Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition