Pothole Videos
Description
Contains a video dataset of road potholes for semantic segmentation purposes. Data is taken from a point of view perpendicular to the road. Contains a total of 619 videos (image, mask) with an aspect ratio of 1:1, 1080x1080 resolution and 48 image frames for each video. The data was divided manually with a ratio of 60/20/20 with 372 for training, 124 for validation, and 123 for test. Ground truth is made manually using Adobe After Effects. The videos in this dataset were captured in Hulu Sungai Tengah Regency, Indonesia. If you use this dataset, please cite the following paper: M. Ihsan, M. A. Amrizal, and A. Harjoko, “A pothole video dataset for semantic segmentation,” Data in Brief, vol. 53, 2024.
Files
Steps to reproduce
The video was taken using the Xiaomi Mi 10t smartphone. Top-down camera position with hole. The distance between the camera and the hole is around 130 cm. The video data was adjusted by cropping to have a 1:1 aspect ratio with a resolution of 1080x1080 pixels. The video duration is also adjusted so that each existing video data has 48 image frames each with a duration of 2 seconds. To create a ground-truth mask, manual annotation is done on one of the frames then the rest is created automatically using the Adobe After Effects mask tracking tool.