Segmentation with Noisy Labels via Diffusion Models
label-efficient segmentation using Stable Diffusion
Undergraduate thesis research at HKUST, advised by Prof. Yincong Chen.
Overview
Enhancing Segmentation from Datasets Containing Noisy Labels Using Generative Models
This work improves image segmentation when training data contains noisy labels, using generative diffusion models. Built on the label-efficient semantic segmentation framework with diffusion models (ICLR 2022).
Key Components
- Noisy-label learning: a novel approach for learning from noise-infused labels in both unsupervised and semi-supervised settings
- Stable Diffusion: unlabeled and noisy-labeled datasets utilized to improve segmentation performance
- Evaluation: segmentation quality validated on benchmark datasets
Resources
- Code: ddpm-segmentation (fork of the label-efficient segmentation framework)
Segmentation results. To be replaced with a figure from the project.