Reproduce “Boundary IoU: Improving Object-Centric Image Segmentation Evaluation” 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
Bowen Cheng, Ross Girshick, Piotr Dollár, Alexander C. Berg, Alexander Kirillov
A new evaluation metric called Boundary IoU is introduced to specifically assess and reward the quality of mask boundaries in object-centric image segmentation. This metric, and its derived task-level counterparts Boundary AP and Boundary PQ, demonstrate enhanced sensitivity to boundary errors for large objects, providing a more balanced assessment of mask fidelity across various scales than standard Mask IoU.
Key insights
- Standard Mask IoU is inherently biased towards large objects and less sensitive to their boundary errors, causing a disconnect between reported scores and perceptual mask quality.
- Boundary IoU provides a symmetric and robust measure that maintains consistent sensitivity to boundary quality across various object scales, effectively addressing the shortcomings of previous metrics.
- The new task-level metrics offer a clearer quantitative signal for mask boundary fidelity, which can help redirect research efforts towards developing more precise segmentation models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.