R-CNN Bounding Boxes Regressor

[MarcoSaku]

Hi all! First of all congratulations for the R-CNN work and his implementation.
In my thesis (computer science) i'm using my approach to detect some boxes in an image containing people. I would use the Pool5 layer of R-CNN to predict the bounding box from my boxes.
I think that I have to use the function "rcnn_predict_bbox_regressor(model, feat, ex_boxes)".
I want to use the pretained model, so I think that the model is in the mat file "bbox_regressor_final.mat" (bbox_reg.models{15} for the person category in voc_2012) and I used the function "rcnn_features(im, boxes, rcnn_model)" to obtain the feat.
In this way it doesn't work because the size of feat is different from the size of variable Beta.
Thanks all.

[MarcoSaku]

Ok I installed the last caffe with the latest Matlab Wrapper: with this I can access to middle layers so I accessed to Pool5 layer features of my boxes. How can I use these features to do the bounding regression? I used the function rcnn_predict_bbox_regressor(model, feat, ex_boxes) but I obtain very strange results and I don't think that are correct. Please help me