Further Exploration of Deep Aggregation for Shadow Detection

Shadow detection is a fundamental challenge in the field of computer vision. It requires network to understand global semantics and local details image. All existing methods depend on aggregation features multi-stage pre-trained convolution neural but comparison high-level capabilities, low-level capabilities provide less performance. Using not only increases complex difficulty also reduces time efficiency. In this article, we propose new shadow detector, which uses explores complementary information between adjacent feature layers. Experiments show that technique paper can accurately detect shadows perform well compared with most advanced methods. The detailed experiments performed three public datasets SUB, UCF, ISTD demonstrate suggested method efficient stable.