VALSE 论文速览 第94期：基于Probabilistic teacher的跨域目标检测算法 ...

论文题目：Learning Domain Adaptive Object Detection with Probabilistic Teacher

作者列表：Meilin Chen (Zhejiang University), Weijie Chen (Zhejiang University, Hikvision Research Institute), Shicai Yang (Hikvision Research Institute), Jie Song (Zhejiang University), Xinchao Wang (National University of Singapore), Lei Zhang (Chongqing University), Yunfeng Yan (Zhejiang University), Donglian Qi (Zhejiang University), Yueting Zhuang (Zhejiang University), Di Xie (Hikvision Research Institute), Shiliang Pu (Hikvision Research Institute)

B站观看网址：

https://www.bilibili.com/video/BV1oe4y1d7Xg/

论文摘要：

Self-training for unsupervised domain adaptive object detection is a challenging task, of which the performance depends heavily on the quality of pseudo boxes. Despite the promising results, prior works have largely overlooked the uncertainty of pseudo boxes during self-training. In this paper, we present a simple yet effective framework, termed as Probabilistic Teacher (PT), which aims to capture the uncertainty of unlabeled target data from a gradually evolving teacher and guides the learning of a student in a mutually beneficial manner. Specifically, we propose to leverage the uncertainty-guided consistency training to promote classification adaptation and localization adaptation, rather than filtering pseudo boxes via an elaborate confidence threshold. In addition, we conduct anchor adaptation in parallel with localization adaptation, since anchor can be regarded as a learnable parameter. Together with this framework, we also present a novel Entropy Focal Loss (EFL)to further facilitate the uncertainty-guided self-training. Equipped with EFL, PT outperforms all previous baselines by a large margin and achieve new state-of-the-arts.

To be specific, the contributions are summarized as:

(1) We propose a threshold-free framework to explore cross-domain object detection via an uncertainty-driven self-training paradigm. It firstly unifies classification, localization as well as anchor adaptations into one framework.

(2) We design an EFL loss for PT framework to further facilitate uncertainty-guided cross-domain self-training.

(3) We draw several interesting yet novel experimental findings, which can inspire the future works in UDA-OD.

(4) Our framework achieves the new state-of-the-art results on multiple source-based / free UDA-OD benchmarks, and surpasses previous approaches by a large margin.

论文信息：

[1] Chen, Meilin and Chen, Weijie and Yang, Shicai and Song, Jie and Wang, Xinchao and Zhang, Lei and Yan, Yunfeng and Qi, Donglian and Zhuang, Yueting and Xie, Di and others. Learning Domain Adaptive Object Detection with Probabilistic Teacher. ICML 2022.

论文链接：

[https://arxiv.org/pdf/2206.06293.pdf]

代码链接：

[https://github.com/hikvision-research/ProbabilisticTeacher]

视频讲者简介：

陈梅林，浙江大学在读硕士研究生，曾获浙江省政府奖学金、浙江大学一等奖学金、优秀研究生、三好研究生等荣誉。在机器学习、计算机视觉等领域于国际顶级学术会议发表多篇论文，曾担任ICML, ECCV等国际学术会议审稿人。

特别鸣谢本次论文速览主要组织者：

月度轮值AC：冯尊磊 (浙江大学)，徐易 (大连理工大学)

季度责任AC：张姗姗 (南京理工大学)