Computing LM filterbank response took more than two hours

[Daniel Liu]

In question 2a), we need to computer the LM response for every picture
in "train" folder, so the output size is #pictures * X_DIM * Y_DIM *
48 = 238 x 166 x 166 x 48. It took me more than two hours to do this
in Matlab. All I did was:
for all pictures (238 of them)
for j = 1:48
response(picture_index, :, :, j) = conv2(im, F(:,:,j),
'valid');

Matlab ran out of RAM after about 150 pictures after about 1.5 hours,
so my hard drive started spinning like crazy and the remainder took
forever to finish.

Has anyone else experienced this problem or am I doing something wrong?

[DerekHoiem]

Did you initialize "response"? If not, then every new image it will
have to create a new larger matrix to store the values. That will
make a huge difference. You could also reduce the storage requirement
by converting to "single" format, though even in double format it
should only use up about 500MB.

Best,
Derek

[DerekHoiem]

Sorry, I miscalculated. It uses 2.4 GB in double format, so it's a
good idea to use single format if you are going to keep all of the
filter responses in memory.

Also, you're right that it will be slow. A couple to speed it up:
(1) Downsample your filters by a factor of 2. This makes it go from
15 sec to 2-3 sec in my tests.
(2) Try running the MR8fast code from the same web site.
(3) Use a subset of the filter bank (maybe just 24 filters)

Thanks for bringing these issues to attention.

Derek