Fixing the train-test resolution discrepancy

Hugo TouvronAndrea VedaldiMatthijs DouzeHervé Jégou

   Papers with code   Abstract  PDF

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)

Benchmarked Models

RANK
MODEL
REPO
CODE RESULT
PAPER RESULT
ε-REPRODUCED
BUILD
2
fixEfficientNet_b7_ns
87.1%
--
3
fixEfficientNet_b6_ns
86.7%
--
4
fixEfficientNet_b5_ns
86.4%
--
5
FixResNeXt-101 32x48d
86.4%
86.4%
6
fixEfficientNet_b4_ns
85.9%
--
9
fixEfficientNet_b3_ns
85.0%
--
13
FixPNASNet-5
83.7%
83.7%
14
fixEfficientNet_b2_ns
83.6%
--
16
fixEfficientNet_b1_ns
82.6%
--
19
fixEfficientNet_b0_ns
80.2%
--
20
FixResNet-50 CutMix
79.8%
79.8%
22
FixResNet-50
79.1%
79.1%