Dataset for Unsupervised Region Growing Dynamic Object Segmentation in Turbulence (URG-T)

*Arizona State University,
&Clemson University,
^George Mason University

Dataset visualization with segmentation ground truth annotation.

Dataset introduction

URG-T consists of 20 videos, all collected outdoors in hot weather using long focal length settings. All videos contain instances of moving objects, such as vehicles, aircraft, and pedestrians. We manually annotate the video to provide per-frame ground truth masks for segmenting moving objects. Specifically, we use a Nikon Coolpix P1000 to capture the videos. The camera has an adjustable focal length of up to 539mm (125X optical zoom), which is equivalent to 3000mm focal length in 35mm sensor format. We record videos with a resolution of 1920X1080. In total, our dataset has 20 videos with 934 frames. We annotate moving objects in each video frame using the Computer Vision Annotation Tool (CVAT). Our dataset is the first of its kind to provide a ground truth moving object segmentation mask in the context of long-range turbulent video. URG-T is specifically designed for motion segmentation tasks, but it can be used for other tasks as well (e.g., turbulent video restoration).

BibTeX

@INPROCEEDINGS{saha,
  author={Ripon Kumar Saha, Dehao Qin, Nianyi Li, Jinwei Ye, Suren Jayasuriya},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, 
  title={Turb-Seg-Res: A Segment-then-Restore Pipeline for Dynamic Videos with Atmospheric Turbulence}, 
  year={2024},
  }