How can fg/bg likelihood merge? about fcis HOT 10 CLOSED

Sorry for the confusion. Actually, a more accurate description would be inside/outside likelihood, which measures the likelihood of a pixel that is inside/outside of the category-specific foreground mask in a region of interest. There is an additional ``null" category for normalizing with the object categories.

from fcis.

I've got a similar question. How do you produce final class-specific mask from these inside/outside maps? Do you explicitly subtract the inside and outside values and then apply sigmoid for every cell?

You mention " For mask prediction, a softmax operation produces the per-pixel foreground probability ([0 ;1])." and "For a positive ROI, for each (inside, outside) score pair, one should be high and the other should be low, depending whether the corresponding pixel is inside or outside the object boundary", but I failed to understand how exactly this conversion is done.

Probably I am missing something very obvious.

from fcis.

Hi, basically, a softmax is applied on the 2-d inside/outside maps of the corresponding ground-truth category, to produce the instance mask.

from fcis.

Please let me rephrase the question. If I get the report well, we have a tensor of shape 2 * k * k * (C + 1), i.e. for a given category it's of shape 2 * k * k and we need to regress a single k * k one (the foreground mask), correct?

In your response I don't understand how you merge two softmax'd inside and outside maps into a single mask. Sorry for re-asking.

from fcis.

The 2-d score maps are pre softmax. They are fed into a softmax layer to produce 1-d foreground probability.

from fcis.

And then you apply softmax for every spatial location separately, right?

(I am asking this because from my understanding, the softmax layer SoftMax(X_i) = exp(X_i) / [sum_j exp(X_j)] does not change the tensor dimensions)

from fcis.

I think @daijifeng001 says that apply softmax over inside and outside map and calculate the likelihood of inside:
softmax(F_ij) = exp(F_ij) / (exp(F_ij) + exp(B_ij)),
where F_ij denotes a pixel at i,j on Foregroud(Inside) Map, and B for Backgroud(Outside) Map.

In this case, for each pixel on both maps you get a single value so 2*k*k reduced to 1*k*k

from fcis.

We have updated the description in our Arxiv paper https://arxiv.org/abs/1611.07709

And we have just released the code. We sincerely apologize for the delay. This is due to switching from our internal Caffe version to the public MXNet, which provides good support of fast multi-GPU training & inference.

Enjoy! It is worth noting that:

-FCIS provides a simple, fast and accurate framework for instance segmentation.

-Different from MNC, FCIS performs instance mask estimation and categorization jointly and simultaneously and estimates class-specific masks.

-We did not exploit the various techniques & tricks in the Mask RCNN system, like increasing RPN anchor numbers (from 12 to 15), enlarging the image (shorter side from 600 to 800 pixels), utilizing FPN features and aligned ROI pooling. These techniques & tricks should be orthogonal to our simple baseline.

from fcis.

Thanks! Sorry for asking too many questions by email!

from fcis.

@vadimkantorov Thanks a lot for your interest!

from fcis.