Single Image Shadow Detection using Multiple Cues in a Supermodular MRF - Paper Metadata

Aug 26, 2026

This paper presents a complete methodology for single image shadow detection driven by learned shadow appearance. It introduces a multi-kernel single-region SVM classifier for shadow-region classification, which supplies unary potentials for an MRF optimization augmented by pairwise terms over neighboring regions. The work also designs a novel boundary classifier for shadow boundaries on same-material surfaces, plus improved paired-region classifiers for same-material regions under identical or different illumination. The MRF optimization uses sparse pairwise potentials and reduces the balanced error rate by 53% versus state of the art on a benchmark dataset.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages12

Reading Time30 min

AccessFree to download

Abstract

1 Introduction

Single Image Shadow Detection using Multiple Cues in a Supermodular MRF Tomás F. Yago Vicente [tyagovicente@cs.stonybrook.edu](tyagovicente@cs.stonybrook.edu)[ ](tyagovicente@cs.stonybrook.edu)Chen-Ping Yu [cheyu@cs.stonybrook.edu](cheyu@cs.stonybrook.edu) Dimitris Samaras [samaras@cs.stonybrook.edu](samaras@cs.stonybrook.edu) Image Analysis Lab Computer Science Dept. Stony Brook University Stony Brook, NY, USA Abstract In this paper we propose a complete methodology for single image shadow detection based on the learned appearance of shadows. The basis of our method is a novel single region SVM classiﬁer with a multi-kernel model speciﬁcally tailored for shadow region classiﬁcation. This classiﬁer, which already outperforms much more complex methods, provides the unary potentials for an MRF optimization that also includes pairwise potentials encoding the relationships between neighboring regions in the image. We introduce a novel boundary classiﬁer for shadow boundaries cast over surfaces with the same material, and two improved paired regions classiﬁers; one for adjacent regions of the same material taken under the same illumination, and one for regions of same material taken under different illumination. The strength of the unary classiﬁer means that our MRF requires only relatively sparse pairwise potentials, resulting in a more efﬁcient and accurate optimization as can be seen in our experimental results. We reduce the balanced error rate by 53% compared to the state of the art on the latest shadow detection image dataset. 1 Introduction Shadows are commonly found in natural scenes. They result from occlusions of the scene's illumination source(s) . As such, shadows are an outcome of the complex interactions between the illuminants, the geometry, and the reﬂectance properties of the materials present in a scene. Shadows provide useful visual cues to infer the illumination of the scene, or the size and shape of objects. However, shadows are also well known to confound many computer vision tasks such as image segmentation, scene understanding, or object recognition. Single image shadow detection is a well studied problem. Earlier methods based on comparing gradients in the original image and its illumination invariant representation (where the effects of illumination changes are suppressed from the image), such as [7] [8], showed remarkable results in high quality images. However, their performance degrades signiﬁcantly with consumer grade photographs or web quality pictures [18] . More recent methods ([25],[18],[13],[10]) use image datasets with labelled shadow ground truth to learn the appearance of shadows in images. For instance, Zhu et al. [25] propose a set of shadow variant 􀀍c 2013. The copyright of this document resides with its authors. It may be distributed unchanged freely in print or electronic forms. and invariant features in monochromatic images to learn a shadow region classiﬁer, and reﬁne the results with a CRF. Lalonde et al. [18] focused on shadow boundaries on the ground. They train a shadow boundary classiﬁer (using Adaboost[5]) based on color and texture features, and combine it with scene layout cues from [12] using a CRF to encourage boundary continuity. Huang et al. [13] use a set of physically inspired features to train a shadow boundary pixel classiﬁer using a SVM. They join pixels conﬁdently predicted as shadow boundaries with weakly predicted adjacent pixels in a Canny-like manner. Shadow boundary methods ([13],[18]) achieve acceptable results but they fail to segment closed shadow contours consistently. Guo et al. [10] train two pairwise classiﬁers to ﬁnd pairs of regions inan image that share the same material and are viewed under the same illumination conditions (both in shadow or both not in shadow), and same material but illuminated differently (only one region in shadow) . They minimize an energy functional that combines the predictions of a single region classiﬁer

FAQs

What is the core idea of the proposed single-image shadow detection method?

The method learns shadow appearance using a multi-kernel single-region SVM and uses it as unary evidence within an MRF that also incorporates contextual boundary and illumination relationships between regions.

What classifiers are introduced to support the MRF optimization?

It introduces a novel boundary classifier for shadow boundaries over same-material surfaces and two improved paired-region classifiers: one for adjacent same-material regions under the same illumination, and one for same-material regions under different illumination.

How does the paper achieve more efficient optimization in the MRF?

Because the unary classifier is strong, the MRF needs only relatively sparse pairwise potentials, improving both efficiency and accuracy of the optimization.

Aug 26, 2026

Uploaded by 8796093062539