Datastores for logical files when training mask R-CNN?

19 views (last 30 days)
When making training data for a Mask RCNN, you need the following things:
4 column cell array, where column 1 is an imageDatastore of the images, 2-3 is a boxLabelDatastore with bounding boxes and what they contain, and column 4, an imageDatastore with a logical array, each row Height x Width x Number of labels.
The first 3 are no problem, but number 4 vexes me. Here it suggests using poly2mask to generate the logical arrays, and a custom read function to put it into an imageDatastore, which to me feels like the ??? step in
1. Be poor. 2. ???, 3. Profit!
I've scoured the documentation and can find nothing about how to make matlab/imageDatastores manage logical arrays as images, or how to use it like a GroundTruth and get the file locations. I think I have an idea, but it seems so bulky and annoying, I hope there's a straightforward way.
My question:
If I have a folder My_Data, that only contains My_Sample_XXX.mat files which are Height x Width x Number of labels logical arrays, how do I get this into a datastore, like in the links?

Accepted Answer

Clive Fox
Clive Fox on 21 Apr 2023
OK I think I found something which works
Don't the logical mask as a .mat file but as a binary.png
imwrite(mask_img,'mask_1.png'); where mask_img is the logical array
Then ...
mask_ds = imageDatastore('mask_1.png');
Seems to work so far.
  1 Comment
Alex
Alex on 28 Apr 2023
This indeed seems to do the trick! Thank you so much!
For anyone in my exact situation, trying to get a maskRCNN to work:
  • I produced a GT with polygons
  • I used the usercreated function MPolyToMask + ImWrite (plus a function for adding the first polygon coordinate to the end of each polygon to "close the shape") to get data and images
  • For rectangles, I used my polygon coordinates and regionprops(CC,'BoundingBox'); to get those values and saved them in a table
This can be used to build the datastores.
Onwards to new roadblocks!

Sign in to comment.

More Answers (1)

Kevin Willeford
Kevin Willeford on 22 Sep 2023
Hi everyone,
I'm stuck on this one too. I have M X N X numObjects logical arrays. I only have two object categories per image; therefore, when I try to save the arrays as a .png, it doesn't work.
So, how to convert logical arrays with multiple objects per image into a datastore? The line saying "just create a custom read function" is baffling me.

Categories

Find more on Recognition, Object Detection, and Semantic Segmentation in Help Center and File Exchange

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!