PetFace: A Large-Scale Dataset and Benchmark for Animal Identification (2407.13555v2)

Abstract: Automated animal face identification plays a crucial role in the monitoring of behaviors, conducting of surveys, and finding of lost animals. Despite the advancements in human face identification, the lack of datasets and benchmarks in the animal domain has impeded progress. In this paper, we introduce the PetFace dataset, a comprehensive resource for animal face identification encompassing 257,484 unique individuals across 13 animal families and 319 breed categories, including both experimental and pet animals. This large-scale collection of individuals facilitates the investigation of unseen animal face verification, an area that has not been sufficiently explored in existing datasets due to the limited number of individuals. Moreover, PetFace also has fine-grained annotations such as sex, breed, color, and pattern. We provide multiple benchmarks including re-identification for seen individuals and verification for unseen individuals. The models trained on our dataset outperform those trained on prior datasets, even for detailed breed variations and unseen animal families. Our result also indicates that there is some room to improve the performance of integrated identification on multiple animal families. We hope the PetFace dataset will facilitate animal face identification and encourage the development of non-invasive animal automatic identification methods.

• The paper introduces PetFace, a large-scale dataset with over one million annotated images covering 13 animal families and 319 breeds.
• It details fine-grained annotations such as sex, breed, color, and pattern to enable precise recognition and behavioral analysis.
• Benchmark experiments reveal ArcFace models achieving 51.23% top-1 accuracy and 92.17% AUC, underscoring the dataset's robust evaluation protocols.

PetFace: A Comprehensive Dataset and Benchmark for Animal Face Identification

The paper "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" authored by Risa Shinoda and Kaede Shiohara presents PetFace, a significant contribution to the field of automated animal face identification. This dataset aims to mitigate the limitations imposed by the scarcity of extensive datasets and benchmarks available for the domain. PetFace encompasses 257,484 unique individuals across 13 animal families and 319 breed categories, offering over one million images with fine-grained annotations such as sex, breed, color, and pattern.

Key Contributions

1. Dataset Creation:
   • PetFace far surpasses prior publicly available datasets like DogFaceNet and MacaqueFaces in terms of individual scale and species diversity. For instance, PetFace includes over 110 times more individuals than the previously largest dataset for animal faces.
   • The dataset encompasses a varied range of species, meticulously curated from pet shop and animal adoption websites to ensure a comprehensive collection with unique individual identifiers.
2. Fine-Grained Annotations:
   • Beyond conventional identifiers, PetFace includes sex, breed, and detailed color annotations, which are essential for nuanced research in animal behavior and health diagnostics.
   • This granular labeling facilitates detailed investigations into fine-grained recognition, which is vital for recognizing intricate differences within and across species.
3. Benchmark Establishment:
   • The paper introduces two primary evaluation protocols: re-identification for seen faces and verification for unseen faces, thus covering a broader spectrum of real-world application scenarios.
   • Models trained on PetFace consistently outperform those trained on existing datasets, highlighting the enhancement brought by the comprehensive nature of PetFace. For instance, ArcFace models trained on PetFace achieved 51.23% top-1 accuracy for seen faces and 92.17% AUC for unseen face verification.

Experimental Methodology and Results

The robustness of the dataset is evaluated through diverse experimental setups wherein models were tested on various tasks such as re-identification and verification.

1. Re-Identification:
   • Evaluations reveal that ArcFace-based models exhibit superior performance with an average accuracy of 51.23%. Models trained jointly on multiple families within PetFace show even better performance, achieving an overall accuracy of 53.80%.
   • However, these unified models sometimes underperform on specific families, indicating potential areas for improvement in integrated identification across diverse animal categories.
2. Verification:
   • For unseen individual recognition, ArcFace trained models achieve an average AUC of 91.30%, showcasing robust generalization capabilities.
   • Comparative experiments using models trained on other datasets such as ImageNet and CLIP underline PetFace's advantage, demonstrating a higher AUC score across varied animal families.

Comparative Evaluation

Comparisons with previous datasets, specifically for Chimpanzee and Dog categories, underscore the effectiveness of PetFace. When models trained on PetFace were tested against CTai and CZoo (Chimpanzee datasets) and DogFaceNet and Flickr-dog (Dog datasets), they consistently outperformed models trained solely on these older datasets. This demonstrates the higher generalization and discrimination ability provided by PetFace, attributed to its extensive and detailed data.

Implications and Future Work

The implications of the research are multifaceted:

• Practical Applications: The dataset can greatly improve non-invasive monitoring, behavior analysis, and health diagnostics in both pets and wild animals.
• Theoretical Advancements: It fosters the development of more robust and generalized models in computer vision and AI, particularly in the animal identification domain.

Looking forward, the limitations observed in integrated family identification and the necessity for improved representation learning strategies highlight avenues for future research. For instance, exploring more advanced and tailored deep learning architectures or hybrid models may address the existing performance gaps observed in integrated datasets.

In conclusion, PetFace stands as a significant resource that bridges the gap between human and animal face recognition, enabling researchers to push the boundaries of what is achievable in animal identification. The dataset's availability promotes further advancements and applications, potentially altering current methodologies in both academic and practical realms of animal paper.