ivl by Example A Scientific Computing Library, Applied to Learning and Vision

Researchers, engineers and professionals in any discipline of scienti c computing often prefer a computing and visual-ization environment like Matlab1 for everyday work, es-pecially when trying out a new idea. On the other hand, joint development of large projects with emphasis on perfor-mance, ease of integration, or genericity, typically requires a full- edged language like C.++, particularly so for open source, cross-platform software. Transition between these two worlds often ends up in painful code re-writing. Our approach is to bring, among others, elements of the Matlab syntax into C.++, where they can become even more powerful, blended e.g. with classes, templates, or the standard library (STL). Our open source template library ivl extends language syntax towards mathematical nota-tion and serves as a foundation in scientic computing rather than a collection of problem-oriented algorithm implementa-tions. As such, we choose to illustrate it through a concrete example in the eld of machine learning and computer vi-sion, followed by a brief outline of its design principles and major features. 1. ORIENTATION Instead of explaining what ivl is, we shall attempt in sec-tion 2 to show what it looks like by building a randomized decision forest classier from scratch. Trained by one input image and associated ground truth of pixel-level class labels, the classier learns to classify individual pixels of a new im-age into one of 23 object classes like `building', ` ower', or `car', plus a `void ' (unknown) class, according to the MSRC dataset2. In particular, we follow the formulation of Shotton et al. [3] and Moosman et al. [2]. We assume familiarity with Matlab and a good knowl-edge of C.++. On the other hand, no background on the classi cation method itself is assumed; we shall explain it along the way, using blocks of actual working C.++ cod