DeepLab-Weak-Bbox-Rect is trained on PASCAL using weak bounding box annotations. See our provided dataset in which the bounding box segmentations are used in this model. Please also see our paper, Weakly- and Semi-Supervised Learning of a Deep Convolutional Network for Semantic Image Segmentation). The model is based on DeepLab basic model (i.e., SmallFOV).
After DenseCRF, the model (trained with 10.5K weak bbox labels) yields 52.5% performance on the PASCAL VOC 2012 val set, and 54.2% on test set.
CRF parameters: bi_w = 5, bi_x_std = 70, bi_r_std = 3, pos_w = 3, pos_x_std = 3
Pretrained models and corresponding prototxt files
Please download from this link.
(1) Please change the variable, TRAIN_SET_WEAK, in run_pascal.sh to TRAIN_SET_WEAK_BBOX so that you can train the model with the provided list of bounding box annotations.
(2) The model should be initialized with “vgg16_128.caffemodel”.