DeepLab-COCO-LargeFOV exploits (1) extra annotations from MS-COCO dataset (i.e., pretrained on COCO), and (2) large Field-Of-View (as DeepLab-LargeFOV).
After DenseCRF, the model yields 72.7% performance on the PASCAL VOC 2012 test set.
CRF parameters: bi_w = 5, bi_xy_std = 69, bi_rgb_std = 5, pos_w = 3, pos_xy_std = 3
Pretrained models and corresponding prototxt files
Please download from here.