These are chat archives for beniz/deepdetect

19th
Aug 2016
Jacky Yang
@anguoyang
Aug 19 2016 07:32
hi, @beniz , I used the imgsearch demo for medicine similar image search, "python imgsearch.py --search /path/your/image.png --search-size 1 " tested for 2 images: test.jpg and test2.jpg , the result images: testresult.jpg and test2result.jpg, the results are all wrong, which is bad than SIFT feature comparison. please see the images:
the image above is test.jpg, and the testresult.jpg is:
and the search result of test2.jpg is:
https://share.weiyun.com/24a2f93de9e277f3596ebfa1971aad5f
Jacky Yang
@anguoyang
Aug 19 2016 07:46
there are 5000 images in the pool
which i indexed all of them yesterday
Emmanuel Benazera
@beniz
Aug 19 2016 09:39
We use search in several customer applications and it works extremely well. At this point it's for you to tweak your system, we can't do much for you.
Jacky Yang
@anguoyang
Aug 19 2016 09:40
thanks , you also used google net or resnet?
Emmanuel Benazera
@beniz
Aug 19 2016 09:41
yes, resnet doesn't give good results for search, or at least not as good as you could expect IIRC
Jacky Yang
@anguoyang
Aug 19 2016 09:42
ok, maybe I need to read more on documentation, the result is also strange for me
Jacky Yang
@anguoyang
Aug 19 2016 09:48
BTW, what does IIRC mean?
Matt Zeiger
@mattzeiger
Aug 19 2016 15:12
anyone done any benchmarks on a per "/predict" POST server-side? I'm averaging around 1.3 seconds per image hitting /predict on a c3.xlarge - does GPU shave off a few ms or more?
Emmanuel Benazera
@beniz
Aug 19 2016 15:13
gpu = 50 x CPU
between 10xCPU and 50xCPU in practice.