heyoeyo
commented
1 year ago If you have access to the segmentation mask itself (i.e. the binary black/white image output from the model), you can get the mask area by taking the sum of all 'pixels' in the mask. The white pixels should have a value of either 1.0 or 255 (depending on the format of the image), so the sum of all the pixels will either be exactly the 'pixel area' or otherwise 255 times larger (so you'd just need to divide it by 255). You also can compare that to the total number of pixels (given by the width x height of the mask) to get the area as a percentage.
Alternatively, the code for generating that dictionary of data can be found in the segment_anything/automatic_mask_generator.py file, so you could maybe take the parts of that code to come up with the same sorts of outputs.
I see that automatic mask generator is able to return a dictionary of data about the mask such as area, bounding box, stability scores and etc.
I was wondering if the masks generated from using prompts also returns similar data information? I am interested in getting the mask pixel area for translation into an equivalent physical dimension. Is there a way to convert the output masks from the predictor to the same output format as the automatic mask generator?