UMAD-Dataset-Usage-Guide-Doc
September 6, 2024 ยท View on GitHub
You can download the UMAD-1.0 dataset from Google Drive, 10.73GB.
We may update the UMAD dataset in the future, but currently, UMAD-1.0 is the latest version.
Dataset Directory Structure
UMAD # UMAD Dataset folder
|
--- UMAD-1.0 # Version
| |
| ---- data # image data
| | |
| | ---- 1.N6 # Scene
| | | ---- 00 # Sequece
| | | | ---- G1 # reference image
| | | | | ---- rgb # 00000000.jpg - xxxxxxxx.jpg
| | | | ---- Q1 # query image
| | | | | ---- rgb # 00000000.jpg - xxxxxxxx.jpg
| | | ---- 01
| .
| .
| | | ---- N
| | ---- 2.Bridge
| .
| .
| | ---- 3.Central-Avenue
| .
| .
| | ---- 4.Border-Road-1
| .
| .
| | ---- 5.Border-Road-2
| .
| .
| | ---- 6.N2
| .
| .
| |
| ---- gt # Ground truth binary mask .png files
| | ---- 1.N6
| | | ---- 00 # Sequece
| | | | ---- rgb # 00000000.png - xxxxxxxx.png
| | | ---- 01
| .
| .
| | | ---- N
| | ---- 2.Bridge
| .
| .
| | ---- 3.Central-Avenue
| .
| .
| | ---- 4.Border-Road-1
| .
| .
| | ---- 5.Border-Road-2
| .
| .
| | ---- 6.N2
| .
| .
|
--- README.md
Dataset Ground truth Mask Label
We have annotated 8 different semantic labels on the UMAD dataset. You can refer to the UMAD-IROS-2024 paper for more details.
Label description:
| Label_index | Label description | Label_binary_intensity_value |
|---|---|---|
| 1 | Anomalous Objects (Only in Query Image) | 128 |
| 2 | Horizon | 64 |
| 3 | People in Query Image | 32 |
| 4 | Dynamic Vehicles in Query Image | 16 |
| 5 | Moved Objects in Query Image | 8 |
| 6 | People in Query Image | 4 |
| 7 | Dynamic Vehicles in Reference Image | 2 |
| 8 | Moved Objects in Reference Image | 1 |
Since reference .jpg and query .jpg may each contain objects, and these objects might overlap if placed in a single mask image, we use a binary ground truth mask .png file to describe them. (Note: A set of data, reference .jpg and query .jpg, share the same size ground truth mask .png file.)
For a set of data, assume reference-00000001.jpg and query-00000001.jpg, as well as their ground truth mask file mask-00000001.png.
For example, when the pixel value at a point in the mask-00000001.png is 196, which is represented in binary as 11000100, so this point belongs to the Anomalous Objects(128, 1000000) in the query data, the Horizon(64, 01000000), and the People in Query Image(4, 00000100).
11000100 =
10000000 + # 128, Anomalous Objects (Only in Query Image)
01000000 + # 64, Horizon
00000100 # 4, People in Query Image
Citation
This project is part of UMAD. If you find this work useful, please consider citing the paper:
@article{li2024umad
author = {Li, Dong and Chen, Lineng and Xu, Cheng-Zhong and Kong, Hui},
title = {UMAD: University of Macau Anomaly Detection Benchmark Dataset},
journal = {arXiv preprint arXiv:2408.12527},
year = {2024},
}
Note
You can contact Dong Li via email(lidong8421bcd@gmail.com) or open an issue on UMAD repo directly If you have any questions.