[keghn]

 Well in the first video your program is rip roaring fast. I the next video it look like it is going really fast at doing nothing? A bug?
What is really going on?   

[yotamarker]

it detects the objects in the picture then the objects in those detected objects.

in the second video it runs faster because I took out marking the detected outline pixels.

in the case of the leaf girl nothing was detected because of all the leaves, I think it is best to use motion detection
for her cases  because it is faster and offers detection for more special cases than specific color detection

[keghn]

 Can you change the thresh hold higher. Sorry do not remember if you are using a company's software, edge
detector or are using your own. That is based on neighboring kernel squares. Like was discussed earlier? 

[yotamarker]

[keghn]

In Raw Numpy: t-SNE: 

https://nlml.github.io/in-raw-numpy/in-raw-numpy-t-sne/

[keghn]

Object Detection vs. Object Recognition: 

http://www.learnopencv.com/selective-search-for-object-detection-cpp-python/

[yotamarker]

todays puzzle :
how do you make an A.EYE recognize groups such as :
employees , customers, dolls, bots, Japanese people, palm trees

[yotamarker]

today I learned I can't use class attributes(vars) in shared procedures 

[ivan.moony]

I it is about object oriented programming, there should be public and private modifiers. Private would be visible only inside the object, while public should be visible everywhere. There should also be other modifiers for i.e. visibility over inheritor classes.

[keghn]

 The difference between people and palm trees can be just done by there outline.

 Each out line are made up of sub features. Such as straight lines, corners, and curves.
 Which bring up a important topic of what is a straight line. It is a line that goes in between two parallel lines. And you want to use the parallel lines for data not the real line. Because there can be thousands of type of crooked lines that cans
stay withing the parallel lines. 
  Which will end up being a chain code of two points that the line goes threw.

 In video out line jump around and do crazy thing. So you need to track outlines back to the selfie for the classification:

https://selff.ee/

[yotamarker]

but the question was about groups.

[Korrelan]

how do you make an A.EYE recognize groups such as :
employees , customers, dolls, bots, Japanese people, palm trees

You use feature detection not the objects outline… pretty sure I mentioned this really early on in this thread.

Outlines are very unreliable; the object can be partially occluded or colour/ contrast boundaries can deform/ impose a false outline etc.

 

[yotamarker]

I was thinking sampling less pixels for the classifier

[keghn]

 if you mean looking at to blobs of crowds from a distance then to randomly pinpoint sampling within each blob. Then yes. 
 Using a uncle Bob classifier will not detect other people. Because it is "over fitting". It needs
more "biasing" to generalize on others.

How Support Vector Machines work / How to open a black box: 

[yotamarker]

for a refreshing and unexpected change the A.eyes next level was reached easily.
to be continued