Abstract

In this work we empirically study the multi-scale boundary detection problem in natural images. We utilize local boundary cues including contrast, localization and relative contrast, and train a classi fier to integrate them across scales. Our approach successfully combines strengths from both large-scale de- tection (robust but poor localization) and small-scale detection (detail-preserving but sensitive to clutter). We carry out quantitative evaluations on a variety of boundary and object datasets with human-marked groundtruth. We show that multi-scale boundary detection offers large improvements, ranging from 20% to 50%, over single-scale approaches. This is the first time that multi-scale is demonstrated to improve boundary detection on large datasets of natural images.