- How Triplet loss trained system would be relate to One shot learning problem. Basically would like to know the flow of art in real time ? can you please share your thoughts ? Because for example , if we take mobile face lock detection system, first time system will capture only one image of mine and everytime for unlock NN do 1:1 comparison. So here how triplet loss playing a role because for mobile face lock we are not going to store 1K pictures of same person right. can u please share your thoughts ?
Because in the Triplet loss lecture video saying that, we must have 1k pictures for each person in the training set. If so for face lock should i collect 1k pictures of mine, it does not make sense sir. Also for turnstile system, everytime new employee comes should we collect 1k pictures of that new employee
2.What is the idea to choose hard triplet such that d(A,P) quite close to d(A,N) ?
There is a “meta” lesson to be learned here: when you ask a well formed and meaningful question, you get an answer. The mentors are unpaid volunteers here. If you form the questions in such a way that I have to expend a bunch of effort to figure out what you are saying, then I can find better uses of my time. Did I mention that we don’t get paid to do this?
I think Prof Ng was quite clear in the lectures about how to deal with the different use cases for face recognition: how to recognize a specific person out of a given training database versus how to tell if two faces that you don’t have any specific training data for are the same or not. I don’t know anything more about this than what I learned from listening to Prof Ng. I suggest you watch the lectures again. Please note that it doesn’t count as “watching the lectures” if you just play them as background music while you do other things. It requires concentrated attention to fully absorb what Prof Ng is saying.