Delving into Dark Regions for Robust Shadow Detection

TL;DR

This paper tackles the challenge of shadow detection, particularly the poor performance in dark regions, by proposing a dark-region focusing framework. The approach first builds global contextual understanding via a Global Context Network, then uses a Dark-Region Recommendation module to identify error-prone dark areas, followed by a Dark-Aware Shadow Analysis module that learns shadow features specifically within those regions. Experimental results on three datasets show state-of-the-art BER, validating the effectiveness of focusing on dark regions for robust shadow detection, with code made available for reproducibility. The method has practical implications for improving downstream vision tasks that rely on accurate shadow masks, especially under challenging illumination conditions.

Abstract

Shadow detection is a challenging task as it requires a comprehensive understanding of shadow characteristics and global/local illumination conditions. We observe from our experiment that state-of-the-art deep methods tend to have higher error rates in differentiating shadow pixels from non-shadow pixels in dark regions (ie, regions with low-intensity values). Our key insight to this problem is that existing methods typically learn discriminative shadow features from the whole image globally, covering the full range of intensity values, and may not learn the subtle differences between shadow and non-shadow pixels in dark regions. Hence, if we can design a model to focus on a narrower range of low-intensity regions, it may be able to learn better discriminative features for shadow detection. Inspired by this insight, we propose a novel shadow detection approach that first learns global contextual cues over the entire image and then zooms into the dark regions to learn local shadow representations. To this end, we formulate an effective dark-region recommendation (DRR) module to recommend regions of low-intensity values, and a novel dark-aware shadow analysis (DASA) module to learn dark-aware shadow features from the recommended dark regions. Extensive experiments show that the proposed method outperforms the state-of-the-art methods on three popular shadow detection datasets. Code is available at https://github.com/guanhuankang/ShadowDetection2021.git.

Figures (13)

• Figure 1: Distribution of BER scores (the smaller the better) of state-of-the-art shadow detection methods and Ours (red) on SBU DBLP:conf/eccv/VicenteHYHS16. It shows that dark ranges contain more errors than bright ranges. It also reveals that our method performs better at low-intensity ranges, e.g., intensity-32 to intensity-48, than state-of-the-art methods DBLP:conf/cvpr/Chen0WW0H20DBLP:conf/cvpr/ZhengQCL19DBLP:conf/cvpr/Hu0F0H18DBLP:conf/eccv/ZhuDHFXQH18.
• Figure 2: Architecture of the proposed shadow detection framework. First, the global context network extracts multi-scale global contextual cues from the entire input image. Second, the dark-region recommendation (DRR) module recommends dark regions from the input image and forwards them to all dark-aware shadow analysis (DASA) modules. Third, the DASA modules at different layers take the corresponding global contextual cues and the recommended dark regions to produce dark-aware shadow features. Finally, the dark-aware shadow features across different layers are concatenated to predict the final shadow mask.
• Figure 3: Global Context Network. We choose ResNeXt-101 DBLP:conf/cvpr/XieGDTH17 as the backbone to extract features from the input image. To reduce memory footprint, we perform a dimensionality reduction on the backbone features to obtain $DF_i$, $i\in \{1,2,3,4,5\}$. We then employ a series of short connections DBLP:journals/pami/HouCHBTT19 to connect the outputs of the deeper fusion blocks to the shallower fusion blocks in the fusion stage. We further adopt a deep supervision strategy on the fusion features $FF_i$, $i\in \{1,2,3,4\}$, to accelerate the convergence. Finally, we concatenate $DF_i$ and $FF_i$ in each layer ($i\in \{1,2,3,4\}$) as our global contextual cues for the DASA modules.
• Figure 4: Fusion strategy comparison. Our fusion strategy allows more high level semantics of upper-stream features being exploited by down-stream features.
• Figure 5: Image-level average intensity histograms of two popular datasets, SBU and ISTD.