DOC: Deep OCclusion Recovering From A Single Image.
Peng Wang, Alan Yuille · arXiv (Cornell University) · 2015
Recovering the occlusion relationship between stuffs and things is a fundamental ability of human vision, and yields information about the 3D world from 2D images. In this paper, we propose deep occlusion (DOC), which is an end-to-end deep learning strategy that recovers the occlusion boundaries from a single image. Firstly, for each pixel along the boundaries, we introduce an orientation variable, indicating the occlusion relationship between two adjacent regions. Then, deep features, long range context and object-level knowledge are exploited by an end-to-end deep convolutional neural network (DCNN) to accurately locate the boundaries and recover the occlusion relationships. To evaluate and train the DOC network, we construct a large-scale instance occlusion boundary dataset from the PASCAL images, which we call the PASCAL instance occlusion dataset. It includes 20k images (100$\times$ bigger than the existing occlusion datasets for outdoor images) and so is big enough to train and evaluate DCNNs for occlusion learning. Finally, extensive experiments are performed to demonstrate that our approach has strong generalization ability and outperform state-of-the-art strategies.