“CVPR 2017 best paper, this is the other exception, it's unique because it doesn't require _any_ labeled data”
Learning from Simulated and Unsupervised Images through Adversarial Training
用 GAN 式对抗训练把模拟图像变真实,无需真实标注
论文提出 Simulated+Unsupervised(S+U)学习:用无标注的真实图像,通过类似 GAN 的对抗网络提升模拟器输出的真实感,同时保留模拟器自带的标注信息。方法在标准 GAN 上加入自正则项、局部对抗损失和用历史精修图像更新判别器三处改动,用精修图像训练视线估计和手部姿态估计模型,在 MPIIGaze 上无需任何真实标注即达到当时最优结果。
它是 Apple 首篇公开的机器学习论文,也被评为 CVPR 2017 最佳论文,HN 上多在讨论 Apple 博客文章和这篇论文的关系,以及它“无需任何标注数据”的特点。
学习路径
读这篇之前,按顺序读书架上的 3 篇 · 参考文献来自 OpenAlex,顺序和要点由 Claude Code 排出这条路径先建立对抗生成网络的架构经验,再看残差网络,以及基于局部块的对抗损失,最后读目标论文如何用它们把模拟图像变真实。
- 1
目标论文的对抗训练沿用卷积式 GAN 的架构经验和稳定训练技巧。
读给生成器和判别器加卷积结构约束的原则;其余的表示分析可略读。
- 2
- 3
目标论文的局部对抗损失(逐图块判别)与这篇的马尔可夫式判别思路相近。
读判别器作用在局部图块上的思路,以及它为什么适合纹理和局部细节。
- 4Learning from Simulated and Unsupervised Images through Adversarial Training
目标论文:用对抗训练让模拟器输出更真实,同时保留标注。
读S+U 学习的设定,自正则项(保留标注)、局部对抗损失,以及用历史精炼图像稳定判别器的做法。
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“The only "paper" on the "journal" right now looks like it's just a watered-down, html-only version of Learning from Simulated and Unsupervised Images through Adversarial Training!”
“they're clear in their first technical post that it is just an overview of work that the researchers are presenting at CVPR [1].”
“The article is based on a paper submitted to the prestigious CVPR, which is obviously not anonymous.”
“* Learning from Simulated and Unsupervised Images through Adversarial Training [1]”
“Apple's first published ML paper[1] described how they trained a neural network to estimate eye gaze.”
看全部 7 条 →
“It's a similar premise and method than Apple's SUGAN paper [3], that came out around the same time, to much more fanfare (it won CVPR best paper award iirc).”
- Learning from Simulated and Unsupervised Images through Adversarial Training ↗shreyassaxena · 2016-12-26 · 103 分