# What is UAVid?

The UAVid dataset is an UAV video dataset for semantic segmentation task focusing on urban scenes. It has several features:

• Semantic segmentation
• 4K resolution UAV videos
• 8 object categories
• Street scene context

### High resolution quality

The images are captured in very high resolution with detailed scenes.

### What are the categories?

There are 8 categories in total:

• Building
• Static car
• Tree
• Low vegetation
• Human
• Moving car
• Background clutter

# News

UAVid 2020 version has 42 sequences in total (20 train, 7 valid and 15 test). Besides the original 30 sequences (UAVid10 version), another 12 sequences have been collected to further strenghthern the dataset.
• Evaluation server is online. Both of the UAVid10 and the UAVid2020 can be evaluated on the Codalab. Experiments on UAVid2020 are recommended. Go to benchmark page for more details.

# Citation

Please cite our paper if you find our UAVid dataset useful.
Bibtex references are as follows,

@article{LYU2020108,
author = "Ye Lyu and George Vosselman and Gui-Song Xia and Alper Yilmaz and Michael Ying Yang",
title = "UAVid: A semantic segmentation dataset for UAV imagery",
journal = "ISPRS Journal of Photogrammetry and Remote Sensing",
volume = "165",
pages = "108 - 119",
year = "2020",
issn = "0924-2716",
doi = "https://doi.org/10.1016/j.isprsjprs.2020.05.009",
url = "http://www.sciencedirect.com/science/article/pii/S0924271620301295",
}

@misc{1810.10438,
Author = {Ye Lyu and
George Vosselman and
Guisong Xia and
Alper Yilmaz and
Michael Ying Yang},
Title  = {The UAVid Dataset for Video Semantic Segmentation},
Year   = {2018},
Eprint = {arXiv:1810.10438},
}


# Organization

## Contact

y.lyu@utwente.nl

michael.yang@utwente.nl

## Semantic Labelling with Video Support

The UAVid dataset provides images and labels for the training and validation set, and images only for the testing set.

All sequences are provided with the corresponding videos. Image and label files are named according to the 0-based index in the video sequence.

## Semantic Labelling with Images Only

If you only need images and labels from the UAVid dataset without video support, please use the following link.

## UAVid Toolkit (python)

UAVidToolKit provides basic tools for easier usage of the UAVid dataset. Including label conversion, label visualization, performance evaluation

## Semantic Labelling

The task for UAVid dataset is to predict per-pixel semantic labelling for the UAV video sequences. The original video file for each sequence is provided together with the labelled images. Currently, UAVid only supports image level semantic labelling without instance level consideration.

## Evaluation Metric

The semantic labelling performance is assessed based on the standard Jaccard Index, more known as the PASCAL VOC intersection-over-union metric.

$$IoU = {TP \over TP+FP+FN}.$$

TP, FP and FN are the numbers of true positive, false positive and false negative respectively, which can be calculated through the confusion matrix determined over all data from test split.
The goal for this task is to achieve as high IoU score as possible. For UAVid dataset, clutter class has a relatively large pixel number ratio and consists of meaningful objects, which is taken as one class for both training and evaluation rather than being ignored.

## Evaluation

The evaluation server and the benchmark table are held on Codalab platform.

Prediction results should be compressed into a zip file before sending to the evaluation server.

Here is an example zip file result to submit: [Example Submission]

Please click on the following button to the UAVid page.

## Benchmark

#### UAVid 2020

 Method Video Clutter Building Road Tree Vegetation Moving Car Static Car Human meanIoU MSD No 57.0 79.8 56.5 74.6 73.7 31.8 62.4 19.8 57.0

#### Legend

• MSD: UAVid: A semantic segmentation dataset for UAV imagery. Ye Lyu, George Vosselman, Gui-Song Xia, Alper Yilmaz and Michael Ying Yang. ISPRS Journal, 2020. [Paper] [Project Page]

## Annotation Method

All the labels are acquired with our home-made video labeller tool.
Three annotation methods are provided:

• Pixel annotation: label pixels in basic scribbling style.
• Super-pixel annotation: automatically partition the image into super-pixels first, and label super-pixels instead.
• Polygon annotation: draw polygons intead of pixels. Pixels in the polygon are labelled the same class.

## Video Labelling Tool

This tool is used for making video semantic labeling ground truth data. It has been used for UAVid dataset.