U-CFR reduces manual clicks by 10% for interactive segmentation

Interactive image segmentation often requires many manual clicks for precise annotation, with existing methods relying on passive refinement that converges slowly. Researchers propose Uncertainty-Guided Cascade Forward Refinement (U-CFR), an inference-time framework that enables models to autonomously self-correct after each user interaction. U-CFR introduces a boundary-aware uncertainty score that combines segmentation uncertainty, contour gradients, and explicit edge predictions to place internal pseudo-clicks on the most ambiguous boundary regions. These self-generated clicks provide strong corrective signals without extra manual input. A dual-head network with a shared encoder-decoder backbone supports this: a segmentation head ensures region consistency, while an edge head sharpens boundary alignment. During inference, U-CFR launches a cascade of refinement steps, where each stage uses uncertainty-driven pseudo-clicks to progressively improve the mask.

Experiments on standard benchmarks demonstrate U-CFR's effectiveness in improving click efficiency, initial mask quality, and boundary accuracy. On challenging datasets like Berkeley, it reduces the required clicks by over 10%. The method is detailed in a 12-page paper accepted at ICPR 2026, with 3 figures and 4 tables. For tech professionals, U-CFR represents a practical advance for image annotation tools, reducing manual effort significantly without complex retraining. By making the AI actively self-correct rather than passively respond, it lowers the barrier for high-quality segmentation in applications like medical imaging, autonomous driving, and satellite imagery analysis.