A fresh look on Vision
Event details
| Date | 10.11.2014 |
| Hour | 11:00 › 12:00 |
| Speaker | Dr Michael Herzog, EPFL SV BMI LPSY |
| Location | |
| Category | Conferences - Seminars |
Textbooks on vision explain that information processing proceeds in a hierarchical and feedforward manner. After photo-transduction in the retina, features of an object are analyzed by neurons in primary visual cortex V1 that are sensitive to lines and edges. These neurons serve as the input to neurons in higher visual areas that code for more complex features, such as shapes. For example, a neuron in V4, coding for a square, is modelled by two V1 neurons coding for vertical and two V1 neurons coding for horizontal lines, that all project to the V4 neuron. In this perspective, information processing proceeds from low level to high level feature analysis, well in line with experimental and clinical results. Such hierarchical, feedforward models are not only the basis for most vision models in neuroscience but also for engineering and computer science. However, this perspective struggles to explain why state of the art vision systems are still unable to find a pair of scissors on a cluttered desk, which is an easy exercise for humans. With various examples, I will show experimentally and conceptually that the textbook model misses an important step in the analysis of object recognition: flexible Gestalt based figure-ground segmentation. In this alternative perspective, high level processing determines low level processing as much as the other way around.
Links
Practical information
- General public
- Free
Organizer
- SV Faculty
Contact
- M. Mary / Dr Hirling