DeepLab-Weak-EM-Adapt is trained on PASCAL using only weak image-level labels and the adaptive EM algorithm (please see our paper, Weakly- and Semi-Supervised Learning of a Deep Convolutional Network for Semantic Image Segmentation). We employed kernel size = 4×4 at the first fully connected layer of VGG-16 and input stride = 4, resulting in a receptive filed size of 128.
Performance
After DenseCRF, the model yields 39.0% performance on the PASCAL VOC 2012 test set.
CRF parameters: bi_w = 35 bi_xy_std = 61 bi_rgb_std = 10 pos_w = 15 pos_xy_std = 19
Pretrained models and corresponding prototxt files
Please download from here