Friday, March 23, 2012

Calibration Time is Too Damn High (part 2)

Following up on Part I:

The other meeting I attended was the annual JIONC that stands for Journées Imagerie Optique Non Conventionelle, a workshop focused on Unconventional Optical Imagery. I just could not pass attending something with unconventional imaging in the title. As we all know, Mark Neifeld said it well in this slide

even conventional imaging systems are already compressive, so unconventional imagery has to have more compressive sensing built in. I was not surprised that it did but one of the presentations that I particularly liked was that of Pauline Trouve, who is involved in building a depth sensing camera using the fact that the PSF of the RGB colors are different when using a coded aperture.

In particular, she mentioned to me that the calibration of that camera, taking into the Bayer pattern, required about 9 PSFs evenly distributed over the whole CCD (with the hope that they are not drastically changing) and that a whooping 70 measurements were needed for distances between 1 to 5 meters for the determination of each PSF. Let us recall,. that if one would want to do the same for a random lens imager, we could not even count on the symmetry on the CCD and one would have to evaluate more than 9 PSFs (each of which requiring more than 70 measurements), I am really thinking that we ought to use robots to help at that stage. There is a calibration club on LinkedIn, let's start a conversation....

A related publication include: Single Image Local Blur Identification by P. Trouvé, F. Champagnat and G. Le Besnerais, J. Idier. The abstract reads:
We present a new approach for spatially varying blur identification using a single image. Within each local patch in the image, the local blur is selected between a finite set of candidate PSFs by a maximum likelihood approach. We propose to work with a Generalized Likelihood to reduce the number of parameters and we use the Generalized Singular Value Decomposition to limit the computing cost, while making proper image boundary hypotheses. The resulting method is fast and demonstrates good performance on simulated and real examples originating from applications such as motion blur identification and depth from defocus

Liked this entry ? subscribe to Nuit Blanche's feed, there's more where that came from. You can also subscribe to Nuit Blanche by Email, explore the Big Picture in Compressive Sensing or the Matrix Factorization Jungle and join the conversations on compressive sensing, advanced matrix factorization and calibration issues on Linkedin.

No comments: