备用

• Bag of Freebies for Training Object Detection Neural Networks

Learning Data Augmentation Strategies for Object Detection

https://github.com/tensorflow/tpu/tree/master/models/official/detection

• 比较有效的数据增强方法
• Rotate rotates the whole image and the bounding boxes
• Equalize flattens the histogram of the pixel values
• BBox Only TranslateY translates only the objects in bounding boxes vertically, up or down with equal probability
• learned data augmentation policy is effective across all data sizes considered, with a larger improvement when the training set is small. We also observe that the improvement due to a learned data augmentation policy is larger on harder tasks of detecting smaller objects and detecting with more precision.
• We carried out several experiments with Input Mixup, Manifold Mixup
  and Dropblock. For all methods we found that they either did not help nor hurt model performance
• 和其他一些正则化方法一起使用，无效，学习的数据增强方法，有一定的正则化作用
• This is an interesting result as the proposed method independently
  outperforms these regularization methods, yet apparently these regularization methods are not needed when applying a learned data augmentation policy.
• 不止是学习到的数据增强策略的作用，还有增加 anchors，增大图片尺寸

'cannot load library \‘libvips.so.42\’: libvips.so.42: cannot open shared object file: No such file—# import pyvips vip的原因，注释有关 pyvips 代码

https://github.com/fizyr/keras-models/releases/download/v0.0.1/ResNet-101-model.keras.h5