Generalizable and Consistent Granular Edge Prediction

ICLR 2026 Conference Submission6772 Authors

Keywords: Edge detection; Granular edge prediction

TL;DR: We propose a new task, Granular Edge Prediction, supported by a synthetic dataset and a novel model with consensus-based training, achieving state-of-the-art zero-shot generalization and consistent, human-aligned edge predictions

Abstract: We introduce a new task in edge detection: Granular Edge Prediction. Unlike traditional binary edge maps, this task aims to predict a categorical edge map, where each edge pixel is assigned a granularity level reflecting the likelihood of being recognized as an edge by a human annotator. Our contributions are threefold: 1) we construct a large-scale synthetic dataset for granular edge prediction, where each edge is labeled with a quantized granularity level, and introduce a graph-based edge representation to enforce consistency in edge granularity across the dataset, 2) we develop a novel edge consensus loss to enforce granularity consistency within individual edges, and 3) we propose a comprehensive evaluation framework, including granularity-aware edge evaluation and two quantitative metrics to assess the consistency of granular edge prediction. Extensive experiments demonstrate that our method generalizes well in zero-shot evaluation across four standard edge detection datasets, closely aligns with human perception of edge granularity, and ensures high consistency in edge-wise granularity estimation.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 6772