Fix ROI conversion from instance segmentation mask

[ajay1685]

This commit fixes two issues:

1. Corrects the ROI orientation for the instance segmentation masks.
2. Corrects the scaling of ROI where Mask size is normalized to full Image (YOLOs) vs bounding box (Mask-RCNN) with different instance segmentation models.

I created an issue: #41

[petebankhead]

Thanks, can you provide any specific examples of models and code that shows that this works?

[ajay1685]

Sure, here is an example of Yolo11n-seg using pytorch

And, example of mask_rcnn_resnet18_v1b_coco using mxnet

I am running the current snapshot for QuPath and QuPath-extension-djl with the fix implemented.

Here is the script:

import qupath.ext.djl.*
boolean allowDownsamples = false
var artifacts = DjlZoo.listInstanceSegmentationModels();
var artifact = artifacts[0].getDefaultArtifact()
println "Model: ${artifact}"

var criteria = DjlZoo.loadModel(artifact, allowDownsamples)

var imageData = getCurrentImageData()
var detected = DjlZoo.detect(criteria, imageData)
println "Detected objects: ${detected.orElse([])}"

[ajay1685]

This test shows the issue, if I run it without the proposed fix:

[ajay1685]

While this seems to address the issue with the orientation and scaling, things to consider:

1. Iterating over each element for raster might not be efficient if the mask dimensions are large (not an issue at the moment since both types of models return fixed size mask).
2. Scaling of a fixed size mask (i.e. 14x14 for MASK-RCNN) to bounding box of a large object could benefit from interpolation before tracing.

Let me know your thoughts.

[petebankhead]

Thanks, the changes look good to me!
I haven't looked at this extension for a while and notice the docs were outdated as well, so I've just made an update to them as well qupath/qupath-docs#168

[ajay1685]

Thanks pete, I can also add an example of instance segmentation to the docs. I just finished setting up local repo and a python environment for the qupath-docs, please let me know.