Research Detection & Recognition
Dilated Convolutions for Image Classification and Object Localization
Dilated convolutions are very effective in dense prediction problems such as semantic segmentation. In this work, we propose a new ResNet based convolutional neural network model using dilated convolutions and show that this model can achieve lower error rate for image classification than ResNet with reduction of the number of the parameters of the network by 94% and that this model has high ability to localize objects despite being trained on image level labels. We evaluated this model on ImageNet which has 50 class labels randomly selected from 1000 class labels.