AttriBE paper quantifies attribute expressivity in body embeddings for ReID

Researchers have developed a new framework called AttriBE to quantify how much specific attributes are encoded within body embeddings used in person re-identification systems. This method uses a secondary neural network to measure the mutual information between learned features and attributes like gender, pose, and BMI. Their analysis of transformer-based ReID models revealed that BMI is consistently the most expressed attribute in deeper layers, followed by pitch, gender, and yaw, with expressivity changing throughout training and across different network depths. The study also extended to cross-spectral identification, showing increased reliance on structural cues like pitch and BMI when bridging infrared modalities. AI

影响 Introduces a novel method for analyzing attribute encoding in ReID models, potentially improving fairness and generalization by understanding implicit biases.

排序理由 This is a research paper detailing a new framework and analysis of attribute expressivity in body embeddings for person re-identification.

在 arXiv cs.CV 阅读 →

• arXiv
• AttriBE
• transformer

• 论文
• 其他

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →