Authors: Changxu Cheng, Wuheng Xu, Qiuhui Huang and Hao Wang at Huazhong University of Science and Technology

Description: We make multi-scale predictions by exploiting GAP and GMP at different feature maps with a 10-layer convolutional network. Specifically, feature maps of 5 scales are used to have 5 sub-branches. In the training stage, all 10 predictions (5x2) are utilized for softmax loss with different weights. Grouping resizing strategy and data augmentation are used on training images. While in the test phase, only the last 2 branches are used to have the final prediction.
The final version