Multiactivation Pooling Method in Convolutional Neural Networks for Image Recognition

Figure 7

The left most three images are raw images from CIFAR-10 datasets. The first image is from class 0 (plane). The second image is from class 1 (car) and the third image is from class 7 (horse). Each raw image has 5 pairs of images on the right. The left image in each pair contains all channels’ output feature maps from its corresponding convolutional layer. The right image in each pair is the reflection map of one feature map casually selected from all channels. The upper-left corner pairs are extracted from the first convolutional layer. The upper-right corner, lower-left, lower right, and the very right are extracted from the second, third, fourth, and fifth convolutional layer, respectively. In feature maps, black corresponds to 0 pixel value.