DeepLab-COCO-Strong is the model that has been fine-tuned on MS-COCO and then on PASCAL VOC 2012. The employed kernel size at the first fully connected layer of VGG-16 is 4×4 and input_stride = 4, resulting in a receptive file size of 128.

Performance

After DenseCRF, the model yields 70.4% performance on the PASCAL VOC 2012 test set.

CRF parameters: bi_w = 5, bi_xy_std = 67, bi_rgb_std = 3, pos_w = 3, pos_xy_std = 3.

Pretrained models and corresponding prototxt files

Please download from here