Good article explaining the math on how to transform a screen space position (including depth) back into a world space position. But, it does not discuss how to get the depth value. Also, getting the depth value on consoles may not be practical. Also, picking on a translucent part of an object may fail to pick that object since alpha blended geometry normally does not update the depth buffer.

But, as it says, you don't need the depthvalue. Just construct a ray instead, and use it to intersect with your geometry (like in raytracing). I belive this is the most common method, exactly because it can be difficult/slow getting the real depth.
Constructing the ray is explained in the text.

A good text, by the way, very clear and understandable if you know a bit a about the transformations beforehand.

Help a n00b understand this.

I'm guessing that what I'm describing would be the "third approach" that Mr. Hook mentions, but he doesn't go into any details about its "numerous limitations that makes it impractical for widespread use" ....

Why not render your scene flat shaded, with aliasing, without effects and frills, encoding the object reference in the ARGB? Whether it's to a back buffer or a texture, it seems like that would be a quick look up.

Anyone willing to describe the limitations here? I'm sure I'm not the only weenier who's wondering.

Yeah, that's actually what I did in the past for an isometric engine. I see no reason why using color as an index would be bad... since you're taking less of a hit than rendering 1 normal frame.