Video Semantic Segmentation and Tracking in Low Light Videos

点击翻译按钮获取中文摘要

Natural History filmmaking often requires filming in low-light environments which results in noisy, low-resolution and low-contrast videos. Due to the quality degradation, it can be difficult to do computer vision tasks like object detection and object tracking. These computer vision tasks are beneficial for production as they can automate and aid content acquisition and post-production. This research proposes to investigate various methods for video semantic segmentation and tracking in low-light videos to improve creative media production workflows. This research aims to explore the following research questions and problems: 1. Developing a semantic segmentation approach for object detection in the context of low-light videos 2. Developing a method that tackles that task of Video Semantic Segmentation (VSS) and Video Object Tracking (VOT) in low-light videos 3. Investigating the impacts of using various low-light enhancement techniques to improve semantic segmentation and tracking 4. Addressing the challenge of occlusion in object tracking specifically for low-light videos