Learning from few images for image recognition

Paper Title:

Towards Few-Annotation Learning in Computer Vision: Application to Image Classification and Object Detection tasks

Published on:

8 November 2023

Primary Category:

Computer Vision and Pattern Recognition

Paper Authors:

Quentin Bouniot

Bullets

Key Details

•

Enforces assumptions from multi-task representation learning theory to improve meta-learning algorithms

•

Introduces localization information in contrastive learning objectives for pretraining object detectors

•

Shows improved generalization and sample efficiency when training models with few annotated images

•

Proposes regularization techniques to guide meta-learning algorithms

•

Leverages object proposals in images as positives examples for contrastive learning

Explore the topics in this paper

contrastive learning

image recognition

meta-learning

sample efficiency

transfer learning

AI generated summary

Learning from few images for image recognition

This paper proposes methods to learn image recognition models more efficiently from limited annotated data. The authors develop techniques in meta-learning and transfer learning to improve model generalization and sample efficiency when training with few images per class. Key ideas include enforcing assumptions from multi-task representation learning theory in meta-learning algorithms, and introducing localization information in contrastive learning objectives for pretraining object detectors.