1. I am using the faster r-cnn function with a pretrained AlexNet. Is it possible to speed up the detection, by using the GPU Coder on the retrained net? As it is explained in the documentation, a standard CNN can be compiled with the GPU Coder but how is it with the faster r-cnn network and the subsequent use of the detect() function
2. When detecting small objects (80x80 pixels) with the AlexNet based faster r-cnn detector (image 800x600 pixel), the function predicts in the best case boxes which are at least 105x105 pixels. Is this Minimum Object size caused by the AlexNet structure itselfe (because of large/many pooling, stride operations), or does the input layer (227x227x3) is to large and has to be reduced to the order of the objects (80x80). As the original paper of Faster R-CNN states, the image input layer should not fulfill any function in faster r-cnn. Hence the description in the MATLAB exmaple "Object Detection Using Faster R-CNN Deep Learning" does not provide a conclusive explanation why the input size is chosen similar to the minimum object size of the ground truth.