Welcome to Ocula from Foundry, my name’s Jon. In this tutorial, we are looking at setting up and reviewing vertical alignment in Ocula 3.0. The Solver node calculates the stereo geometry of your input footage. We are going to use that information in the O_VerticalAligner node to correct the horizontal alignment of these plates. So we saw how to set up the Solver in the previous tutorial by defining some analysis keys on the timeline. On each analysis keyframe, the Solver calculates some feature matches between the left and right view. It uses the feature matching data to calculate the stereo geometry. It also uses it to calculate the alignment for the plates, and you can preview that directly inside the Solver to QC (quality-check) your feature matches. The O_VerticalAligner node applies that alignment data to the plates and interpolates it smoothly between the keys.

For this footage, I have combined several VerticalAligner nodes. If we take a look at the first one, you will see you can set the left and right view, you can choose whether to align both views or to align one view to a hero view, you can do a local per pixel warp which we will look at later, or a global alignment which acts like a corner pin. For the global alignment, you can choose different options for how it calculates the alignments for the plates, and you can choose the filter that’s used when it warps the images. So here you will see I have got four Aligners in a row. All of the global methods concatenate, so this will only generate a single filter hit. You are free to choose the most appropriate methods to align your plates. I generally have a Simple Shift upfront to align vertically first, I then apply a Scale to the plate to take out any focal length differences, and I then have a Rotation and this will take out any roll between the cameras. You can also choose to use Camera Rotation here, and this will take out roll and pitch between cameras. Finally, I then use a Vertical Skew, which you can use to take out any keystoning with a conversion camera rig. It’s important to remember though that skew can’t be represented by a corner pin. So if you do want to export a corner pin, which we will look at later, then use the Perspective Warp method as the final step.

Ok, let’s have a look at how the final alignment works. To QC alignment, you can use an Anaglyph node to overlay the left and right nodes and check for any vertical mismatches that remain in your plates. You can also use the StereoReviewGizmo that comes with Ocula. Here it’s set to show a Difference image. I am calculating the disparity for the aligned plates, and the gizmo has a ReConverge node inside. You can pick up the converge point and move it around to check for vertical mismatches in different parts of the image. So here it looks like there’s a difference on the lights in the background. You can correct this by going back and reviewing the alignment options you chose using the VerticalAligner nodes. Alternatively, what we're going to do here is add some user-defined matches in the analysis key in O_Solver. So this locks down the alignment exactly at those user-defined feature matches. I am going to add some on the lights in the background on both sides here, taking great care to make sure they are matched exactly in the same position. We can go back to our Anaglyph node and see how well the alignment now works. The lights now match up.

It’s important to remember that the alignment is weighted to ensure the user matches are horizontal now, so if you have any change in alignment with depth, you need to place those user matches at different depths. Now we have a way to QC the alignment and correct the analysis keys in O_Solver. You can see that the ghosting effect has now disappeared in that light in the background. We now need to check the alignment in between the keys to see how well it’s interpolated. There are a lot of frames to get through here, so a trick is to look at the disparity that’s being calculated for the aligned plates. I have got an Expression node that’s set up to look at the vertical component of disparity and check it against the threshold to see if there is any significant vertical discrepancy between the left and right view. It is worth noting here that I have increased the Strength in my DisparityGenerator to make sure I do pick up any mismatches in the alignment.

I have rendered that out and I can read it back in and play it back to see if there are any bad frames that I need to check first. So there’s a small discrepancy on this frame. If I switch back, you will see here, around frame 105 and 106, that some large discrepancies are being shown. So now I actually need to go back to Solver and add a new Analysis Key to re-calculate the alignment here - interpolating isn't good enough.

Now let’s look at the Local Alignment option. This does a per-pixel update based on the vertical component and disparity. I have got a DisparityGenerator and I am calculating the vertical discrepancy between left and right views. I am correcting that locally, so there’s no vertical offset left in the plates. So here I have simulated a local distortion left in the plate, for example, if there was a distortion in the mirror for the mirror rig. This is the input footage and I have distorted this region over here. It’s exaggerated, so you can see it more easily. I am going to do a global correction as we did before. Here’s the global result. This can’t correct the local distortion in the plates, so you can still see that the lights are still displaced, and if I switch to view the local correction afterwards, you can see this gets corrected. So if I switch between the two, you can see the correction is only being applied in a local region where the plate was distorted to begin with. So the local alignment can be used to correct distortions. You can also use it to take out any high-frequency jitter in the rig without having to key every frame in O_Solver. The important thing to remember though is to correct globally first, so that you are only taking out small differences with the local alignments. If you correct large vertical discrepancies with this method, it can distort the pixel aspect in your footage.

Finally, if you have a match-moved camera for your shot, you can pull that into O_Solver before setting the analysis key. The feature matches in stereo geometry will then be based on those input cameras. You can then use the Camera Rotation method in the VerticalAligner to correct the matched-moved cameras, so the plates are aligned horizontally. Here are the aligned plates. Also, in VerticalAligner you can analyse the whole sequence and create a corner pin if you want to export the alignment outside Nuke. If you have that match-moved camera on the input, you can also create an aligned rig to use with the correct plates. You can do this with all the global alignment methods, except Vertical Skew, which you can’t represent with a camera or a corner pin. So here I have analysed the shot and created an aligned rig. If I switch to 3D, here’s the original rig and here’s the aligned rig next to it. So you can see its been rotated to match the aligned plates.

That wraps up this tutorial. We have taken a look at how to use the O_VerticalAligner node to align stereo footage horizontally based on the analysis keys in the O_Solver node. We have looked at how to QC the alignment, how to add user-defined matches to correct alignment at the keyframes in O_Solver, and you can use disparity for the aligned plates to review how well the alignment works when it’s interpolated between the keyframes. We have also looked at the Local Alignment option to correct small distortions, and you can use this to remove high-frequency jitter in our rig. And we have pulled in a match-moved rig to define the alignment of the plates. Finally, we saw how you can analyse a sequence to output an aligned camera rig for the plates or to create a corner pin to export outside Nuke.