DeepLab-DT-COCO employs a discriminatively trained domain transform to refine the segmentation results. The model has been (1) trained on trainval_aug set and (2) pretrained on MS-COCO dataset. Note the Domain Transform componenet and the EdgeNet are trained on train_aug, and thus train_iter_1000.caffemodel is provided.
Performance
After DenseCRF, the model yields 73.6% performance on PASCAL VOC 2012 test set.
Domain transform parameters: num_iter = 3 (5: during test), spatial_sigma = 130, range_sigma = 0.1
CRF parameters: bi_w = 3, bi_xy_std = 99, bi_rgb_std = 3, pos_w = 3, pos_xy_std = 3.
Note that using 5 iterations during test for domain transform brings extra 0.1% improvment over using 3 iterations.
Pretrained models and corresponding prototxt files
Please download from this link.