Reproduce “Promoting Segment Anything Model towards Highly Accurate Dichotomous Image Segmentation” 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
Xianjie Liu, Keren Fu, Yao Jiang, Qijun Zhao
A two-stage framework called DIS-SAM is introduced, which enhances the Segment Anything Model (SAM) to achieve highly accurate dichotomous image segmentation while retaining its interactive, promptable design. The model demonstrates an 8.5% improvement in maximum F-measure over original SAM and superior zero-shot generalization capabilities across diverse datasets, coupled with reduced human correction efforts.
Key insights
- Integrating SAM's interactive, coarse segmentation with a specialized, high-precision refinement network effectively combines the benefits of foundation models with the demands of highly accurate segmentation.
- Utilizing the prompt box and SAM's coarse mask as explicit inputs to the refinement stage provides synergistic guidance for both object localization and fine-grained boundary correction.
- Data enrichment techniques and architectural regularizations like Parameter Orthogonalization are essential for improving the robustness and generalizability of the refinement model when fine-tuning on diverse datasets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.