📄 中文摘要
自动化动物识别是帮助失宠物与主人团聚的实用任务，但现有系统常因数据集规模有限和依赖单一视觉线索而面临挑战。研究提出了一种多模态验证框架，通过合成文本描述衍生的语义身份先验来增强视觉特征。为支持这一研究，构建了一个包含1.9百万张照片和695,091种独特动物的大型训练语料库。通过系统消融研究，识别出SigLIP2-Giant和E5-Small-v2作为最佳视觉和文本骨干网络。此外，评估了从简单连接到自适应门控的多种融合策略，以确定最佳的特征集成方法。
从视觉到多模态：动物识别中的编码器和融合策略的系统消融研究
📄 English Summary
From Visual to Multimodal: Systematic Ablation of Encoders and Fusion Strategies in Animal Identification
Automated animal identification serves as a practical solution for reuniting lost pets with their owners, yet existing systems often face challenges due to limited dataset sizes and reliance on unimodal visual cues. This study proposes a multimodal verification framework that enhances visual features with semantic identity priors derived from synthetic textual descriptions. A massive training corpus comprising 1.9 million photographs of 695,091 unique animals was constructed to support this investigation. Systematic ablation studies identified SigLIP2-Giant and E5-Small-v2 as the optimal vision and text backbones. Additionally, various fusion strategies, ranging from simple concatenation to adaptive gating, were evaluated to determine the most effective method for feature integration.