Faster RCNN Roi proposals access from the model for ouput bboxes.

[abinav719]

Hi guys, I am trying to augment the current faster rcnn for openset detection. I wanted to access ROI proposals for a trained objector during inference time to use the ROI (7X7X256) tensor for each proposal (ROI feature extraction) and post process to detect whether it is known class or unknown class. Is there a way to access this tensor and take it out for inference just like we have scores and bbox. One way is appending this to class scores and try to parse through multiple functions and making minor changes in these muliple functions. But its not a clean idea. I am all ears for some input !