July 29, 2016 12:00 PM - 1:30 PM
Title: A Multi-scale Consensus Model for Low-level Vision. Abstract: During the past ten years my research group has been using various tools, from applied mathematics and signal processing to computer graphics and computational imaging, to help build a catalogue of how local patterns of image brightness can be used by vision systems to constrain their local interpretations of shape, lighting, and materials. Supposing we have a complete catalogue of such local, low-level constraints, and supposing we can apply them at multiple scales across a visual field, how should we combine their predictions for vision? I will try to answer this question by suggesting a computational framework for incorporating diverse, noisy scene constraints at dense locations and scales. The framework can be understood as a large collection of distributed computational units, with each unit corresponding to a spatial region of the visual field (a “receptive field”) having a certain position and size. Regardless of its location and scale, each of the computational units iteratively performs the same set of simple calculations, and it iteratively trades messages with other units through sparse feed-forward and feed-back connections across scales. I will show results of applying this framework to estimate disparity from stereo images and surface normals from monocular diffuse shading. Then, I will give you some time to tell me what in this framework, if anything, resembles what is known about neural implementations for similar visual tasks.
Metcalf Research Bldg, Room 305
Dept: CLPS, Departments, Other, Other Events
Subscribe to our events calendar with our calendar feed:
- For Google Calendar, right-click on the calendar feed icon above, choose 'Copy link address', then visit your Google Calendar page and 'Add by URL.'
- For iCal/Mac Calendar, right-click on the calendar icon above, choose 'Copy link address', then open iCal and create a 'New Calendar Subscription'. Paste the copied address into the Calendar URL field, and set an Auto-refresh setting of 'Every Day'.