重点提炼：
- 多目标检测需要保持空间变换信息（相反，分类任务需要对空间变化不敏感），因此不能直接用原始的（用于分类任务）mixup。本文提出了 visual coherent mixup。
We first explore the mixup technique on object detection. Unlikewe recognize the special property of multiple object detection task which favors spatial preserving transforms,and thus proposed a visually coherent image mixup methods for object detection tasks
- one-stage object detection 框架缺乏图像的空间变化，因此空域数据扩增对于one-stage的算法十分重要。
For example, due to the lack of spatial variation in single stage pipelines, spatialdata augmentation is crucial to the performance as proven in Single-Shot MultiBox Object Detector (SSD)
- 基于采样的two-stage算法在feature map上生成了众多的proposal，这替代了图中的随机裁剪。因此two-stage算法不需要在training阶段进行大量的几何形状的数据扩增。
Since sampling-based approaches repeat enormous crop like operations on feature maps, it substitutes the operation
of randomly cropping the input image, therefore these networks do not require extensive geometric augmentations applied during the training stage.
-
one-stage和two-stage对于data preprocessing的需求差异较大，作者比较了YOLO v3以及faster RCNN的各种效果，见下图
image.png
- Warm-up对于一些目标检测算法是非常重要的，比如YOLO v3。这类算法在迭代开始的时候，负样本会产生较大的梯度。
Warm up learning rate is another common strategy to avoid gradient explosion during the initial training iterations. Warm-up learning rate schedule is critical to several object detection algorithms, e.g., YOLO v3, which has a dominant gradient from negative examples in the very beginning iterations where sigmoid classification score is initialized around 0.5 and biased towards 0 for the majority predictions
- 证明了在YOLO v3上，配上合适的warm-up后，cosine schedule好于 step schedule
image.png
- mixup 目标检测方面可以从两个角度使用: 1. 用于backbone的预训练，2. 用于目标检测器的训练。作者做了实验，无论用于哪个阶段，都对最终的结果有所帮助。而且，如果两个阶段都是用mixup，则可以取得1+1>2的效果。
While the results proved the consistent improvements by adopting mixup to either training phases, it is also notable that applying mixup in both phases can produce more significant gains as 1 + 1 > 2.
9.本文的bag of freebies对于one-stage框架具有更大的提升，通过在COCO2017上面的实验可以看出：
