Reproduce “Re-labeling ImageNet: from Single to Multi-Labels, from Global to Localized Labels” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sangdoo Yun, Seong Joon Oh, Byeongho Heo, Dongyoon Han, Junsuk Choe, Sanghyuk Chun
NAVER AI Lab developed 'ReLabel,' a machine-driven approach to generate pixel-wise, multi-label annotations for ImageNet's training set, and 'LabelPooling,' a novel training scheme to leverage these localized labels. This methodology improves ImageNet classification accuracy by up to 1.4 percentage points on ResNet-50 and enhances model robustness and transfer learning capabilities.
Key insights
- ImageNet's single-label and global-label assumptions, particularly when combined with random cropping, introduce substantial noise that hinders model training.
- Automated, machine-driven re-labeling can efficiently provide cleaner, localized multi-label supervision at scale, overcoming the prohibitive cost of human annotation.
- Utilizing both multi-label and localized information is crucial for improving classification accuracy, robustness, and transferability of deep learning models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.