Init Models

Here, we provide the modified VGG-16 network, which are used as initial pretrained models in our experiments.

Models

  • vgg16_128.caffemodel (link): first fully connected layer has kernel size 4×4 and 4096 filters. Used as init.caffemodel, e.g., for DeepLab basic model.
  • vgg16_20M.caffemodel (link): first fully connected layer has kernel size 3×3 and 1024 filters. Used as init.caffemodel, e.g., for DeepLab-LargeFOV.
  • vgg16_20M_coco.caffemodel (link): first fully connected layer has kernel size 3×3 and 1024 filters. The model has been pre-trained on MS-COCO dataset. Used as init.caffemodel, e.g., for DeepLab-COCO-LargeFOV.