Train RCNN Object Detector

As the example in DeepLearningRCNNObjectDetector from MATLAB, the ROI of the image is a rectangular shape.

Similarly, the requirement of the ground truth to use to train a RCNN Object Detector must be in the format of

[x,y,width,height]

My question is that any possible way to make the ROI in a trapezium shape by modify the ways of the trainer takes the ground truth?

For example, the ground truth in the format of

[x1,y1,x2,y2,x3,y3,x4,y4]

and the trainer only take the features within this ROI.