自动数据增强 对抗学习

自动数据增强，对抗学习（adversarial）

• 现象：在自动驾驶场景中，经常遇到一些corner cases,如树影投到路面上，导致模型误检测，造成错误刹车。在阴天时，对黑色车辆的漏检。
• 原因：模型对某些数据泛化性差，或者对训练数据过拟合。
• 方法：1.减少模型对数据过拟合，如何充分利用已有的数据，不断产生新数据

2d关键点估计(adversaial algorithm)

• 问题:Human pose estimation is a fundamental yet challeng
  ing task in computer vision, which aims at localizing hu
  man anatomical keypoints.However, unlike human vision
  that is robust to various data corruptions such as blur and
  pixelation, current pose estimators are easily confused by
  these corruptions. 关键点估计器容易受data corruption 影响。
• 贡献： Our work
  has several unique benefits. (1) AdvMix is model-agnostic
  and capable in a wide-spectrum of pose estimation mod-
  els. (2) AdvMix consists of adversarial augmentation and
  knowledge distillation.
• 方法： Adversarial augmentation contains
  two neural network modules that are trained jointly and
  competitively in an adversarial manner, where a generator network mixes different corrupted images to confuse a
  pose estimator, improving the robustness of the pose estimator by learning from harder samples. To compensate for the
  noise patterns by adversarial augmentation, knowledge distillation is applied to transfer clean pose structure knowledge to the target pose estimator

2021_CVPR_When Human Pose Estimation Meets Robustness:
Adversarial Algorithms and Benchmarks