Inconsistencies with displacement within a shot

In this blog, I extend the ‘cinemetrics’ with depth data and discuss the inconsistencies with the depth maps I have generated so far.

My research will use the ‘cinemetrics’ used to perform statistical analysis of 2D films. My aim is to extend this discipline with data that is unique to 3D films: depth. As detailed in my first blog, I am using the Depth Map Automatic Generator (DMAG) by Ugo Capeto to create data from 3D film stills.

As detailed in my previous two blogs, I have collected a 48 stills (in 3D) – one from each shot within a scene. These have been catalogued and, against each, the Average Shot Length (ASL) and Shot Scale recorded. In addition, I have run each image pair through DMAG and embellished the data with the min and max displacement reported by DMAG. A copy of this catalogue may be found here.

When I came to analyse the displacements, I find that there are inconsistencies between two very similar shots. For example, here are two mid shots of Hugo which are the second and fourth shots in the scene which I am analysing:

Shot 2.

Shot 4.

Between these two shots, the min displacement (white) varies from -22.4 to -20.3; the max displacement (black) varies from 30.6 to 34.6. The first cause for this inconsistency may be the quality of the source material i.e. the image pairs. Here are the depth maps for those same two shots:

Depth map for Shot 2.

Depth map for Shot 4.

We can see that DMAG has performed quite well in isolating Hugo in the depth map. There are a couple of minor issues: an artefact produced by the lamp and, in Shot 4, the picture on the wall on the right hand side of the image. But why is Hugo nearer (darker) in Shot 4? On initial inspection, it looks as though the two shots were filmed as a single shot (with Shot 3 intercut). If this was originally a single shot, I would have expected the depth bracket to remain consistent between Shot 2 and Shot 4.

I need to investigate whether the change in depth is genuine, perhaps as a result of the slight pan to the right between the two shots as the camera follows Hugo's movement into the room. My confidence with the data will have a significant bearing on how I am to use in my analysis. It may be that I establish an acceptable margin of error.

When I came to look at the image pairs close up, I notice that they haven't been cropped perfectly. In my next blog, I will return to the themes of confidence and consistency in the depth data once this issue has been resolved.