DOC: Deep OCclusion Estimation from a Single Image
In this paper, we propose a deep convolutional network architecture, called DOC, which detects object boundaries and estimates the occlusion relationships (i.e. which side of the boundary is foreground and which is background). Specifically, we first repr
- PDF / 4,077,002 Bytes
- 17 Pages / 439.37 x 666.142 pts Page_size
- 52 Downloads / 196 Views
University of California, Los Angeles, USA [email protected] John Hopkins University, Baltimore, USA [email protected]
Abstract. In this paper, we propose a deep convolutional network architecture, called DOC, which detects object boundaries and estimates the occlusion relationships (i.e. which side of the boundary is foreground and which is background). Specifically, we first represent occlusion relations by a binary edge indicator, to indicate the object boundary, and an occlusion orientation variable whose direction specifies the occlusion relationships by a left-hand rule, see Fig. 1. Then, our DOC networks exploit local and non-local image cues to learn and estimate this representation and hence recover occlusion relations. To train and test DOC, we construct a large-scale instance occlusion boundary dataset using PASCAL VOC images, which we call the PASCAL instance occlusion dataset (PIOD). It contains 10,000 images and hence is two orders of magnitude larger than existing occlusion datasets for outdoor images. We test two variants of DOC on PIOD and on the BSDS ownership dataset and show they outperform state-of-the-art methods typically by more than 5AP. Finally, we perform numerous experiments investigating multiple settings of DOC and transfer between BSDS and PIOD, which provides more insights for further study of occlusion estimation.
1
Introduction
Humans are able to recover the occlusion relationships of objects from single images. This has long been recognized as an important ability for scene understanding and perception [4,15]. As shown on the left of Fig. 1, we can use occlusion relationships to deduce that the person is holding a dog, because the person’s hand occludes the dog and the dog occludes the person’s body. Electrophysiological [18] and fMRI [13] studies suggest that occlusion relationships are detected as early as visual area V2. Biological studies [9] also suggest that occlusion detection can require feedback from higher level cortical regions, indicating that long-range context and semantic-level knowledge may be needed. Psychophysical studies show that there are many cues for occlusion including edge convexity [23], edge-junctions, intensity gradients, and texture [35]. Electronic supplementary material The online version of this chapter (doi:10. 1007/978-3-319-46448-0 33) contains supplementary material, which is available to authorized users. c Springer International Publishing AG 2016 B. Leibe et al. (Eds.): ECCV 2016, Part I, LNCS 9905, pp. 545–561, 2016. DOI: 10.1007/978-3-319-46448-0 33
546
P. Wang and A. Yuille
Fig. 1. Left: Occlusion boundaries represented by orientation θ (the red arrows), which indicates occlusion relationship using the “left” rule where the left side of the arrows is foreground. Right: More examples from our Pascal instance occlusion dataset (PIOD). (Color figure online)
Computer vision researchers have also used similar cues for estimating occlusion relations. A standard strategy is to apply machine learning techniques to combine cues li
Data Loading...
