Detectron: Recognize same object in a video

[ciolo]

Hi everyone,
I'm working with Detectron for a university project. All things in documentation that I followed works fine.
My issue is to detect an object in a video and recognize if the object at frame x is the same at frame x+1.
Since this is my first time with Detectron, I don't know how to implement it.
Any ideas?

System information

• Operating system: Ubuntu 14.04
• CUDA version: 8.0
• cuDNN version: 7.1.1
• NVIDIA driver version: 367.48
• GPU models: 2 GeForce GTX 980

[kampelmuehler]

This is a tracking problem, which is not usually covered in a detection framework.
You can use a tracker on RGB frames (find implementations in OpenCV), or use correlations on feature maps.
Either way you are going to have to get quite some implementation work done here as far as I can tell.

[ciolo]

@kampelmuehler Thank u for reply. Coding is not a problem, but as I said, since is the first time with Detectron I don't know where to extract "useful things" for my work.

[xmyqsh]

https://arxiv.org/pdf/1710.03958.pdf
@ciolo
There are many branches in tracking
Maybe what you need is this

[kampelmuehler]

@ciolo I also strongly suggest you come up with a theoretical plan on what you wanna achieve and how you wanna go about it before starting to think about implementations.
Read about detection and tracking. https://arxiv.org/pdf/1710.03958.pdf is definitely a good place to start.

[ciolo]

I'll read this and then I'll think how to implement my idea. If I'll encounter some problems I'll write here.
Thank u all for now!

[kampelmuehler]

@ciolo you can also find code here: https://github.com/feichtenhofer/Detect-Track