Reinforcement learning-based interactive video search

• Main Authors: MA, Zhixin, WU, Jiaxin, HOU, Zhijian, NGO, Chong-wah
• Format: text
• Language: English
• Published: Institutional Knowledge at Singapore Management University 2022
• Subjects: Feature enhancement; Interactive video retrieval; Query understanding; Reinforcement learning; Artificial Intelligence and Robotics; Graphics and Human Computer Interfaces
• Online Access: https://ink.library.smu.edu.sg/sis_research/7503; https://ink.library.smu.edu.sg/context/sis_research/article/8506/viewcontent/reinforcement_learning.pdf
• Institution: Singapore Management University

Despite the rapid progress in text-to-video search due to the advancement of cross-modal representation learning, the existing techniques still fall short in helping users to rapidly identify the search targets. Particularly, in the situation that a system suggests a long list of similar candidates, the user needs to painstakingly inspect every search result. The experience is frustrated with repeated watching of similar clips, and more frustratingly, the search targets may be overlooked due to mental tiredness. This paper explores reinforcement learning-based (RL) searching to relieve the user from the burden of brute force inspection. Specifically, the system maintains a graph connecting shots based on their temporal and semantic relationship. Using the navigation paths outlined by the graph, an RL agent learns to seek a path that maximizes the reward based on the continuous user feedback. In each round of interaction, the system will recommend one most likely video candidate for users to inspect. In addition to RL, two incremental changes are introduced to improve VIREO search engine. First, the dual-task cross-modal representation learning has been revised to index phrases and model user query and unlikelihood relationship more effectively. Second, two more deep features extracted from SlowFast and Swin-Transformer, respectively, are involved in dual-task model training. Substantial improvement is noticed for the automatic Ad-hoc search (AVS) task on the V3C1 dataset.