< Previous | Contents | Next >
One thing to be aware of is that aux channels tend to consume a lot of memory. A float-32 1080p image containing just RGBA uses about 32 MB of memory, but with all the aux channels enabled it consumes around 200 MB of memory.
Stereo and Optical Flow Best Practices
How you create your composition, the images you are using, and the type of shot you are working on can all have an impact on the success of the Disparity generation and Optical Flow analysis. Below, we’ll look at some of the situations to be aware of and how you can avoid some pitfalls when dealing with optical flow.
Semi-Transparent Objects
The Optical Flow and Disparity generation algorithms Fusion uses assume there is only one layer per pixel when tracking pixels from frame to frame. In particular, transparent objects and motion blur will cause problems. For example, a shot flying through the clouds with the semi-transparent clouds in the foreground and a distant landscape background will confuse the Optical Flow/Stereo algorithms, as they do not recognize overlapping objects with different motions. Usually the optical flow will end up tracking regions of one object or the other. If the transparent object and the background are near the same depth and consequently have the same disparity, then it is not a problem.
Motion Blur
Motion blur is also a serious problem for the reason explained in the previous point. The Disparity and Optical Flow algorithms are unsure whether to assign a pixel in the motion blur to the moving object or the background pixel. Because the algorithms used are global in nature, not only the vectors on the motion blur will be wrong, but it will confuse the algorithm on regions close to the motion blur.
Depth of Field
Depth of field is also another problem related to the above two problems. The problem occurs when you have a defocused foreground object over a background object that is moving (Optical Flow case) or shifts between L/R (Stereo Disparity case). The blurred edges will confuse the tracking because they can’t figure out that the edges are actually two separate objects.