Recently, we introduced a state-of-the-art object detection approach referred to as Multi-Expert R-CNN (ME R-CNN) that featured multiple expert classifiers, each being responsible for recognizing objects with distinctive geometrical features. The ME R-CNN architecture consists of multiple components: a shared convolutional network, Multi-Expert classifiers (ME), and Expert Assignment Network (EAN). Both ME and EAN take as a common input the output from the convolutional network and also use each other's output during training. Thus, it is quite challenging to properly train all the components simultaneously to globally optimize the network parameters. The main innovation of the proposed work is to optimize the entire architecture by using a novel training strategy in which manually associated 'RoI-to-expert' mapping is used instead of using the direct output of ME for training EAN. Our experiments show that the proposed training strategy speeds up training time at least 4.2x while maintaining comparable object detection accuracy.
|