Hi!
For those of you interested in nupic.vision, this past December I created
several simulated images based on the mnist database.

It is a subset (10 categories of 10.000 training images and 1000 testing
images) of the original 28x28 MNIST images went through the ISET [1]
simulation pipeline. The simulation consists of images displayed in a 8mm
LED display with different luminance (Light variable in the tar file name)
and distance in feet (DistFt variable in the tar file name) values, and
captured in a camera sensor before the image processing pipeline (f/# = 4,
noise flag = 2, exposure  = 15ms, pixel size = 1.25e-06, sensor and image
size = 64x64).

http://scarlet.stanford.edu/validation/SCIEN/ISETBIO/ISET_MNIST/

Although the original idea was to use the retinal images for the
simulations and nupic to classify them, I didn't have enough time for the
Hackathon, so at the end I used the sensor simulated images and Tensorflow
to do the training/testing (the main conclusion was that the system trained
with the 'real life' simulated data was better at classifying any data than
the system trained only with the original data (for those of you that are
interested I can send the poster presented at the Stanford SCIEN [2]
meeting the 4th of Dec. as well).

Best,
Gari


[1] https://github.com/isetbio/isetbio
[2] https://scien.stanford.edu/index.php/scien-affiliates-meeting/

Reply via email to