I don’t use Survielance station package. But I have read about the DVA3219 / DS1419dva, equipped with Nvidia GTX1050Ti just for this DVA feature. Reasonable step from them, because I can’t imagine an Atom based (and similar) NAS will provide enough power for such performance eater as DVA is.
I received my DVA3219 yesterday and got it set up.
I don't know if I am not setting it up correctly, but after I set up one of the DVA features, Intrusion detection, it'd trigger motion alerts on everything that crossed the line, but just automatically tag people/cars/animals only in the DVA results page. That seems kind of useless in some ways. I only want it to trigger motion alerts when it's detected people/cars/animals, not everything like a piece of paper crossing the line. What's the point of just tagging people/cars/animals that I need to review the DVA results but continue to receive all the false positives alerts?
this NAS is in my watched list. But it is pretty new for share a heavy experiences. Then still watching.
To be sure, deep learning is about learning, of course. In the case of such "detection" from images (the video source is composed from static images), all the detection outputs are based on hierarchically defining specific features of images . No one knows what kind of algorithms are in usage in this NAS model. But we need such pioneers like you @outie - to know, if the current Synology approach is good or not. @outie - thanks for next experiences sharing to future.
It is not easy as for Machine learning, where you need just structured data relationship evaluation. The Deep learning is really challenge or a misery with pieces of paper crossing the line.
BTW: Classification of True vs. False and Positive vs. Negative is one of powerful approach in Machine learning practice. Then Question is, if there is an input, how to tell to the algorithm, that it was TP, TN, FP, FN user's evaluation - because it is still about the learning (or not).
And we know, that immature sw or precocious issued/not tested deeply sw is part of Synology approach. But sometime they provide really good job!
The issue is not that the machine has difficulty differentiating objects. The issue is that it triggers motion alerts on all objects. It’s basically the dumb motion detection plus AI tagging the clips that is not related to “smart detection” of the motion alerts. Again I don’t know if I am just not setting it up right or synology hasn’t released a “full feature” surveillance station for the DVA units, or in worst case scenario, this is how it’s supposed to work??
did you read this doc: Deep Motion Detection in Deep Learning NVR Administration guide?
Customize sensitives: ignore small objects. You have to olay with the % amount. Seems to be very sensitive (what is better for patient people). Finally there is still opened question about the algorithm.
Smart tags (your point) is not about filtering of unwanted objects (such papers, etc). It is about better understanding of the captured events by tagging of them, e.g. it is courier.
Deep learning is about patience.
I did read them yes. I have tweaked the ignore small object parameter as well. However, my point is that if all this does is detect motion using a set size parameter (usually called a threshold value), this is not AI or smart detection. This is as dumb as my regular camera/nvr combo that doesn't require any fancy GPU power.
The AI DVA of the DVA3219 is supposed to detect people/animals/vehicles in real-time and only trigger motion alert on such objects. At least that's how I understood it from all the marketing material. However, as it currently is, it only automatically "tags" them as such after the fact for your review. Whatever that triggered the motion, you get the motion alert irregardless of what kind of object it was. Again, my dumb camera/nvr system can do this just as well.
as I told you, deep learning is more than single “%” value tune
then you have got the answer, maybe additional product targeted for 80% of users from Mass market - few setup steps for immediate action. An achievement? Frequently not enough for rest of 20%.
Btw, there is no deep dive knowledge base (in Synology web). Just 9 pages of mentioned pdf doc for administration of Deep learning, where:
- 5 pages about title, intro, ending
- rest of them are really for the immediate action (not enough for scientists)
my interest for this NAS model wasn’t about the surveillance services. If there is a chance to install Tensorflow and utilize the GPU by train of ML models, then it is better way of the usage this NAS.
@outie or anybody here, can you confirm that the "AI" Feature is just tagging the images after the recording? I was thinking about getting this NAS or doing my own roll with TensorFlow. If what you guys wrote still holds, then I will go with TensorFlow.