Animal Kingdom Dataset

This is the official repository for
[CVPR2022] Animal Kingdom: A Large and Diverse Dataset for Animal Behavior Understanding
Xun Long NG, Kian Eng ONG, Qichen ZHENG, Yun NI, Si Yong YEO, Jun LIU
Information Systems Technology and Design, Singapore University of Technology and Design, Singapore

Dataset and Codes

Download dataset and codes here

NOTE: The codes of the models for all tasks have been released. Codes are included in the folder of the dataset. After you download our dataset, you can find the corresponding codes for each task. Helper scripts are provided to automatically set up the environment to directly run our dataset. The Animal_Kingdom GitHub codes are the same as the codes in the download version, hence there is no need to download the GitHub codes.

README

Please read the respective README files in Animal_Kingdom for more information about preparing the dataset for the respective tasks.

• Overview of dataset
• Meta-data: 140 Action Description, List of >850 animals and classification
• Action Recognition
• Pose Estimation
• Video Grounding

Paper

• CVPR2022
• ResearchGate
• arXiv

Citation

@InProceedings{
    Ng_2022_CVPR,
    author    = {Ng, Xun Long and Ong, Kian Eng and Zheng, Qichen and Ni, Yun and Yeo, Si Yong and Liu, Jun},
    title     = {Animal Kingdom: A Large and Diverse Dataset for Animal Behavior Understanding},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {19023-19034}
 }

Abstract

Understanding animals' behaviors is significant for a wide range of applications. However, existing animal behavior datasets have limitations in multiple aspects, including limited numbers of animal classes, data samples and provided tasks, and also limited variations in environmental conditions and viewpoints. To address these limitations, we create a large and diverse dataset, Animal Kingdom, that provides multiple annotated tasks to enable a more thorough understanding of natural animal behaviors. The wild animal footages used in our dataset record different times of the day in extensive range of environments containing variations in backgrounds, viewpoints, illumination and weather conditions. More specifically, our dataset contains 50 hours of annotated videos to localize relevant animal behavior segments in long videos for the video grounding task, 30K video sequences for the fine-grained multi-label action recognition task, and 33K frames for the pose estimation task, which correspond to a diverse range of animals with 850 species across 6 major animal classes. Such a challenging and comprehensive dataset shall be able to facilitate the community to develop, adapt, and evaluate various types of advanced methods for animal behavior analysis. Moreover, we propose a Collaborative Action Recognition (CARe) model that learns general and specific features for action recognition with unseen new animals. This method achieves promising performance in our experiments.

Action Recognition

eg_ar.mp4

Results of action recognition

	mAP
Method	overall	head	middle	tail
Baseline (Cross Entropy Loss)
I3D	16.48	46.39	20.68	12.28
SlowFast	20.46	54.52	27.68	15.07
X3D	25.25	60.33	36.19	18.83
Focal Loss
I3D	26.49	64.72	40.18	19.07
SlowFast	24.74	60.72	34.59	18.51
X3D	28.85	64.44	39.72	22.41
LDAM-DRW
I3D	22.40	53.26	27.73	17.82
SlowFast	22.65	50.02	29.23	17.61
X3D	30.54	62.46	39.48	24.96
EQL
I3D	24.85	60.63	35.36	18.47
SlowFast	24.41	59.70	34.99	18.07
X3D	30.55	63.33	38.62	25.09

Collaborative Action Recognition (CARe) Model

Results of action recognition of unseen animals

Method	Accuracy (%)
Episodic-DG	34.0
Mixup	36.2
CARe without specific feature	27.3
CARe without general feature	38.2
CARe without spatially-aware weighting	37.1
CARe (Our full model)	39.7

Pose Estimation

Results of pose estimation

		PCK@0.05
Protocol	Description	HRNet	HRNet-DARK
Protocol 1	All	66.06	66.57
Protocol 2	Leave-k-out	39.30	40.28
Protocol 3	Mammals	61.59	62.50
	Amphibians	56.74	57.85
	Reptiles	56.06	57.06
	Birds	77.35	77.41
	Fishes	68.25	69.96

Video Grounding

Results of video grounding

	Recall@1				mean IoU
Method	IoU=0.1	IoU=0.3	IoU=0.5	IoU=0.7
LGI	50.84	33.51	19.74	8.94	22.90
VSLNet	53.59	33.74	20.83	12.22	25.02

Acknowledgement and Contributors

This project is supported by AI Singapore (AISG-100E-2020-065), National Research Foundation Singapore, and SUTD Startup Research Grant.

We would like to thank the following contributors for working on the annotations and conducting the quality checks for video grounding, action recognition and pose estimation.

• Ang Yu Jie
• Ann Mary Alen
• Cheong Kah Yen Kelly
• Foo Lin Geng
• Gong Jia
• Heng Jia Ming
• Javier Heng Tze Jian
• Javin Eng Hee Pin
• Jignesh Sanjay Motwani
• Li Meixuan
• Li Tianjiao
• Liang Junyi
• Loy Xing Jun
• Nicholas Gandhi Peradidjaya
• Song Xulin
• Tian Shengjing
• Wang Yanbao
• Xiang Siqi
• Xu Li