An Innovative Approach to the Dual Problems of High-Resolution Input and Video Object Detection
ABOUT THE PROJECT
At a glance
Camera-sensor feeds for an autonomous vehicle can support a stream of high-resolution UHD (3840 × 2160) videos at 30 frames-per second. This high resolution offers the promise of accurate object detection, most pressingly for objects that are difficult or impossible to recognize at lower resolutions. In addition, video offers many advantages over standalone images. This work explores how to leverage attention and redundancy across space and time to more efficiently and accurately detect objects in high-resolution video. Our group has been actively developing more efficient neural architectures for single frame processing of high-resolution data as well as temporal fusion for video data. However, we have not yet connected these projects.
|object detection, video, high-resolution, fast inference|