Distributed gradient-domain processing of planar and spherical images

Abstract

Gradient-domain processing is widely used to edit and combine images. In this paper we extend the framework in two directions. First, we adapt the gradient-domain approach to operate on a spherical domain, to enable operations such as seamless stitching, dynamic-range compression, and gradient-based sharpening over spherical imagery. An efficient streaming computation is obtained using a new spherical parameterization with bounded distortion and localized boundary constraints. Second, we design a distributed solver to efficiently process large planar or spherical images. The solver partitions images into bands, streams through these bands in parallel within a networked cluster, and schedules computation to hide the necessary synchronization latency. We demonstrate our contributions on several datasets including the Digitized Sky Survey, a terapixel spherical scan of the night sky.

Hybrid CPU-GPU Solver for Gradient Domain Processing of Massive Images

“…After processing an image's gradient field for the desired effect, these methods attempt to find a smooth image that is closest to the guiding gradient with a minimal least squared error. Gradient domain processing has been used for applications such as seamless cloning [8], drag-and-drop pasting [9], matting [10] and seamless panorama stitching [4], [6]-[8], [11], [12]. Other examples include compressing HDR (High Dynamic Range) images for display on standard monitors [13] and applications such as detection of lighting [14] or shapes [15] from images, shadow removal [16], reflections [17], and artistic editing in the gradient domain [18].…”

“…These are sequential methods and though efficient, still take several hours to compute solutions for gigapixel images. Hence, some work has been done recently to extend these methods to work under a distributed environment in [6] and [7] respectively. The implementation described in [7] has been shown to be portable across different systems and has good scalability in both the strong and weak sense.…”

Target Visualization at the National Ignition Facility

“…An image stitching algorithm could overcome these issues. [21], which deals with the combination of spherical images looks especially interesting. A more robust technique for determining pixel visibility using ray sphere intersection suggested during the review of this work may also be explored.…”

Massive Image Editing on the Cloud

“…For out-of-core processing of large images, the streaming multigrid method of Kazhdan and Hoppe [29] and the progressive Poisson method [39] have so far provided the only solutions. Recently, streaming multigrid has been extended to a distributed environment [30] and has reduced the time to process gigapixel images from hours to minutes. Out-of-core methods often achieve a low memory footprint at the cost of significant disk storage requirements.…”

“…Our approach solves directly from the original image data in one MapReduce pass by computing gradients on the fly. The multigrid method [29,30] may also be limited by main memory, since the number of iterations of the solver is directly proportional to the memory footprint. For large images, this limits the solver to only a few Gauss-Seidel iterations and therefore may not necessarily converge for challenging cases.…”