Fixing the train-test resolution discrepancy

14 Jun 2019Hugo TouvronAndrea VedaldiMatthijs DouzeHervé Jégou

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... (read more)

PDF Abstract

Evaluation results from the paper


 SOTA for Image Classification on ImageNet (using extra training data)

     Get a GitHub badge
Task Dataset Model Metric name Metric value Global rank Uses extra
training data
Compare
Fine-Grained Image Classification Birdsnap FixSENet-154 Accuracy 84.3% # 1
Fine-Grained Image Classification CUB-200-2011 FixSENet-154 Accuracy 88.7% # 4
Image Classification ImageNet FixResNet-50 CutMix Top 1 Accuracy 79.8% # 32
Image Classification ImageNet FixResNet-50 CutMix Top 5 Accuracy 94.9% # 24
Image Classification ImageNet FixResNet-50 CutMix Number of params 25.6M # 1
Image Classification ImageNet FixPNASNet-5 Top 1 Accuracy 83.7% # 9
Image Classification ImageNet FixPNASNet-5 Top 5 Accuracy 96.8% # 8
Image Classification ImageNet FixPNASNet-5 Number of params 86.1M # 1
Image Classification ImageNet FixResNet-50 Billion Top 1 Accuracy 82.5% # 16
Image Classification ImageNet FixResNet-50 Billion Top 5 Accuracy 96.6% # 10
Image Classification ImageNet FixResNet-50 Billion Number of params 25.6M # 1
Image Classification ImageNet FixResNeXt-101 32x48d Top 1 Accuracy 86.4% # 1
Image Classification ImageNet FixResNeXt-101 32x48d Top 5 Accuracy 98.0% # 1
Image Classification ImageNet FixResNeXt-101 32x48d Number of params 829M # 1
Image Classification ImageNet FixResNet-50 Top 1 Accuracy 79.1% # 38
Image Classification ImageNet FixResNet-50 Top 5 Accuracy 94.6% # 33
Image Classification ImageNet FixResNet-50 Number of params 25.6M # 1
Image Classification iNaturalist FixSENet-154 Top 1 Accuracy 75.4% # 1
Fine-Grained Image Classification NABirds FixSENet-154 Accuracy 89.2% # 1
Fine-Grained Image Classification Oxford 102 Flowers FixInceptionResNet-V2 Top-1 Error Rate 4.3% # 1
Fine-Grained Image Classification Oxford 102 Flowers FixInceptionResNet-V2 Accuracy 95.7% # 1
Fine-Grained Image Classification Oxford-IIIT Pets FixSENet-154 Top-1 Error Rate 5.2% # 1
Fine-Grained Image Classification Stanford Cars FixSENet-154 Accuracy 94.4% # 2