Reproduce “Low-Biased General Annotated Dataset Generation” 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
Dengyang Jiang, Haoyu Wang, Lei Zhang, Wei Wei, Guang Dai, Mengmeng Wang, Jingdong Wang, Yanning Zhang
Researchers at Northwestern Polytechnical University and collaborators developed lbGen, a framework for generating low-biased general annotated datasets by fine-tuning diffusion models. This approach consistently enhanced the generalization capacity and robustness of deep neural networks across various downstream tasks and bias evaluations, particularly benefiting few-shot learning scenarios.
Key insights
- Language information, as processed by multimodal foundation models like CLIP, provides a representation closer to ideal semantic information and can effectively define a 'low-biased semantic space' for guiding image generation.
- A bi-level semantic alignment strategy is crucial for both ensuring overall dataset-level consistency and precise individual image-to-category matching, preventing class ambiguity and model collapse during generation.
- Explicitly incorporating an image quality assessment mechanism is necessary to prevent visual quality deterioration when relying primarily on text-semantic supervision for generating synthetic data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.