3d reconstruction from axis-aligned photos

Question

Having poked through the other threads on this topic, I believe there may be a reasonably simple answer to my question. I'm given a collection of six axis-aligned photographs of an object, such as this:

http://www.flickr.com/photos/nnenn/4479290754/in/photostream

Are there any algorithms that would reconstruct a 3d approximation of this object? One thing that I tried with marginal success was to create six 3-dimensional "prisms" (one for each picture) that are infinite in the x-, y-, or z-direction, and take the intersection of these. This works OK, but it's a bit sensitive to the inputs and I'd like to know if there's something more robust that's well-known. I should add that I expect that some manual intervention is clearly going to be required, but I'd like to know how to minimize the amount of such intervention necessary.

One subroutine that might be useful is something like the following: suppose I were to take each pixel in the six photos and identify it to a component, so e.g. all pixels belonging to the cockpit were labelled "1", all pixels belonging to the leftmost missile were labelled "2", and so forth. Would this help in the reconstruction?

Thanks!

Ragan

(by the way, this was totally inspired by this recent blog post: link ) – Ragan

If you don't consider some kind of similarity recognition within the objects, I don't think you'll get much further than what you have described. Create the outline, make an infinite extruded volume and determine the intersections between the 6. However, perhaps you can create some 3d point clouds by recognizing similarities in the images and considering epipolar geometry? MVG might be inspiring. Just a ramble on a lack of coffee, so a comment rather than an answer. ;) – Herra

Good point, and I'll check out that book. Thanks! Identifying similarities among the six pictures is something I would like to leverage, and I suspect a lot of this will be manual, but I don't know anything about how such relationships are typically represented, so any suggestions there are most welcome. – Ragan

Answer 1

This is almost a textbook example for a "space carving" or volume intersection solution. Here is a reasonable introduction to the paper.

Greatuncle