Timezone: »
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 size of the objects seen by the classifier at train and test time: in fact, a lower train resolution improves the classification at test time!
We then propose a simple strategy to optimize the classifier performance, that employs different train and test resolutions. It relies on a computationally cheap fine-tuning of the network at the test resolution. This enables training strong classifiers using small training images, and therefore significantly reduce the training time. For instance, we obtain 77.1% top-1 accuracy on ImageNet with a ResNet-50 trained on 128x128 images, and 79.8% with one trained at 224x224.
A ResNeXt-101 32x48d pre-trained with weak supervision on 940 million 224x224 images and further optimized with our technique for test resolution 320x320 achieves 86.4% top-1 accuracy (top-5: 98.0%). To the best of our knowledge this is the highest ImageNet single-crop accuracy to date.
Author Information
Hugo Touvron (Facebook AI Research)
Andrea Vedaldi (Facebook AI Research and University of Oxford)
Matthijs Douze (Facebook AI Research)
Herve Jegou (Facebook AI Research)
More from the Same Authors
-
2021 : ResNet strikes back: An improved training procedure in timm »
Ross Wightman · Hugo Touvron · Herve Jegou -
2022 : Direct LiDAR-based object detector training from automated 2D detections »
Robert McCraith · Eldar Insafutdinov · Lukas Neumann · Andrea Vedaldi -
2023 Poster: Contrastive Lift: 3D Object Instance Segmentation by Slow-Fast Contrastive Fusion »
Yash Bhalgat · Iro Laina · João Henriques · Andrea Vedaldi · Andrew Zisserman -
2023 Poster: Improving Category Discovery When No Representation Rules Them All »
Sagar Vaze · Andrea Vedaldi · Andrew Zisserman -
2023 Poster: EPIC Fields: Marrying 3D Geometry and Video Understanding »
Vadim Tschernezki · Ahmad Darkhalil · Zhifan Zhu · David Fouhey · Iro Laina · Diane Larlus · Dima Damen · Andrea Vedaldi -
2022 Poster: Unsupervised Multi-Object Segmentation by Predicting Probable Motion Patterns »
Laurynas Karazija · Subhabrata Choudhury · Iro Laina · Christian Rupprecht · Andrea Vedaldi -
2021 : Spotlight talk: ResNet strikes back: An improved training procedure in timm. »
Hugo Touvron -
2021 : Image Similarity Challenge + Q&A »
Matthijs Douze · Zoe Papakipos · Cristian Canton · Lowik Chanussot · Giorgos Tolias · Filip Radenovic · Ondrej Chum -
2021 Poster: XCiT: Cross-Covariance Image Transformers »
Alaaeldin Ali · Hugo Touvron · Mathilde Caron · Piotr Bojanowski · Matthijs Douze · Armand Joulin · Ivan Laptev · Natalia Neverova · Gabriel Synnaeve · Jakob Verbeek · Herve Jegou -
2021 : Billion-Scale Approximate Nearest Neighbor Search Challenge + Q&A »
Harsha Vardhan Simhadri · George Williams · Martin Aumüller · Artem Babenko · Dmitry Baranchuk · Qi Chen · Matthijs Douze · Ravishankar Krishnawamy · Gopal Srinivasa · Suhas Jayaram Subramanya · Jingdong Wang -
2020 Poster: Quantifying Learnability and Describability of Visual Concepts Emerging in Representation Learning »
Iro Laina · Ruth Fong · Andrea Vedaldi -
2020 Poster: RELATE: Physically Plausible Multi-Object Scene Synthesis Using Structured Latent Spaces »
Sebastien Ehrhardt · Oliver Groth · Aron Monszpart · Martin Engelcke · Ingmar Posner · Niloy Mitra · Andrea Vedaldi -
2019 Poster: Large Memory Layers with Product Keys »
Guillaume Lample · Alexandre Sablayrolles · Marc'Aurelio Ranzato · Ludovic Denoyer · Herve Jegou -
2019 Spotlight: Large Memory Layers with Product Keys »
Guillaume Lample · Alexandre Sablayrolles · Marc'Aurelio Ranzato · Ludovic Denoyer · Herve Jegou -
2018 Poster: Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks »
Jie Hu · Li Shen · Samuel Albanie · Gang Sun · Andrea Vedaldi -
2018 Poster: Modelling and unsupervised learning of symmetric deformable object categories »
James Thewlis · Hakan Bilen · Andrea Vedaldi -
2018 Poster: Unsupervised Learning of Object Landmarks through Conditional Image Generation »
Tomas Jakab · Ankush Gupta · Hakan Bilen · Andrea Vedaldi -
2017 Workshop: Interpreting, Explaining and Visualizing Deep Learning - Now what ? »
Klaus-Robert Müller · Andrea Vedaldi · Lars K Hansen · Wojciech Samek · Grégoire Montavon -
2017 Poster: Learning multiple visual domains with residual adapters »
Sylvestre-Alvise Rebuffi · Hakan Bilen · Andrea Vedaldi -
2017 Spotlight: Learning multiple visual domains with residual adapters »
Sylvestre-Alvise Rebuffi · Hakan Bilen · Andrea Vedaldi -
2017 Poster: Unsupervised learning of object frames by dense equivariant image labelling »
James Thewlis · Hakan Bilen · Andrea Vedaldi -
2017 Oral: Unsupervised learning of object frames by dense equivariant image labelling »
James Thewlis · Hakan Bilen · Andrea Vedaldi -
2016 Poster: Learning feed-forward one-shot learners »
Luca Bertinetto · João Henriques · Jack Valmadre · Philip Torr · Andrea Vedaldi -
2016 Poster: Integrated perception with recurrent multi-task neural networks »
Hakan Bilen · Andrea Vedaldi -
2013 Poster: Deep Fisher Networks for Large-Scale Image Classification »
Karen Simonyan · Andrea Vedaldi · Andrew Zisserman -
2013 Spotlight: Deep Fisher Networks for Large-Scale Image Classification »
Karen Simonyan · Andrea Vedaldi · Andrew Zisserman -
2011 Poster: Pylon Model for Semantic Segmentation »
Victor Lempitsky · Andrea Vedaldi · Andrew Zisserman -
2010 Poster: Simultaneous Object Detection and Ranking with Weak Supervision »
Matthew B Blaschko · Andrea Vedaldi · Andrew Zisserman -
2009 Poster: Structured output regression for detection with partial truncation »
Andrea Vedaldi · Andrew Zisserman -
2006 Poster: A Rate-Distortion Approach to Joint Pattern Alignment »
Andrea Vedaldi