search

amazon-science / siam-mot

SiamMOT: Siamese Multi-Object Tracking
Apache License 2.0
478 stars 61 forks source link

maskrcnn-benchmark vs. Detectron2 #12

Closed SkalskiP closed 3 years ago

SkalskiP commented 3 years ago

I have a question rather than an issue. Why did you choose to use maskrcnn-benchmark, which is no longer actively supported? Do you think SiamMOT could use Detectron2 instead? If so, could you give me some tips on how to do this?

bingshuai2019 commented 3 years ago

There is no particular technique reason that we prefer maskrcnn-benchmark over Detectron2. The codes can be easily refactored to be within Detectron2 infrastructure.

If you've already dealt with Detectron2 before, you just need to follow the standard training scripts in Detectron2 to do the refactor of model training.

One thing to be aware is that the __get_item__ function in the dataset class needs to be refactored accordingly (e.g. returned value and the data structure of the bounding box in an image, etc.)

I'm happy to chat more if you have more specific questions for code refactor.

SkalskiP commented 3 years ago

Hi @bingshuai2019 sounds like a pretty interesting free-time project. I see that you published your code under Apache-2.0 License. In fact, I would be willing to build an alternative version of the SiamMot just based on the Detectron2. I'd also be happy to talk about how to do it, I'll probably need some time to read through the paper and the published code. But then if you could find the time I'd be happy to talk. :)

fanq15 commented 2 years ago

@SkalskiP Hi, did you build the detectron2 version of SiamMot?