I was wondering if anyone could inform me of a good method to recognise 3d objects from a 2d interface (eg. a picture) in a world where there are finitely many objects which might be in the 2d viewpoint. So in other words, I need a good method to recognise objects in a picture that takes into account the possiblities of rotation and distance from viewpoint.

If you find a good method, let me know (quietly please) and we''ll make a lot of money together!

Currently there isn''t a ''good method'' as this is an open research topic in computer vision. Current methods incorporate a variety of techniques that include edge detection, depth perception of points of interest, lighting/shading analysis and object classification. One of the hardest tasks is identifying which parts of an image belong to a single object.

I believe there have been some recent efforts to use camera motion and parallax to help identify distinct objects, but I don''t have any specific references for that... just something I recall hearing at some time.

I suggest you scour the web and try citeseer (www.citeseer.com).

Cheers,

Timkin

Thanks Timkin - I suspected that might be the case, although obviously I was hoping it wasn''t.

Robots have to do similar things in a universe of essentially infinitely many objects, so I was hoping for some relatively efficient method to be well-known.

Are neural nets a good tool for these problems?

The techniques you may find useful have shape-from-X names:
shape from shading, shape from texture, etc. Most computer vision book will have a chapter devoted to these things.