ive been looking around the net for explanations on object picking in opengl. unfortunately they seem all a little too tedious (compared to what i came up with) or imprecise. some articles ive read mention that the selection buffer in opengl only offers 4bits for color. i came up with the idea to store a color index with each object i draw into the buffer. as im storing the objects i will increment a 24bit variable (or 3 ints). red would be first, then add green into the mix, and finally blue. this would give me up to 2^24 colors. i doubt ill be putting more than 16 million triangles into a scene at once, so color repetition doesnt seem to be a problem. when i go to point the cursor, i can do a real quick glReadPixels() of a 1x1 pixel to get the color that was found. using that color will give me the offset into the array of objects. i cant see any flaws in this plan. if anybody has experimented with this in the past, or has heard of flaws, could you let me know? id hate to spend a lot of time on this only to find out that its performance blows :P thanks -adam

performance isnt much of an issue. im trying to design a terrain editor. i like the idea of saving a copy of the framebuffer whenever the scene changes. if that can let me get rid of using glReadPixels(), it will definitely help speed it up. i plan to disable everything but coloring, like texturing, lighting, multisampling and shaders.

i appreciate the help!

Quote:Original post by remdul
Reading back stuff (glReadPixels) stalls the pipeline, so it may not be very fast compared to a raytest on the CPU, but if you only do the picking routine once every while when the user clicks inside the viewport and not constantly there should be performance issue. If you want something like a hover effect where the mouse cursor changes, you can keep a copy of the entire framebuffer on the application side (read back the framebuffer only if the scene changes).

how exactly would i go about doing raytesting? do you know of any places i can read up on it?

I'm personally not a huge fan of the item buffer method, for a number of reasons:

1) Along the lines of what remdul said, all but the latest graphics cards (namely the G80) work exclusive with floating-point values, meaning that there's a chance you won't get exact equality when comparing items or experience inconsistent behavior across hardware.

2) Any render-based method can potentially suffer from Z-fighting on distant objects.

3) You only get pixel-perfect selection. While this can be beneficial if there are many objects real close together or stacked on top of each other, if the object is really far away or at such an angle that only a tiny sliver is visible, it's difficult to select.

3) You'd need to either render your objects again in a second pass to the item buffer or require MRT support to do it in a single pass (which I hear can be real slow on some cards).

4) If you wanted to select multiple objects ala a selection rectangle, you'd have to iterate over each pixel in the returned item buffer and merge their values into a unique list. Also, it would be impossible to select completely hidden objects should that be something you want to add later.

The only real benefit is that it scales extremely well when you have a lot of objects in the scene. Instead of performing a ray-object intersection test for every object, you just render them and only the top-most ones are selectable. The graphics card can tear through simple flat-shaded triangles that have no special shaders or other processing like nobody's business. And I suppose you could always use depth peeling to get more objects underneath the selection.

You can probably just Google for more information on ray casting and ray-picking, or ray-sphere/ray-triangle/ray-object etc. intersection references. For selecting against a frustum (which is what you get when you back-project a selection rectangle), look up the SAT test.