AINeutralarXiv – CS AI · 7h ago6/10
🧠
Towards Resolving Optimization Conflicts Between Image- and Text-Based Person Re-Identification
Researchers propose a decoupled two-stage training pipeline to resolve optimization conflicts when jointly training image-based and text-based person re-identification systems. The approach uses a single vision encoder with separate training stages to prevent cross-task interference, improving performance in both retrieval modalities.