Hi, > I suspect you want the polygons to be coincident with the faces of 3D > polyhedra.
Yes > This problem is hard for 2D image data. Obviously... > It is unsolved > in almost > all practical cases. Hmmm ... yeah, I was afraid of that ;-) > Fortunately, converting 3D images to polyhedra is relatively straight > forward and commercial packages exist to do this. Yes, that makes sense, I can think of a lot of simple ways to do this. > 3D images usually come > from a "laser range finder" This make sense too, and I have read some things about it before. I am aware that the military uses a combination of camera input with lidar (laser range finder) in many of its vehicles, for exactly this purpose. Do you happen to know if/where it is possible to purchase relatively inexpensive lidar equipment to use in a simple home-built robot? Or is this not possible yet? > Even so, your AGI > might be swamped with an awful lot of triangular facets. You will probably > want to use polyhedral, not curved, objects in your real world scenes and > clean up the 3D polyhedra to impose planarity. There is no way to restrict attention to polyhedral objects in the real world, I suppose... However, you are right that one would want software to clean up the 3D polyhedra. I suppose one could take a polyhedron whose faces consist of many tiny triangles and then approximate it via a series of coarser and coarser polyhedra containing progressively fewer and fewer (and larger and larger) faces. More or less coarse approximations could be used for different purposes. So, based on these ideas, a rough sketch of a Novamente "visual cortex" could look something like: camera + lidar ==> 3D image of world ==> image of world as excessively multifaceted polyhedra ==> coarser images of world as less multifaceted polyhedra The output of this pipeline would be images of the real world looking something like what Novamente now sees in its AGI-SIM simulated world. For most purposes this coarse level of granularity would be fine, but sometimes it would need to be able to penetrate down to the excessively-multifaceted-polyhedra or raw-3D-world-image level. Inside Novamente, separate data tables could be maintained for coarse-grained polygons, fine-grained polygons and actual 3D images. This would enable rapid processing based on coarser-grained data whenever possible. This approach to computer vision would seem the most sensible one for Novamente, as it bypasses a lot of nasty complexity related to stereo vision, 2.5-D reconstructions, and all that. Do you know of any references where someone has taken this kind of approach in a robotics context (preferably a mobile-robotics context rather than just industrial robotics where the goal is to recognize a well-shaped bikky or something...). Googling "lidar robotics" just yields a bunch of (interesting but not germane) NASA stuff... Can you think of any reasons why this would be a stupid way to supply Novamente or another AGI system with a visual cortex? FYI, I am not planning to give Novamente a visual cortex anytime soon, but I have been talking to some humanoid-robotics folks and I'd like to have a coherent story regarding what is the apparently-best-way to give Novamente a visual cortex in principle. -- Ben ------- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
