[Paper Review] Amur Tiger Re-identification in the Wild.

Shuyuan Li, Jianguo Li+2 more|arXiv (Cornell University)|Jun 13, 2019

Wildlife Ecology and Conservation25 citations

TL;DR

This paper introduces the Amur Tiger Re-identification in the Wild (ATRW) dataset, a large-scale collection of over 8,000 video clips from 92 wild Amur tigers with bounding boxes, pose keypoints, and identity annotations. It proposes a novel deep learning method that models precise pose parts to improve re-identification under extreme pose and lighting variations, achieving significant performance gains over existing re-ID methods on this challenging dataset.

ABSTRACT

Monitoring the population and movements of endangered species is an important task to wildlife conversation. Traditional tagging methods do not scale to large populations, while applying computer vision methods to camera sensor data requires re-identification (re-ID) algorithms to obtain accurate counts and moving trajectory of wildlife. However, existing re-ID methods are largely targeted at persons and cars, which have limited pose variations and constrained capture environments. This paper tries to fill the gap by introducing a novel large-scale dataset, the Amur Tiger Re-identification in the Wild (ATRW) dataset. ATRW contains over 8,000 video clips from 92 Amur tigers, with bounding box, pose keypoint, and tiger identity annotations. In contrast to typical re-ID datasets, the tigers are captured in a diverse set of unconstrained poses and lighting conditions. We demonstrate with a set of baseline algorithms that ATRW is a challenging dataset for re-ID. Lastly, we propose a novel method for tiger re-identification, which introduces precise pose parts modeling in deep neural networks to handle large pose variation of tigers, and reaches notable performance improvement over existing re-ID methods. The dataset will be public available at this https URL .

Motivation & Objective

• To address the lack of large-scale, unconstrained wildlife re-identification datasets for endangered species like Amur tigers.
• To develop a re-identification method robust to extreme pose variations and diverse lighting conditions in wild settings.
• To create a benchmark dataset that enables training and evaluation of deep learning models for wildlife monitoring.
• To improve the accuracy of individual animal tracking and population estimation in conservation efforts.

Proposed method

• The authors construct the ATRW dataset, comprising over 8,000 video clips from 92 Amur tigers captured in natural, unconstrained environments.
• Each clip is annotated with bounding boxes, 2D pose keypoints, and individual tiger identities to support detailed analysis.
• A novel deep neural network architecture is proposed that explicitly models pose parts to enhance feature representation under large pose variations.
• The model integrates spatial and structural priors from keypoint annotations to improve feature discrimination across diverse poses.
• Baseline re-ID models are evaluated on ATRW to establish performance benchmarks and demonstrate dataset difficulty.
• The method is trained and tested on the ATRW dataset, with performance measured using standard re-ID metrics such as mAP and rank-1 accuracy.

Experimental results

Research questions

• RQ1 How does the performance of existing re-identification models generalize to wild Amur tigers under unconstrained conditions?
• RQ2 To what extent do pose variations and lighting conditions in natural environments degrade standard re-ID model performance?
• RQ3 Can explicit modeling of pose parts in deep neural networks improve re-identification accuracy for tigers with large pose variations?
• RQ4 How does the proposed method compare to baseline re-ID models on the ATRW dataset in terms of mAP and rank-1 accuracy?

Key findings

• The ATRW dataset contains over 8,000 video clips from 92 individual Amur tigers, annotated with bounding boxes, pose keypoints, and identities.
• Baseline re-ID models show limited performance on ATRW, confirming the dataset's challenge due to extreme pose and lighting variations.
• The proposed pose-aware deep learning method achieves notable performance improvements over existing re-ID methods on the ATRW benchmark.
• The integration of precise pose part modeling significantly enhances feature representation, particularly in cases of large pose variation.
• The dataset is publicly available at the provided URL, enabling broader research in wildlife re-identification and conservation technology.