Image Enhanced Rotation Prediction for Self-Supervised Learning
arXiv:1912.11603
Abstract
The rotation prediction (Rotation) is a simple pretext-task for self-supervised learning (SSL), where models learn useful representations for target vision tasks by solving pretext-tasks. Although Rotation captures information of object shapes, it hardly captures information of textures. To tackle this problem, we introduce a novel pretext-task called image enhanced rotation prediction (IE-Rot) for SSL. IE-Rot simultaneously solves Rotation and another pretext-task based on image enhancement (e.g., sharpening and solarizing) while maintaining simplicity. Through the simultaneous prediction of rotation and image enhancement, models learn representations to capture the information of not only object shapes but also textures. Our experimental results show that IE-Rot models outperform Rotation on various standard benchmarks including ImageNet classification, PASCAL-VOC detection, and COCO detection/segmentation.
arXiv commentAccepted to IEEE ICIP 2021. The title has been changed from "Multiple Pretext-Task for Self-Supervised Learning via Mixing Multiple Image Transformations"
References in corpus (12)
- Very Deep Convolutional Networks for Large-Scale Image Recognition
- Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
- Texture Synthesis Using Convolutional Neural Networks
- Searching for Efficient Multi-Scale Architectures for Dense Image Prediction
- Multi-Task Learning as Multi-Objective Optimization
- Revisiting Unreasonable Effectiveness of Data in Deep Learning Era
- Data-dependent Initializations of Convolutional Neural Networks
- Approximating CNNs with Bag-of-local-Features models works surprisingly well on ImageNet
- Improvements to context based self-supervised learning
- High-Fidelity Image Generation With Fewer Labels
- Self-supervised Label Augmentation via Input Transformations
- Self-Supervised Feature Learning by Learning to Spot Artifacts