Fursee: Hybrid YOLO-DINOv3 Framework for Fursuit Identity Retrieval and Clustering

As the furry fandom grows, the volume of convention photography has created a bottleneck for manual image sorting and identity tagging. Existing general-purpose multimodal models often fail to distinguish between fursuits because they share similar base molds, relying instead on subtle color and texture patterns. This paper addresses the need for an automated, fine-grained identity retrieval and clustering system capable of handling crowded scenes with multiple overlapping subjects.

The author introduces Fursee, a three-stage hybrid pipeline designed to isolate and identify fursuit characters. First, a YOLO detector is trained to locate and crop high-resolution fursuit head patches, effectively removing background noise and isolating individual subjects. Second, these patches are processed by a DINOv3 Vision Transformer backbone, which is fine-tuned using ArcFace loss to maximize the angular separation between different identities in the feature space. Finally, an unsupervised clustering approach using DBSCAN is employed, where the optimal hyperparameters are automatically selected via a silhouette-coefficient-driven search rather than manual tuning.

Experimental results demonstrate that the Fursee pipeline outperforms mainstream multimodal models, including GPT-5.5, Claude Opus 4.8, and Qwen3.7-Plus, across all evaluation metrics. By decoupling localization (YOLO) from feature extraction (DINOv3) and using metric learning (ArcFace), the framework successfully handles the fine-grained visual distinctions required to differentiate between unique fursuit designs. The automatic hyperparameter selection for clustering further ensures that the system can adapt to datasets where the total number of identities is unknown.

This work provides a specialized solution for a niche but high-volume domain, offering a scalable way to organize massive archives of convention photos. By pioneering a detection-guided embedding pipeline for this specific subculture, the research demonstrates how combining off-the-shelf detectors with domain-specific metric learning can solve complex, fine-grained recognition tasks that general-purpose models currently struggle to address.