If one applies the inverse transformation of the motion estimation, theoretically one obtains a "still" scene. With a panning shot, of course it implies that the picture frame constantly moves in the opposite direction of the panning. So I superimposed an overall motion based on the accumulated translation between first and last frame. When I rendered the resulting video, cropping the images to the original video resolution, two interesting phenomenon occur:

Moor

The moorlands are part of the historic landscape in the part of North Germany where I grew up. After long periods of peat digging that came with a drainage of the ground, part of the original marsh is now restored, and in January 2016 I visited this nature reserve. I had a camera with me, and stunned by the beauty I decided to record a panning shot although I did not bring a tripod. I had the vague idea that I could later digitally stabilise the picture, and so my pan also included various "partial repeats" where I felt my hand became too instable.

Motion estimation is a problem of image processing studied under such terms as "image registration", for instance the stitching together of large images from a number of individual images taken possibly from different angles at different times. Sophisticated algorithms are proposed to distort an image so that it is maximally congruent with another. I went for one of the first preprocessing stages, that is the estimation of the horizontal and vertical translation that happened between two images. This can be done using a phase correlation based on spectral transformation.

…………………………………………………………………………………

The choreography of the hand is reconstructed as a moving black boundary that now frames the video. The formerly hidden change in panning speed manifests itself as a change of the video appearing with a left or right focused perspective and reduced size, vascillating with moments where the video becomes "panoramic" and filling the entire screen. This apparently simple effect is not just a forensic retrace of the observer behind the camera but it creates a very rich scenography in its own right.

The second phenomenon occurs when I suddenly move the camera to repeat a partial pan. The motion excursion is relatively high here. Two aspects produce the phenomenon of the image attaining the quality of a piece of paper or membrane that is bent in front of one's eyes, a phenomenon of a "film inside the film": The perspective distortion cannot be neglected any longer, yielding—after the phase correction—a three dimensional displacement. And also the shutter speed is now too large to produce pin sharp images, instead yielding temporarily blurry images, as if one had caught liquid in one's eye. A strange distancing of the scenery results.

input: mentasm-e8646341fd

input: mentasm-63dcf8a8

direct "morass"

spectral "morass"

phase-processed spectral "morass"

…with inputs swapped

This image shows an area of the phase matrix around zero translation (centre). While with most image pairs you have one clear peak, here the camera was moved too fast, producing a blurry image that correlates across a horizontal range. After some experimentation, using a weighted centroid produces a good approximation of the relative motion.

………………………………………………………

[hh 03/02/16]

Sound

<TODO: continue here>