These are chat archives for beniz/deepdetect

26th
Sep 2016
dgtlmoon
@dgtlmoon
Sep 26 2016 10:07
mmm. i think this image similarity problem is too hard for me. I've retrained a network with my imageset (5 categories) which worked fine (12+ hours :) ) and tried indexing with imgsearch.shon upper and lower layers ( pool5/7x7_s1and inception_5b/output).. ive tried also euclidean binarized and angular metrics but the image search results are no better (and usually worse) than a basic SIFT/ORB search. results against the GG model and my own retrained GG model based on my images are not really any better. trying to get some basically "good" results to get some motivation here
think i'll focus on the CSV prediction and go back to some more caffe tutorials, tho i cant see where just i'm going wrong
lots of variables here.. i guess.. retrained with 7500 images on 5 categories, they're all the same 'domain' ie, just photos of tshirts on plain backgrounds, attempting to find similar designs on the tshirts
i did notice the precision in my retrain was down to about 0.18, not sure if thats related or not, sounds a little low to me
Emmanuel Benazera
@beniz
Sep 26 2016 10:25
yes, it's a difficult problem, it all depends on your pictures, etc...
dgtlmoon
@dgtlmoon
Sep 26 2016 10:31
not sure what direction to take, it's too easy to assume the NN is doing something that it's not
maybe go back to ORB and use a NN for making the binary string hamming distance or hash search faster
ORB results were fairly good, tho its a pain getting the 200 descriptors x 256bit features per image to perform well when comparing that same 200x256 across thousands of images, perhaps some kind of LSH hash parked in NN space or... i dont know
dgtlmoon
@dgtlmoon
Sep 26 2016 10:42
also in my case its more important to know the closest neighbour by similarity rather than classifying, tho I suspect the GG net should be able todo this no probs
perhaps i'll try the net reindex on a much larger image dataset
Emmanuel Benazera
@beniz
Sep 26 2016 10:56
I'm too busy to help this week... Similarity search works, you may need to pre-process your tshirt patterns in some ways.
dgtlmoon
@dgtlmoon
Sep 26 2016 10:57
yeah, maybe trimming them down will work better, cut out the noise around them, i'll let you know how i go
will retest with a 'cleaner' data set, perhaps album covers