Test-Time Augmentation is a very efficient way to improve the results of your model at testing time.
After having trained our network, we can show it a test image and see it prediction.
Here is the query image :
Its prediction is :
While the true label is :
So the test image is misclassified.
Meaning that we will show the model several slightly modified versions of the image and average its predictions.
The 5 modified images are :
And the corresponding predictions :
By now taking the average we have:
And the image is correctly classified !
Further explanation can be found here