The Fast Bilateral Solver
We present the bilateral solver, a novel algorithm for edge-aware smoothing that combines the flexibility and speed of simple filtering approaches with the accuracy of domain-specific optimization algorithms. Our technique is capable of matching or improv
- PDF / 2,762,789 Bytes
- 16 Pages / 439.37 x 666.142 pts Page_size
- 32 Downloads / 265 Views
2
Google, Mountain View, USA [email protected] Stanford University, Stanford, USA [email protected]
Abstract. We present the bilateral solver, a novel algorithm for edgeaware smoothing that combines the flexibility and speed of simple filtering approaches with the accuracy of domain-specific optimization algorithms. Our technique is capable of matching or improving upon state-ofthe-art results on several different computer vision tasks (stereo, depth superresolution, colorization, and semantic segmentation) while being 10–1000× faster than baseline techniques with comparable accuracy, and producing lower-error output than techniques with comparable runtimes. The bilateral solver is fast, robust, straightforward to generalize to new domains, and simple to integrate into deep learning pipelines.
1
Introduction
Images of the natural world exhibit a useful prior – many scene properties (depth, color, object category, etc.) are correlated within smooth regions of an image, while differing across discontinuities in the image. Edge-aware smoothing techniques exploit this relationship to propagate signals of interest within, but not across edges present in an image. Traditional approaches to edge-aware smoothing apply an image-dependent filter to a signal of interest. Examples of this include joint bilateral filtering [37,40] and upsampling [20], adaptive manifolds [12], the domain transform [11], the guided filter [16,17], MST-based filtering [41], and weighted median filtering [30,44]. These techniques are flexible and computationally efficient, but often insufficient for solving more challenging computer vision tasks. Difficult tasks often necessitate complex iterative inference or optimization procedures that encourage smoothness while maintaining fidelity with respect to some observation. Optimization algorithms of this nature have been used in global stereo [34], depth superresolution [10,19,24,26,29,32], colorization [25], and semantic segmentation [6,22,28,45]. These approaches are tailored to their specific task, and are generally computationally expensive. In this work we present an optimization algorithm that is 10–1000× faster than existing domain-specific approaches with comparable accuracy, and produces higher-quality output than lightweight filtering techniques with comparable runtimes. Electronic supplementary material The online version of this chapter (doi:10. 1007/978-3-319-46487-9 38) contains supplementary material, which is available to authorized users. c Springer International Publishing AG 2016 B. Leibe et al. (Eds.): ECCV 2016, Part III, LNCS 9907, pp. 617–632, 2016. DOI: 10.1007/978-3-319-46487-9 38
618
J.T. Barron and B. Poole
(a) Input (MAE = 6.00, RMSE = 38.8)
(b) Output (MAE = 3.02, RMSE = 17.9)
(c) Input Confidence
(d) Input Reference
Fig. 1. The bilateral solver can be used to improve depth maps. A depth map (a) from a state-of-the-art stereo method [43] is processed with our robust bilateral solver using a reference RGB image (d). Our output (b) is smooth with respect to the reference image,
Data Loading...
