These are chat archives for beniz/deepdetect

25th
Nov 2016
Jacky Yang
@anguoyang
Nov 25 2016 12:44
@beniz , long time no see! wonderful update on deepdetect, thanks for your great contribution.
Emmanuel Benazera
@beniz
Nov 25 2016 13:50
np, enjoy ^^
dgtlmoon
@dgtlmoon
Nov 25 2016 13:50
hmm, interesting, ever used this for hamming distance comparison? (k2NN)
Jacky Yang
@anguoyang
Nov 25 2016 13:57
Have anyone tested dede on NVidia Jetson TX1?
Emmanuel Benazera
@beniz
Nov 25 2016 13:57
not sire what this is, have a link ?
Emmanuel Benazera
@beniz
Nov 25 2016 13:59
a link for k2NN @dgtlmoon
Jacky Yang
@anguoyang
Nov 25 2016 13:59
ok:)
Emmanuel Benazera
@beniz
Nov 25 2016 14:00
@anguoyang yes dd runs on tk1 and tx1. We will provide ready to use kits in a near future but they won't be free. Look I the closed issues, there s on Jetson build.
Jacky Yang
@anguoyang
Nov 25 2016 14:05
@beniz Ok, thank you, it is reasonable. What does these kits include, I suppose that it includes hardware/software develop kit?
dgtlmoon
@dgtlmoon
Nov 25 2016 14:07
@beniz a link? :) ps: long time no see too, hope all is well
Emmanuel Benazera
@beniz
Nov 25 2016 14:07
what is k2nn ?
all is very well, thank you!
dgtlmoon
@dgtlmoon
Nov 25 2016 14:08
@beniz https://github.com/kevinlin311tw/caffe-cvprw15 seems to require matlab.. would be quite interesting to see an abstracted deepdetect version
@beniz nearest neighbour, something like what annoy providers, but for binary strings/descriptors
hamming distance, can be used by image search and DNA matching
Emmanuel Benazera
@beniz
Nov 25 2016 14:09
@anguoyang building on ARM takes a long time, or cross building needs to be setup. We'll provide previous, including basic models for image classification and objects detection. There will be on version with face detection.
dgtlmoon
@dgtlmoon
Nov 25 2016 14:10
i guess one could train deepdetect on an existing set of data matched by existing known (slow) hamming distance matchers
would be quite easy to generate millions of rows of training data
Emmanuel Benazera
@beniz
Nov 25 2016 14:17
ah, hamming distance is simple yes, levenshtein distance on strings is a good one too. Not sure what exact problem you are solving, image similarity search ?
dgtlmoon
@dgtlmoon
Nov 25 2016 14:18
@beniz yes correct, image similarity search based on ORB/FAST/etc type descriptors
ie, which other groups of (200 features@256bits each) match this current set of (200 features @256bits each)
Emmanuel Benazera
@beniz
Nov 25 2016 14:21
that is applied to t-shirts or clothes? not sure I remember correctly...
dgtlmoon
@dgtlmoon
Nov 25 2016 14:25
yeah just searching for similar/same designs on tshirts
ive got a huge training set (100,000+ images)
and the images on the tshirts are categorised into about 5,000 categories
ive had fairly good results with ORB type matching, tho i feel it could be improved :D
Jacky Yang
@anguoyang
Nov 25 2016 14:45
@beniz the model file on jetson is somewhat compact and differed from pc version? I suppose the source code maybe same.
Emmanuel Benazera
@beniz
Nov 25 2016 14:52
@dgtlmoon I'm in a train right now but remind me that there's a service for clothes detection in images we run and that could be useful to solve the framing of you t-shirts in images iirc. We could grant you access to it if there are chances it'd be useful.
@anguoyang this is part of the job we're doing yes, providing a range of models and measures so that there can be expectations regarding a variety of usages
There's also official AWS AMIs but we re still waiting for Amazon final approval for them to be available publicly