Hi @epnev. few more questions! 1) what does the pixel intensities of the spatial filters correspond to exactly 2) why does the register_roi algorithm binarize masks? have you considered running intersection/union metric on the spatial masks after some normalizing 3) if i'm updating the spatial filter using a seed, but there's no activity at the seed nor the surrounding area, what would happen to the updated spatial filter?
@pwang724 1) It's relative intensity (or expression level) over the extent of this spatial component. For example a donut shaped component will have lower values inside the nucleus. The absolute units do not matter here. What matters is the product of each spatial and each temporal component that fits the fluorescence data. 2) The intersection over union metric requires binary masks. A similar metric to operate on non-binarized values would be for example to consider correlation values between spatial filters. I wouldn't expect big differences over the two but haven't tested this. 3) I would expect that the component will try to overfit noise but since they are penalized for sparsity it will tend to be shrunk towards zero.