Abstract:
Detecting objects in high-resolution videos in real-time has proven extremelydifficult. The large size of high-resolution images makes traditional object- detection methods impractical within the short time period between video frames. Previous approaches to this problem have relied on techniques which select re- gions to analyze within a frame through pyramid pooling and attention pipelin- ing. We propose a novel approach which uses historical location information from earlier frames to inform decisions relating to specific regions in later frames. When run on a dataset of 4k videos, this approach has shown significant improve- ments in temporal efficiency without reducing accuracy over both attention- based methods and more naı̈ve approaches. At lower frame rates, this algorithm is able to process high-resolution video data in real time and can be used to monitor video camera footage without human intervention.
