Fixing the train-test resolution discrepancy

Conference item

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 of the network at the test resolution. This enables training strong classifiers using small training images. For instance, we obtain 77.1% top-1 accuracy on ImageNet with a ResNet-50 trained on 128×128 images, and 79.8% with one trained on 224×224 image. In addition, if we use extra training data we get 82.5% with the ResNet-50 train with 224×224 images.
Conversely, when training a ResNeXt-101 32x48d pretrained in weakly-supervised fashion on 940 million public images at resolution 224×224 and further optimizing for test resolution 320×320, we obtain a test top-1 accuracy of 86.4% (top-5: 98.0%) (single-crop). To the best of our knowledge this is the highest ImageNet single-crop, top-1 and top-5 accuracy to date.

Authors: Touvron, H; Vedaldi, A; Douze, M; Jégou, H

• Publication status: Published
• Peer review status: Reviewed (other)
• Publisher: Neural Information Processing Systems Foundation
• Host title: Advances in Neural Information Processing Systems 32 (NIPS 2019) pre-proceedings
• Journal: Advances in Neural Information Processing Systems (NeurIPS), 2019
• Publication date: 2019-11-30
• Acceptance date: 2019-07-22
• Keywords: FFR