I'm experimenting with generating autostereograms and trying to understand the underlying algorithm. The asymmetric approach was easy to implement, but it requires a reference strip with no subject information and causes information to accumulate from one strip to the next. It also makes it difficult to center the subject.
I then tried a symmetric, pixel-linking method. Pixels are connected according to the depth map and linked pixels receive the same color. This avoids reserving a reference region and allows the subject to be centered, but the implementation still seems direction-dependent. When processing links from left to right, later links can overwrite earlier ones; processing from right to left produces a different result. For now, I generate images in both directions and average them.
I tested random-dot and solid backgrounds. The solid pattern looked more abstract than three-dimensional, while the noisy texture worked better for visual alignment. Is there a more principled way to resolve conflicting pixel links and generate a symmetric autostereogram, rather than rendering both directions and combining the results?
1 Answer
The apparent information imbalance in the first method may not actually be a problem. No individual strip contains 3D information on its own; the depth perception comes from comparing neighboring repeated regions. The reference strip is simply the starting pattern, not a region that must remain visually meaningless.
The results also appear to use fairly large depth values. Excessive disparity makes the image difficult to fuse, while a smaller disparity is easier to view. However, reducing it too far can make depth changes look quantized, especially on a curved object such as a sphere. A sphere changes depth continuously in every direction, so it can be a demanding test case; a simpler shape with broad, flat surfaces may make algorithmic issues easier to diagnose.
That makes sense. With a smaller depth range I can fuse the image more easily, but the sphere starts looking like concentric discs because the available depth resolution is limited. I’ll try testing with simpler geometry as well.