"28*28 masks", what is that?

[xuexl]

When i deploy the model in libtorch, I get a pred_mask that is a 28*28`s tensor.
pred_masks: (1,1,.,.) =
Columns 1 to 6 6.7040e-02 8.0095e-02 8.0535e-02 8.8935e-02 1.4882e-01 2.1055e-01
6.6355e-02 8.8262e-02 8.7877e-02 1.0452e-01 1.7061e-01 2.4665e-01
4.9940e-02 7.5910e-02 1.0294e-01 1.1402e-01 2.1387e-01 2.9677e-01
5.3084e-02 9.5872e-02 1.1876e-01 1.3956e-01 2.3670e-01 3.2493e-01
6.6633e-02 1.2497e-01 1.7715e-01 1.9293e-01 3.6666e-01 4.7469e-01
8.7384e-02 1.9130e-01 2.4560e-01 2.8409e-01 4.4529e-01 5.5942e-01
1.0172e-01 2.3206e-01 3.2975e-01 4.0185e-01 6.1649e-01 6.8876e-01
...

What is that ? How to use it ?
I find some information in the site "https://detectron2.readthedocs.io/en/latest/tutorials/deployment.html"
But, I think it is so not clear and so hard to understand.

[tsubauaaa]

I also want to know.
I want to display the mask image from the mask of float [] obtained by libtorch.
Please tell me how to do it. How do I convert this float [] to binary mask?

[xuexl]

You need resize the mask of 28*28 to the size of the ROI. Like this,

cv::Mat cv_mask(28,28,CV_32FC1); std::memcpy(cv_mask.data, pred_masks[i].data<float>(), 28*28*sizeof(float)); cv::resize(cv_mask, cv_mask, cv::Size(x2-x1,y2-y1));

[ppwwyyxx]

It is the mask of the object inside the given bounding box, i.e. left of this image:

detectron2 provides this function https://detectron2.readthedocs.io/en/latest/modules/layers.html#detectron2.layers.paste_masks_in_image that performs this

[tsubauaaa]

Thanks to everyone, I was able to create a mask image.

[Jain-Archit]

Hi,
Can you share the code with which you were able to get the mask image from 28x28 sized mask.