Cannot get the accuracy as expected

I trained on simple googlenet with 224 size on a fine grained dataset of classifying dresses based on sleeve type. Google net gave an accuracy of 82 % while bcnn with two VGG-M models and size 448 gave 72 % accuracy. On top of it, bcnn is much slower than googlenet on 224. So i dont understand what is the point of using bcnn if simple and faster googlenet could give better accuracy.