TTA 测试时数据增强(multi-scale testing)(TEST.AUG)

hxxjxw

已于 2022-11-02 01:16:08 修改

阅读量2.3k

点赞数 1

文章标签：深度学习

于 2022-05-13 20:42:18 首次发布

本文链接：https://blog.csdn.net/hxxjxw/article/details/124759533

版权

TTA, Test-Time Augmentation，测试时数据增强

测试时将原始数据做不同形式的增强,然后取结果的平均值作为最终结果。可以进一步提升最终结果的精度

The input size significantly influences detection accuracy, since high resolution inputs make the detectors “seeing” small objects clearly to increase successful detections. The multi-scale testing can make the detector trained with limited input size (e.g., 320x320) to 'see' those small objects that only can be 'see' by the large input size (e.g., 1000x600).

detectron2和mmdetection中的TEST.AUG指的就是TTA

但是TTA时用的数据增强方法一般都比较简单，比如水平翻转，垂直翻转，90度旋转等

不太会用mixup，光照等这种复杂的

复杂的增强容易导致识别结果下降

YoloV5 Pseudo Labeling + OOF Evaluation | Kaggle