Urban Drone Dataset(UDD)

View project on GitHub

Urban Drone Dataset(UDD)


0.1 Dataset Overview

This is a collection of drone image Dataset collected at Peking University, Huludao city, Henan University and Cangzhou city.

example of UDD: visual_color visual_mask

Class Definitions

  • UDD5
Class Gt Label RGB Suffix
Vegetation 0 (107,142,35) _t.png
Building 1 (102,102,156) _b.png
Road 2 (128,64,128) _r.png
Vehicle 3 (0,0,142) _v.png
Other 4 (0,0,0) N/A
  • UDD6 (Released on 28 Jun 2020)
Class Gt Label RGB Suffix
Other 0 (0,0,0) N/A
Building 1 (102,102,156) _b.png
Road 2 (128,64,128) _r.png
Vegetation 3 (107,142,35) _t.png
Vehicle 4 (0,0,142) _v.png
Roof 5 (70,70,70) _roof.png


Date log
2018.03.15 repo init
2018.03.23 UDD-3 released
2019.11.04 UDD-5 released
2020.06.28 UDD-6 released. Beware of the changing Gt Label!!

now UDD-6 is on air (Vegetation, Building, Road, Vehicle, Roof and Other)! See Download Link below.

This Dataset is only for non-commercial use.


If you benefit from UDD, please cite our paper:

  title={Large-scale structure from motion with semantic constraints of aerial images},
  author={Chen, Yu and Wang, Yao and Lu, Peng and Chen, Yisong and Wang, Guoping},
  booktitle={Chinese Conference on Pattern Recognition and Computer Vision (PRCV)},

1.Labeling Policy (instruction included)

1.0 Vegetation

    1. enter photo shop,press alt+F9 to open Action menu,load action script “ps-annotation.atn” selection
    1. open the src url, and press CTRL+F2,a raw mask of vegetation would be generated selection
    1. adjust the selected area by hand(lasso is recommended, just press shift/alt and drag the mouse)
    1. then press CTRL+F3 to generate bitmap, save it by “_t.png” suffix,”DJI_0285_t.png”,e.g.

Annotation example vegetation

Chinese version of annotation instruction

1.1 Building

    1. new a black layer, using polygon lasso to select building and fill it with black
    1. press CTRL+F3 to generate bitmap, save it by “_b.png” suffix,”DJI_0285_b.png”,e.g.

example of annotated result Building

1.2 Other classes

    1. After filled ROI with black, press CTRL+F3 to generate bitmap. Remember to save it by suffix(see Class Definitions above)

2. Directory Naming Policy

/src origin source image

/gt ground truth

/gt_class groundtruth split by classes

/ori annotation raw result(subfolders containing annotated '_t.png', '_b.png', etc. are all here)

/visualization visualization result

you can name your directories arbitrarily. Just keep them corresponding to envs in main.m

3. Scripts

  • script/main.m

    Processing with raw annotated result. You can DIY your ground truth label here.


visual_mode = 0; % 1 to run gtVisual.m
visual_resizerate=0.25; % downsample rate to accelerate
split_mode = 1; % 1 to run gtSplit.m
split_visualmode = 0;  % 1 to run visualization.m

To visualize the ground truth map.

To generate some split map

After running main.m, you can see the visualization result in/visualization by running this script.


view_mode = 1; % 0 for automatic, 1 for manual

run this to prepare train.txt,val.txt for training in tensorpack.

Convert JPG to PNG.

4. Acknowledgements

Sincerely tribute to all companions who contributed to this Dataset: Xiao Deng(邓枭)Youpeng Gu(顾友鹏)Jianyuan Guo(郭健元)Chen Hou(侯忱)Zhao Jin(金朝)Boning Song(宋博宁)You’er Wen(文佑尔)Yang Yao(姚洋)Kangrui Yi(易康睿)Haotian Zhou(周昊天)Youkun Wu(吴有堃)Xupu Wang(王旭普)Tongwei Zhu(朱彤葳)Zebin Wang(王泽斌)