This repository provides a comprehensive survey and codebase for deep long-tailed learning, addressing the challenge of imbalanced datasets where a few classes dominate while many have few samples. It targets researchers and practitioners in computer vision and machine learning, offering a structured overview of techniques and practical implementations.

How It Works

The project categorizes long-tailed learning methods into three main groups: class re-balancing, information augmentation, and module improvement, with further sub-classifications. It includes an empirical analysis of state-of-the-art methods using novel metrics like Upper Bound Accuracy (UA) and Relative Accuracy (RA) to better assess performance gains specifically from addressing class imbalance, distinguishing it from improvements due to architecture or general data augmentation.

Quick Start & Requirements

• Installation: pip install -r requirements.txt
• Hardware: 4 GPUs with >= 23GB GPU RAM recommended.
• Dataset: Requires ImageNet-1K dataset downloaded and placed in ./data/ImageNet.
• Training: Example commands provided for various methods (e.g., Softmax, Weighted Softmax, ESQL, MiSLAS) using Python scripts within specific subdirectories (Main-codebase, MiSLAS-codebase, etc.).
• Resources: Links to Papers With Code, other GitHub repos, and Star History are available.

Highlighted Details

• Comprehensive survey paper (TPAMI 2023) with a structured taxonomy of long-tailed learning techniques.
• Curated list of top-tier conference papers (NeurIPS, ICML, CVPR, ECCV, ICCV) with links to official codebases.
• Benchmark datasets (ImageNet-LT, CIFAR100-LT, LVIS, etc.) are listed with relevant statistics.
• Empirical studies comparing methods using accuracy, UA, and RA metrics, identifying SADE as a top performer.
• Codebases for multiple state-of-the-art methods are included (e.g., MiSLAS, RSG, ResLT, PaCo, RIDE, SADE).

Maintenance & Community

The project is associated with a TPAMI 2023 survey paper, indicating academic backing. No specific community channels (Discord, Slack) or active development updates are mentioned in the README.

Licensing & Compatibility

The repository does not explicitly state a license. The codebase is provided for community use, but commercial use or linking with closed-source projects would require clarification on licensing terms.

Limitations & Caveats

The README does not specify a license, which could be a barrier for commercial adoption. While multiple codebases are provided, they are in separate subdirectories, potentially leading to dependency conflicts or varied setup procedures. The hardware requirements are substantial.