A Comparison of FPGA and GPU for Real-Time Phase-Based Optical Flow, Stereo, and Local Image Features

Phase-differencing in Stereo Vision - Solving the Localisation Problem

“…For example, Fröhlinghaus and Buhmann (1996) considered disparity estimation an illposed problem which requires regularisation in order to produce smooth results, and thereby lose, not to say sacrifice, good localisation. This effect is clearly visible in most if not all results which employ real images (Pauwels et al, 2012). Solari et al (2001) presented an alternative phase-differencing model (see Section 10), perhaps more in line with biological processing, but without any postprocessing or regularisation.…”

“…There are serious problems involved in obtaining reliable disparity estimates, and these can explain the poor results which have been obtained in previous work (Sanger, 1988;Jenkin and Jepson, 1988;Fröhlinghaus and Buhmann, 1996;Solari et al, 2001;Pauwels et al, 2012). And instead of one "sting" of the model we have identified three stings: (1) phases and phase differences are not localised, which may be an advantage when creating a smooth depth map but it comes at the cost of sacrificing any neat depth transitions;…”

“…It seems that all processing applied targets smooth surfaces. Even more advanced processing schemes, in which disparity estimation starts at a coarse scale and is refined at progressively finer scales, seem to target only smooth surfaces such that well-defined depth transitions at high-contrast object boundaries often produce disastrous results (Pauwels et al, 2012). Focusing on low-level processing, also a solution became obvious for tweezing the stings from the model: the only way to obtain reliable and precise disparity information is to check the local image structure, i.e., to detect lines and edges and to attribute disparity only to detected events.…”

Towards the Optimal Hardware Architecture for Computer Vision

“…With very little hardware overhead and a consumption of few tenths of a watt, outperforms standard software and DSP implementations more than x800 and x5 times respectively. However, these designs have larger development cycles, as (Pauwels et al, 2011) depicts. In this work, some low-level operations (phase-based optical flow, stereo and local image features) are compared both on FPGA and GPU.…”

Incremental learning of an optical flow model for sensorimotor anticipation in a mobile robot

“…We use a GPU implementation of phase-based optical flow [18], which provides a dense flow field in real time.…”