".pbm" and "Foreground/Background segmentation:"

[debidatta]

".pbm" and "Foreground/Background segmentation:"

[debidatta]

For the paper, we had access to Ground Truth masks in the BigBIRD dataset. So we were able to train an FCN as a foreground/background segmentation model.

However in general these masks might not be available. For the repo, we wanted to give some pointers as to how one can get masks for their dataset. So that is why we mentioned GrabCut/Pixel Objectness with Bilinear Filtering, which do not require training an additional segmentation model.

Hope this clears the confusion,

[chho-work]

Thanks for your prompt reply!

You recommended using Pixel Objectness which is based in Caffe framework, if I am not mistaken. As I mentioned, I am new to machine learning and started out learning Pytorch. To learn another framework right now seems a bit challenging for me.

Questions:

1. I would like to know if the DSS-Pytorch that you mentioned in another thread will work as Pixel Objectness.
2. If not, by chance, do you know if Pixel Objectness can be installed in Google Colab, as I have no access to physical GPU nor I subscribe to cloud GPU services.

Many thanks for your support!

[debidatta]

You can stick to one framework and use DSS-Pytorch. If your class of interest exists in PASCAL dataset you can use https://pytorch.org/hub/pytorch_vision_deeplabv3_resnet101/ .

If your images have simple background with object in center you can use https://docs.opencv.org/master/d8/d83/tutorial_py_grabcut.html and use the center pixel as the seed.

[chho-work]

I appreciate your prompt and complete reply!
Have a nice day!