my env is a jetson xavier with 5 ip cameras streaming via rtsp. A graph made by graph compser is running on the jetson with a people, car and traffic sign detectionmodel. So far it is working.
My goal is to increase the people detection by using TAO. I want to use use an amazon ec2 g3 instance, therefor is an howto description in the nvidia quick start. But here I am quite irritated in going forward.
Pleasre give me simple step by step instructions, how to use TAO.
Which data (pictures or videos) do I have to prepare for upload into TAO?
Is TAO usage guided, so that I will follow a line of bread crumbs to the final result?
What should I have in mind when generating a custom model for my jetson with the graph using five different video scenes in live mode?
@Morganh
Thank you!
Please let me ask some further questions:
I have collected several jpeg pictures from the cameras, triggerd by motion detection. On each picture are one or more objects (persons and or animals).
Is the next task to group the pictures in folders (person and animal)?
How do I group the pictures, that inherit both types of objects?
Please take care of my “simple” type of question:
Do I have to manually create a label file for each image?
For classification task, yes, it is needed to split the training dataset into different folders.
One folder for one class. You can download classification notebook for reference.