Developer-friendly segmentation using OpenVL, a high-level task-based abstraction

Abstract

Research into computer vision techniques has far outpaced the development of interfaces (such as APIs) to support the techniques' accessibility, especially to developers who are not experts in the field. We present a new interface, specifically for segmentation methods, designed to be application-developer-friendly while retaining sufficient power and flexibility to solve a wide variety of problems. The interface presents segmentation at a higher level (above algorithms) and uses a task-based description derived from definitions of low-level segmentation. We show that through interpretation, the description can be used to invoke an appropriate method to provide the developer's requested result. Our proof-of-concept implementation interprets the model description and invokes one of six segmentation methods with automatically derived parameters, which we demonstrate on a range of segmentation tasks. We also discuss how the concepts presented for segmentation may be extended to other computer vision problems.

Task-based control of articulated human pose detection for OpenVL

“…colour, texture, intensity), how much it varies (the distinctiveness of each segment) and perhaps the size or quantity of segments expected. (This builds on previous work by Miller et al that created a task-based abstraction for segmentation [19].) The general appearance of the scene can also be derived from this segmentation description -this is used to compare against the target appearance (discussed in the following section).…”

“…This idea has been applied successfully in many other fields, notably graphics with OpenGL, and is the main goal of OpenVL [18] for computer vision. OpenVL is an abstraction framework which hides algorithmic detail and provides mainstream developers with access to sophisticated vision methods, such as face detection [11] and segmentation [19]. The work presented here applies a similar methodology to construct a task-based description of pose estimation at a low enough level to maintain flexibility but high enough for mainstream developers to apply successfully.…”

An Abstraction for Correspondence Search Using Task-Based Controls

“…This idea has been applied successfully in many other fields, notably graphics with OpenGL, and is the main goal of OpenVL [1] for computer vision. OpenVL is an abstraction framework which hides algorithmic detail and provides developers with access to sophisticated vision methods, such as segmentation [2], human body pose estimation [3] and face detection [4]. The work presented here applies a similar methodology to construct a task-based description of correspondence search at a low enough level to maintain flexibility but high enough for mainstream developers to apply successfully, within the OpenVL framework.…”