This website requires JavaScript.

Rethinking Person Re-Identification via Semantic-Based Pretraining

Suncheng XiangJingsheng GaoZirui Zhang ...+4 Yuzhuo Fu
Oct 2021
摘要
Pretraining is a dominant paradigm in computer vision. Generally, supervisedImageNet pretraining is commonly used to initialize the backbones of personre-identification (Re-ID) models. However, recent works show a surprisingresult that CNN-based pretraining on ImageNet has limited impacts on Re-IDsystem due to the large domain gap between ImageNet and person Re-ID data. Toseek an alternative to traditional pretraining, here we investigatesemantic-based pretraining as another method to utilize additional textual dataagainst ImageNet pretraining. Specifically, we manually construct a diversifiedFineGPR-C caption dataset for the first time on person Re-ID events. Based onit, a pure semantic-based pretraining approach named VTBR is proposed to adoptdense captions to learn visual representations with fewer images. We trainconvolutional neural networks from scratch on the captions of FineGPR-Cdataset, and then transfer them to downstream Re-ID tasks. Comprehensiveexperiments conducted on benchmark datasets show that our VTBR can achievecompetitive performance compared with ImageNet pretraining - despite using upto 1.4x fewer images, revealing its potential in Re-ID pretraining.
展开全部
图表提取

暂无人提供速读十问回答

论文十问由沈向洋博士提出,鼓励大家带着这十个问题去阅读论文,用有用的信息构建认知模型。写出自己的十问回答,还有机会在当前页面展示哦。

Q1论文试图解决什么问题?
Q2这是否是一个新的问题?
Q3这篇文章要验证一个什么科学假设?
0
被引用
笔记
问答