Data Augmentation

Data Augmentation( From deeplearning AI)

- Mirror

- Crop

注：左上图中的红框处的裁剪应该不利于训练（推测，不确定）

- Rotation,Shearing,Local warping

- Color shifting

注： PCA色彩增扩 （AlexNet paper): 让全部的色彩，色调一致）

- 多线程：1个线程做数据增强，其它线程做训练

Yolo data augmentation

saturation = 1.5
exposure = 1.5

hue=.1

jitter=.3

random=1

From AlexeyAB

Data augmentation is built-in in both Original repo https://github.com/pjreddie/darknet and my repo https://github.com/AlexeyAB/darknet and enbabled by default.

Automatically will be generated infinity number of images with changed colors (hue, saturation, exposure) randomly croped and resized (jitter of edges (top,bottom,left,right) of images), and neural network will be randomly resized.

Does it preformed automatically during training? - Yes.

How many new images are created? - Infinity - you can only change the amplitude of the changes.

Does it transform the bounding boxes as well? - Yes:

- it transforms image:  https://github.com/AlexeyAB/darknet/blob/65e430588d76adbcf435db6b2e3aec791de651d0/src/data.c#L776-L784

- it transforms bounding boxes:  https://github.com/AlexeyAB/darknet/blob/65e430588d76adbcf435db6b2e3aec791de651d0/src/data.c#L786

But the data augmentation can not completely replace a large set of data with a diverse context