iPanda-50

The iPanda-50 dataset consists of 6,874 images of 50 giant panda individuals with 49 ~ 292 images per panda. The iPanda-50 dataset is used for fine-grained panda identification and it was proposed in Le Wang, Rizhi Ding, Yuanhao Zhai, Qilin Zhang, Wei Tang, Nanning Zheng, and Gang Hua, “Giant Panda Identification”, IEEE Transactions on Image Processing, 2021. In case you do not have convenient access to IEEE Xplore, the preprint PDF can be accessed here from Github.io or here from xjtu.edu.cn. If you find this dataset helpful in your research, please consider citing this publication.

  @ARTICLE{Wang21Giant,
  author={Le Wang and Rizhi Ding and Yuanhao Zhai and Qilin Zhang and Wei Tang and Nanning Zheng and Gang Hua},
  journal={IEEE Transactions on Image Processing}, 
  title={Giant Panda Identification}, 
  year={2021},
  volume={30},
  pages={2837-2849},
  doi={10.1109/TIP.2021.3055627}}

This dataset can be downloaded as ZIP files from the following URLs. Some filenames of the JPG images contain UTF-8 characters, please enable proper encoding while reading these files.

• (Not recommended: single zip file, requiring client software) Baidu NetDisk/Wangpan 百度网盘 Extraction code 提取码: ohln File Size: 885.8M
  The 'iPanda-50.zip' file contains three folders, i.e., "iPanda50-images", "iPanda50-split", "iPanda-50-eyes-labels".
  Note: Baidu NetDisk/Wangpan service may require a client software (Baidu Netdisk) installed on your machine before allowing you to download. If you prefer not to install any software, please try the alternative download links as follows.

• (Recommended: 3 Zip files, direct download links) Please download all three ZIP files for the images, train/test split settings and the additional eye annotations, respectively.

Files	File Size	MD5 checksum	Download Servers
iPanda50-images.zip	883M	46727b709139ad3306491f9c11eaefba	[Google Drive], [Dropbox], [MS OneDrive]
iPanda50-split.zip	137K	b133b5ef63de4b02a7193c377efc4635	[Google Drive], [Dropbox], [MS OneDrive]
iPanda50-eyes-labels.zip	2.5M	5bf41bf5cf2f401664ced7ccf55df870	[Google Drive], [Dropbox], [MS OneDrive]

Figure: Sample images and statistics of the iPanda-50 dataset. Statistics are formatted as identity, #images ( #images-in-training/#images-in-testing )

iPanda50-images

All images are in the "iPanda50-images" folder, with the directory structure as follows.

.
├── 00_aibang
├── 01_aoliao
├── 02_baolan
├── 03_bingdian
├── 04_chengdui
├── 05_chengjiu
├── 06_chengshi
├── 07_chengshuang
├── 08_fulai
├── 09_fushun
├── 10_hexing
├── 11_jiaoao
├── 12_jingjing
├── 13_jingliang
├── 14_maodou
├── 15_maosun
├── 16_maotao
├── 17_meibang
├── 18_miaomiao
├── 19_nannan
├── 20_nike
├── 21_nina
├── 22_nini
├── 23_qiubang
├── 24_qixi
├── 25_qiyi
├── 26_qiyuan
├── 27_rourou
├── 28_sa
├── 29_shuangxiong
├── 30_shuqing
├── 31_shurong
├── 32_susu
├── 33_wuyi
├── 34_xiaoqiao
├── 35_xilan
├── 36_xingda
├── 37_xinger
├── 38_xingfan
├── 39_xinghui
├── 40_xingrong
├── 41_xingxiao
├── 42_xingyi
├── 43_yaxing
├── 44_yayi
├── 45_yayun
├── 46_yazhu
├── 47_yingying
├── 48_yongbang
└── 49_yuanrun 

These 50 subfolders (00_aibang until 49_yuanrun) represent the numeric labels and names of these 50 giant panda individuals. For example, Panda #0 has the name "aibang" and Panda #49 has the name "yuanrun".

iPanda50-split

We randomly divide the iPanda-50 dataset into the training split and the testing split at approximately 3:2 ratio, leading to a training set of 4,106 images and a test set of 2,768 images. This division operation was repeated for a total of 5 times (i.e., 5 Monte Carlo trials), and the results are saved in this iPanda50-split folder.

The directory structure of iPanda50-split is as follows.

.
├── split0_test.txt
├── split0_train.txt
├── split1_test.txt
├── split1_train.txt
├── split2_test.txt
├── split2_train.txt
├── split3_test.txt
├── split3_train.txt
├── split4_test.txt
└── split4_train.txt

iPanda50-eyes-labels

The area around eyes may contribute significantly to the identification of giant pandas, therefore we have provided additional eye annotations (i.e., location) in JSON format in the iPanda50-eyes-labels folder. Only images with at least one eye visible are annotated. A total of 5,696 images are annotated in this folder. These JSON files are organized similar to the iPanda50-images folder, with each JSON file corresponding to an image with the same name.