从多模态信息指导中学习可转移的行人表示（英文版）

英文标题：Learning Transferable Pedestrian Representation from Multimodal Information Supervision中文摘要：本文提出了一种学习可迁移表示以增强各种行人分析任务的新框架 VAL-PAT，并介绍了三个学习目标：自监督对比学习、图像文本对比学习和多属性分类。该框架在无标签人员图像的预训练下实现了良好的性能，并扩展到各种行人分析任务。英文摘要：Recent researches on unsupervised person re-identification~(reID) havedemonstrated that pre-training on unlabeled person images achieves superiorperformance on downstream reID tasks than pre-training on ImageNet. However,those pre-trained methods are specifically designed for reID and s