论文解读:Exploring Object Relation in Mean Teacher for Cross-Domain Detection

论文题目:Exploring Object Relation in Mean Teacher for Cross-Domain Detection(CVPR 2019)

论文主要贡献:该论文主要也是为了解决跨域的目标检测问题,但是与之前的基于判别的特征方法有很大的不同,主要是利用了半监督学习中SOTA方法Mean Teacher(NIPS 2017:Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results),同样的有一篇论文也是利用Mean Teacher解决跨域的图像分类问题(ICLR 2018:Self-ensembling for domain adaptation),而本文是解决检测问题;

论文解读:Exploring Object Relation in Mean Teacher for Cross-Domain Detection_第1张图片

1、将跨域分类的论文(ICLR 2018:Self-ensembling for domain adaptation)中的全局的consistency转成检测中的region的consistency;

2、引入inter-graph和intra-graph consistency,变相将object relation引入Mean Teacher结构;

论文解读:Exploring Object Relation in Mean Teacher for Cross-Domain Detection_第2张图片

网络结构:

论文解读:Exploring Object Relation in Mean Teacher for Cross-Domain Detection_第3张图片

1、

你可能感兴趣的:(域自适应,目标检测,计算机视觉,domain,adaption,目标检测)