Fast Shadow Detection from a Single Image Using a Patched Convolutional Neural Network

Abstract

In recent years, various shadow detection methods from a single image have been proposed and used in vision systems; however, most of them are not appropriate for the robotic applications due to the expensive time complexity. This paper introduces a fast shadow detection method using a deep learning framework, with a time cost that is appropriate for robotic applications. In our solution, we first obtain a shadow prior map with the help of multi-class support vector machine using statistical features. Then, we use a semanticaware patch-level Convolutional Neural Network that efficiently trains on shadow examples by combining the original image and the shadow prior map. Experiments on benchmark datasets demonstrate the proposed method significantly decreases the time complexity of shadow detection, by one or two orders of magnitude compared with state-of-the-art methods, without losing accuracy.

Direction-Aware Spatial Context Features for Shadow Detection

“…We compare our method with four recent shadow detection methods: scGAN [18], stacked-CNN [24], patched-CNN [6] and Unary-Pairwise [5]. Among them, the first three are deep-learning-based methods, while the last one is based on hand-crafted features.…”

“…For a fair comparison, the shadow detection results of other methods are obtained either directly from results provided by the authors, or by generating them using implementations provided by the authors with recommended parameter setting. input image ground truth DSC (ours) scGAN [18] stkd'-CNN [24] patd'-CNN [6] SRM [26] Amulet [28] PSPNet [29] Figure 5: Visual comparison of shadow maps produced by our method and other methods (4th-9th columns) against ground truths shown in 2nd column. Note that stkd'-CNN and patd'-CNN stand for stacked-CNN and patched-CNN, respectively.…”

“…Note that stkd'-CNN and patd'-CNN stand for stacked-CNN and patched-CNN, respectively. Table 1: Comparing our method (DSC) with state-of-thearts methods for shadow detection (scGAN [18], stacked-CNN [24], patched-CNN [6] and Unary-Pairwise [5]), for saliency detection (SRM [26] and Amulet [28]), and for semantic image segmentation (PSPNet [29] [18] 0.90 9.10 0.87 11.50 stacked-CNN [24] 0.88 11.00 0.85 13.00 patched-CNN [6] 0.88 11.56 --Unary-Pairwise [5] 0.86 25.03 --SRM [26] 0.96 7.25 0.94 9.81 Amulet [28] 0.93 15.13 0.92 15.17 PSPNet [29] 0.95 8.57 0.93 11.75 Table 1 reports the comparison results, where we can see that our method outperforms the others in terms of both accuracy and BER for both benchmark datasets. Note that our shadow detection network is trained using the SBU training set [24], but it still outperforms the others also for the UCF dataset, thus demonstrating its generalization ability.…”

“…The last three rows in Table 1 report the comparison results in terms of the accuracy and BER metrics. Although these methods achieve good re-input image ground truth DSC (ours) scGAN [18] stkd'-CNN [24] patd'-CNN [6] SRM [26] Amulet [28] PSPNet [29] Figure 6: More visual comparison results (continue from Figure 5). [24]: "basic" denotes the architecture shown in Figure 4 but without all DSC modules; "basic+context" denotes the "basic" network with spatial context but not direction-aware spatial context; and "DSC" is the overall network in Figure…”

“…Very recently, Hosseinzadeh et al [6] detected shadows using a patch-level CNN and a shadow prior map generated from hand-crafted features, while Nguyen et al [18] developed scGAN with a sensitivity parameter to adjust weights in the loss functions. Although the shadow detection accuracy keeps improving on the benchmark datasets [30,24], existing methods may still misrecognize black objects as shadows and miss unobvious shadows in the testing images.…”