ATRW: A Benchmark for Amur Tiger Re-identification in the Wild

Shuyuan Li, Jianguo Li, Hanlin Tang, Rui Qian, Weiyao LinACM Multimedia 2020 · DOI

Abstract

Wildlife monitoring is transitioning from labor-intensive manual tagging to automated computer vision systems. This paper introduces ATRW (Amur Tiger Re-identification in the Wild), a large-scale benchmark designed to facilitate this shift. The dataset contains over 8,000 video clips of Amur tigers in unconstrained environments with 15-point pose annotations. We propose a novel Pose Part-based Model (PPbM) that leverages skeletal keypoints to address the challenges of non-rigid animal movement and stripe pattern deformation, significantly outperforming standard re-identification (re-ID) algorithms.

Introduction

Accurate population tracking is critical for wildlife conservation. While computer vision has revolutionized re-identification for humans and vehicles, animals present unique challenges. Unlike rigid vehicles or upright pedestrians, tigers are non-rigid, four-limbed objects that move through complex environments with varying lighting. Their identity is primarily determined by unique stripe patterns, which deform significantly based on posture. Existing datasets often lack the scale, pose annotations, or environmental realism necessary for real-world deployment. ATRW addresses these gaps by providing a comprehensive "in the wild" benchmark.

Methodology

The proposed end-to-end pipeline for Amur tiger monitoring, from camera trap detection to web-based trajectory visualization.

The authors propose a multi-stage pipeline consisting of detection, pose estimation, and re-identification.

Detection and Pose Estimation

To ensure viability for battery-powered camera traps, lightweight detectors such as TinyDSOD and SSD-MobileNet were benchmarked. For pose estimation, HRNet and AlphaPose were adapted to a specialized 15-point tiger skeleton. Accurate pose estimation is foundational for the subsequent re-ID task, as it allows the model to normalize for the tiger's non-rigid movement.

Pose Part-based Model (PPbM)

The core innovation is the PPbM, which moves beyond global image analysis. By using estimated pose keypoints, the model defines a "7-part star model" encompassing the trunk, legs, thighs, and shanks.

  1. Regional Average Pooling (RAP): Instead of pooling features at the final low-resolution layer, PPbM extracts local features from an intermediate layer (res3d). This preserves spatial resolution (32x16 vs. 8x4), capturing fine-grained stripe patterns essential for identification.
  2. Aggregation Strategies: Two variants were tested: PPbM-a, which uses simple concatenation of part features, and PPbM-b, which employs a soft-attention mechanism to weigh the most informative body parts for a given frame.

The ATRW Dataset

A sample from the ATRW dataset showing the 15-point skeletal annotation used for pose estimation.

The ATRW dataset serves as the primary contribution of this work, featuring:

  • Scale: 8,076 high-resolution video clips covering 92 individual Amur tigers.
  • Annotations: Comprehensive labels including bounding boxes, 15 skeletal keypoints, and identity labels.
  • Realism: Data captured in unconstrained settings with diverse poses and lighting conditions.
  • Entity Definition: Due to the asymmetry of stripe patterns, the left and right sides of the same tiger are treated as distinct entities for the re-ID task.

Experimental Results

Experimental evaluations demonstrate the effectiveness of the proposed framework:

  • Re-ID Performance: PPbM-a achieved a Mean Average Precision (mAP) of 74.1% in the plain setting, a significant improvement over the standard classification baseline of 59.1%.
  • Pose Estimation: HRNet proved highly effective, achieving 86.9% AP at an Object Keypoint Similarity (OKS) of 0.5.
  • Edge Deployment: TinyDSOD emerged as the optimal balance for edge devices, achieving 0.511 mAP with only 1.1B FLOPs.

Conclusion and Future Work

While ATRW represents a significant step forward, challenges remain. A measurable performance gap exists between "Plain re-ID" (manual crops) and "Wild re-ID" (automated detection), indicating that detection and pose accuracy are current bottlenecks. Furthermore, performance drops in cross-camera scenarios where lighting and angles vary significantly. Future research should focus on integrating "unknown" identity samples into training and improving robustness across diverse camera viewpoints.

Summary updated

BibTeX
@inproceedings{li2020atrw,
  author = {Shuyuan Li and Jianguo Li and Hanlin Tang and Rui Qian and Weiyao Lin},
  title = {{ATRW: A Benchmark for Amur Tiger Re-identification in the Wild}},
  booktitle = {ACM Multimedia},
  year = {2020},
  doi = {10.1145/3394171.3413569},
}
Figure