/AVSD

Aerial Video Segmentation Dataset

Aerial Video Segmentation Dataset (AVSD)

Video segmentation refers to the task of partitioning pixels that exhibit homogeneous appearance and motion into coherent spatial-temporal groups, which can be applied to a wide variety of practical aerial applications, such as landscape detection and recognition, ground scene classification, and urgent safe landing. However, few works about aerial video segmentation for higher-altitude platforms are developed in the current literatures due to the lack of common evaluation dataset. Therefore, we collected a new aerial video dataset, which exhibits the simple and challenging conditions of video complexity and motion, for the evaluation of aerial video segmentation algorithms.

Description

ImgData_0351 ImgData_1702 ImgData_4006
ImgData_0351_gt ImgData_1702_gt ImgData_4006_gt

The dataset consists of 10 diverse sequences and total 525 images, of which 131 images are annotated with pixel-wise segmentation labels. The videos are all captured at 12 fps and 1280*1024 spatial resolution. The images contain a variety of man-made and natural “objects” whose appearance varies across sequences and comprises homogenous and textured parts. We select 6 sequences to be annotated by experts. The primary subjects of the aerial videos are mainly various land surfaces and we define six most common types: bare land, grassland, forest, building, road, and vehicles.

Each folder contains video sequence with same ground scene. For each sequence, there are following files:

  • bmp folder: *.bmp, initial video image;
  • gt_rgb folder: *.bmp, RGB masks of segmentation for visualization;
  • gt_label folder: *.bmp, index masks of segmentation ;
  • gt_mat folder: *.mat, index masks of segmentation .

Download

Representative examples: [GoogleDrive] [BaiduYunPan]

Full Resolution: [GoogleDrive] [BaiduYunPan(code: e4vu)]

Contact

If you have any question, please contact Yufeng Wang (wyfeng@buaa.edu.cn).

Citation

If you find this dataset useful, please cite the following publication: