Abstract:
In recent years, deep neural network based shadow detection approaches achieve high performance on benchmarks, but they require large amounts of labeled data for training to learn the statistical attributes of the shadowed and unshadowed regions. The physical relationship between non-grid regions is ignored due to the character of deep convolution network. In this paper, we seek to analyze the physical principle of the shadowed region and present a paired region based algorithm for shadow detection without extra model training. To be specific, we first segment the image via a region growing based approach to maintain the character of shadowed region. Then the penumbra is detected and used as an important cue. After that we adopt three physics based confidence coefficients when comparing the color of two regions. At last, we design a novel objective function to find the best paired strategy and detect the shadowed region. The proposed approach is tested on three public datasets. The comparative results show that the proposed non-learning approach performs favorably against the state-of-the-art approaches.
