Reproduce “Scaling Open-Vocabulary Image Segmentation with Image-Level 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
Golnaz Ghiasi, Xiuye Gu, Yin Cui, Tsung-Yi Lin
OpenSeg enables open-vocabulary image segmentation by combining class-agnostic mask proposals with visual-semantic alignment learned from image-level captions. This approach achieves superior generalization and significantly outperforms prior open-vocabulary methods, demonstrating up to +19.9 mIoU improvement on PASCAL-20 compared to LSeg.
Key insights
- Explicit class-agnostic mask proposals (visual grouping) are crucial for precise localization in open-vocabulary image segmentation, complementing image-level semantic understanding.
- Training with diverse image-level captions on a large scale is essential for building "generalist" segmentation models that generalize well to unseen categories and datasets.
- The combination of a strong vision-language pre-trained backbone (ALIGN) with a mask proposal network and weak caption supervision effectively bridges the gap between image-level classification and pixel-level segmentation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.