Coherent Depth in Stereo Vision

Опубликовано 17 августа 2016, 3:07

In this talk, I will present my recent PhD work on coherent depth in stereo vision – both in computer vision and human vision. The first half of the talk introduces a real-time stereo matching technique that incorporates temporal evidence in real time (≥14 fps). It is based on a per-frame technique inspired by a reformulation of adaptive support weights (Yoon & Kweon 2006), which achieves a 200 times speedup compared to a standard GPU implementation. The spatio-temporal technique visibly reduces flickering and outperforms per-frame approaches in the presence of image noise. To quantitatively evaluate the depth estimation from stereo video, we created five synthetic stereo videos with ground truth disparity maps. In the second half talk, I will introduce a novel computational model for objectively assessing the visual comfort of stereoscopic 3D imagery. The model integrates research in visual perception with tools from stereo computer vision to quantify the degree of stereo coherence between both stereo half-images. The coherence scores computed by the model strongly correlate with human comfort ratings, as shown by a perceptual study. Based on these experiments, this talk further describes a taxonomy of stereo coherence issues which affect viewing comfort, and how they can be identified and localised in stereoscopic 3D images using computational tools. This talk is based on the following papers: Real-time Spatiotemporal Stereo Matching Using the Dual-Cross-Bilateral Grid Christian Richardt, Douglas Orr, Ian Davies, Antonio Criminisi and Neil A. Dodgson European Conference on Computer Vision 2010 (poster + demo) cl.cam.ac.uk/research/rainbow/... Predicting Stereoscopic Viewing Comfort Using a Coherence-Based Computational Model Christian Richardt, Lech Świrski, Ian Davies and Neil A. Dodgson Computational Aesthetics 2011, Vancouver, 5–7 August 2011 cl.cam.ac.uk/research/rainbow/...