auniquesun/AirBirds

Introduction

1. The resolution of each image is 1920x1080.

2. The zipped label files contain the bounding box annotations of birds captured in the images. Each image corresponds to a single label file, their filenames are same except for the file extension.

3. For the label format in each label file, we follow the YOLO format, that means for each row in a label file, the format is

   <class_id> <normalized_object_x_center> <normalized_object_y_center> <normalized_object_width> <normalized_object_height>
   

   • Since we only have one class bird, the class_id is always 0. If there is no bird in an image, its label file is empty.

   • If needed, you can convert YOLO format to COCO format using the GitHub project Yolo-to-COCO-format-converter.

Train and Test Set Split

1. images0.zip, images1.zip, ..., images9.zip and their corresponding labels are in the train set.

2. The test set includes images_test_D02_20210628090856.zip and images_test_D02_20210721142744.zip. And the corresponding labels_test_D02_20210628090856.zip and label_test_D02_20210721142744.zip are label files.

3. Users can split validation set individually for evaluating their models. For example, treat images9.zip and labels9.zip as the validation set and the remaining images and labels as the train set. Also, users can split multiple validation sets for cross validation for their model.

License and Citation

1. The dataset is for non-commercial research and educational use only and the license is CC BY-NC-SA 4.0.

2. If you use our dataset or find our dataset is helpful for your project/research, please cite our paper as follows.

   @InProceedings{
       author = {Sun, Hongyu and Wang, Yongcai and Cai, Xudong and Wang, Peng and Huang, Zhe and Li, Deying and Shao, Yu and Wang, Shuo},
       title = {AirBirds: A Large-scale Challenging Dataset for Bird Strike Prevention in Real-world Airports},
       booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)},
       month = {December},
       year = {2022},
       pages = {2440-2456}
   }