Fixing the train-test resolution discrepancy

Hugo Touvron, Andrea Vedaldi, Matthijs Douze, Herve Jegou

Abstract: Data-augmentation is key to the training of neural networks for image classification. This paper first shows that existing augmentations induce a significant discrepancy between the typical size of the objects seen by the classifier at train and test time. We experimentally validate that, for a target test resolution, using a lower train resolution offers better classification at test time. We then propose a simple yet effective and efficient strategy to optimize the classifier performance when the train and test resolutions differ. It involves only a computationally cheap fine-tuning at the test resolution. This enables training strong classifiers on small images. For instance, we obtain 77.1% top-1 accuracy on ImageNet with a ResNet-50 trained on 128x128 images, and 78% with our multi-resolution classification. Conversely, when training a PNASNet at resolution 331x331 and further optimizing for test resolution 480x480, we obtain a test top-1 accuracy of 83.7% (top-5: 96.8%) (single-crop).

Code Link: https://github.com/facebookresearch/FixRes

CMT Num: 4477