1. Motivation
数据增强在分类网络中应用广泛,但是在目标检测中under-explored.
- Data augmentation has become a de facto component for training high-performance deep image classifiers, but its potential is under-explored for object detection.
- In this paper, we aim to enhance this learning paradigm for training not only accurate but also robust object detectors.
一般来说,目标检测中有2种方法可以增强detector,第一种是直接获取上游分类任务的模型,然后fine-tune;另一种就是直接在训练过程中增强的detector。
- there are two natural choices to augment a detector: (i) Borrow an already augmented model from the upstream classification task and then fine-tune it for object detection. (ii) Directly augment the detector during training
- we shift to the second option, augmenting during fine- tuning for harvesting data augmentation to improve a detec- tor’s accuracy and robustness.
作者研究了分类网络中(resnet)应用的多种数据增强如何迁移到目标检测中,但是实验效果表明,在fine-tune过后,accuracy以及robustness中的增益都会diminish。
- we first study how the classifiers’ gains from various data augmentations transfer to object detection.
- The results are discouraging; the gains diminish after fine-tuning in terms ofeither accuracy or robustness.
这篇文章通过对抗样本adversarial examples增强目标检测的fine-tuning stage。也就是把对抗样本看做是一个基于模型model-independent 数据增强。
- This work instead augments the fine-tuning stage for object detectors by exploring adversarial examples.
这篇文章是建立在AdvProp的基础之上,提出了Det-AdvProp,适用于目标检测网络。
- Inspired by their success in image classification, we pro- pose Det-AdvProp to build accurate and robust detectors as object detection plays a crucial role in many real-world ap- plications such as autonomous drivin