On neural processing in the ventral and dorsal visual pathways using the programmable Graphics Processing Unit
We describe a system of biological inspiration that represents both pathways of the primate visual cortex. Our model is applied to multi-class object recognition and the creation of disparity maps from stereo images. All processing is done using the programmable graphics processor; we show that the...
| Main Author: | |
|---|---|
| Format: | Others |
| Language: | en |
| Published: |
University of Ottawa (Canada)
2013
|
| Subjects: | |
| Online Access: | http://hdl.handle.net/10393/27660 http://dx.doi.org/10.20381/ruor-12187 |
| Summary: | We describe a system of biological inspiration that represents both pathways of the primate visual cortex. Our model is applied to multi-class object recognition and the creation of disparity maps from stereo images. All processing is done using the programmable graphics processor; we show that the Graphics Processing Unit (GPU) is a very natural platform for modeling the highly parallel nature of the brain.
Each visual processing area in our model is closely based on the properties of the associated area within the brain. Our model covers areas V1 and V2, area V3 of the dorsal pathway and V4 of the ventral pathway of the primate visual cortex. Our model is able to programmatically tune its parameters to select the optimal cells with which to process any visual field. We define a biological feature descriptor that is appropriate for both multi-class object recognition and stereo disparity. We demonstrate that this feature descriptor is also able to match well under changes to rotation, scale and object pose.
Our model is tested on the Caltech 101 object dataset and the Middlebury stereo dataset, performing well in both cases. We show that a significant speedup is achieved by using the GPU for all neural computation. Our results strengthen the case for using both the GPU and biologically-motivated techniques in computer vision. |
|---|