DeepLab-LargeFOV employs large Field-Of-View, which is found to be useful for PASCAL VOC 2012 segmentation task. We employed kernel size = 3×3 at the first fully connected layer of VGG-16 and input stride = 12, resulting in a receptive filed size of 224. The number of filters is also reduced from 4096 to 1024, enabling 3.36 times faster training speed than original DeepLab model.
After DenseCRF, the model yields 70.3% performance on the PASCAL VOC 2012 test set.
CRF parameters: bi_w = 4, bi_xy_std = 121, bi_rgb_std = 5, pos_w = 3, pos_xy_std = 3
Pretrained models and corresponding prototxt files
Please download from here