Training FasterRCNNObjectDetector with LayerGraph error with MinBoxSizes

When running the training instruction for a Faster RCNN with a LayerGraph the 'MinBoxSizes' parameter won't let it start:

Error using vision.internal.cnn.parseInputsFasterRCNN (line 67)
'MinBoxSizes' is not supported when the input network is a LayerGraph.

Error in trainFasterRCNNObjectDetector (line 333)
    vision.internal.cnn.parseInputsFasterRCNN(...

Error in FRCNN (line 86)
detector = trainFasterRCNNObjectDetector(images, lgraph, options, ...

The documentation states that:

"You cannot use this property if you set the network to a LayerGraph object or if you resume training from a detector checkpoint."

https://it.mathworks.com/help/vision/ref/trainfasterrcnnobjectdetector.html

And my options don't have the 'MinBoxSizes' parameter set, so it goes automatically on 'auto', not allowing the training.

detector = trainFasterRCNNObjectDetector(images, lgraph, options, ...
    'NegativeOverlapRange', [0 0.3], ...
    'PositiveOverlapRange', [0.6 1], ...
    'BoxPyramidScale', 1.2);

How do I solve this and get rid of the 'MinBoxSizes' parameter? Or is there a way to convert the LayerGraph into another type of variable for the training as a workaround?

Answer 1

[Angelo Dumitriu]

I found the problem:

'BoxPyramidScale', 1.2

This must also be removed from the parameters.

Answer 2

[Tunai Marques]

That does solve the problem because no more scaling is going to be done in the anchors. However, what if you want to use the 'BoxPyramidScale' and 'NumBoxPyramidLevels' parameters? I am interested in having anchors of vastly different sizes, and I am also using lgraph as input. Did you find a solution for that?

Cheers.