necessitates the development of systems capable of understanding such content. However, building these systems poses significant challenges, including the absence of scalable and robust supervision signals, computational complexity, and multimodal modelling. To address these issues, this thesis explores the role of language as a complementary learning signal for video, drawing inspiration from the success of self-supervised Large Language Models (LLMs) and image-language models.

First, joint video-language representations are examined under the text-to-video retrieval task. This includes the study of pre-extracted multimodal features, the influence of contextual information, joint end-to-end learning of both image and video representations, and various frame aggregation methods for long-form videos. In doing so, state-of-the-art performance is achieved across a range of established video-text benchmarks.

Second, this work explores the automatic generation of audio description (AD) – narrations describing the visual happenings in a video, for the benefit of visually impaired audiences. An LLM, prompted with multimodal information, including past predictions, and pretrained with partial data sources, is employed for the task. In the process, substantial advancements are achieved in the following areas: efficient speech transcription, long-form visual storytelling, referencing character names, and AD time-point prediction. Finally, audiovisual behaviour recognition is applied to the field of wildlife conservation and ethology. The approach is used to analyse vast video archives of wild primates, revealing insights into individual and group behaviour variations, with the potential for monitoring the effects of human pressures on animal habitats.

Authors: Bain, M

• DOI: 10.5287/ora-gpn5kbm4e
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: multimodal; video understanding; vision & language; deep learning
• Subjects: Deep learning (Machine learning); Natural language processing (Computer science); Computer vision