Option to return masks of segemented leaves

[LittleAri]

Hi! I am currently using LeafMachine to segment leaves on Herbarium sheets. I know I can see this in the resulting plots, and the bounding boxes in the output are helpful too, but I think it will be very useful to have the outlines of masks resulting from the segmentation step. I am currently extracting this information by saving Project.project_data_list as a numpy file, within "machine" in machine.py, for each image, and extracting the leaf outlines from the dictionary there, as a post-processing step. Was hoping to know a neater way of doing this! Thank you!

[Gene-Weaver]

Hello! You are correct, the current version is very messy. I have an update that I have not pushed yet that exports all of the leaf masks as a very simple txt file. The new method uses a new key point detection algorithm to orient the leaves vertically (lamina tip at the top, lamina base at the bottom) and then saves the oriented points like this:
BR_1839844133_Erythropalaceae_Heisteria_parvifolia__L__1139-6-1676-1349.txt
Line 1: the angle the contour was rotated from the original orientation

Based on location in the mask:
Line 2: top-most point, not guaranteed to be the lamina tip for an emarginate leaf apex, but would be the true tip for most leaf apex types
Line 3: bottom-most point, not guaranteed to be the lamina base

Based on ML predicted location:
Line 4: closest contour point to the predicted tip, may not be the true tip
Line 5: closest contour point to the predicted base, may not be the true base

Line 6: beginning of the contour starting at the top-most point, all subsequent points are ordered clockwise based on Line 2, but you can write a script to reorder based on Line 4 if you want.

I would use the Line 2 point to be the tip, and Line 5 point to be the base. Line 5 will be more robust to strange lamina base shapes. Line 2 will give the most consistent 'tip' value (unless the leaf has an inverted tip)

Below are visualizations, the code should be pushed this week if all goes well.

[LittleAri]

That's fantastic! And those closest contours would be highly useful! And the text file format in the example is perfect! Right now I've been extracting the coordiantes from ['polygon_closed'] from ['Segmentation_Whole{/Partial}_Leaf'] from the Project file. As these have been normalised to start from the origin, I approximately reparameterize the coordinates so I can visualise them again on the herbarium sheet using the bounding boxes in that same Project file dictionary.

In this example I've found the outlines from the Project.project_data_list dictionary and plotted them with a thin line. I've also plotted the bounding boxes. For one leaf I've then translated the contours using the bounding box information - and this can be seen with a thicker red line.

Thanks for the heads up! Will certainly keep an eye out for the next update!