Scribble-Supervised Semantic Segmentation by Uncertainty Reduction on Neural Representation and Self-Supervision on Neural Eigenspace

Zhiyi Pan, Peng Jiang*, Yunhai Wang, Changhe Tu, Anthony G. Cohn

The paper has been accepted by ICCV2021.

dataset

scribble_shrink and scribble_drop are available at here. The scribble_sup dataset can be downloaded on jifengdai.org/downloads/scribble_sup/.

environment

pip install -r requirements.txt

baseline

Please modify the dataset file path in train_seg_baseline.sh and run:

sh train_seg_baseline.sh

first-stage training with Uncertainty Reduction on Neural Representation

Please modify the dataset file path in train_seg_UR.sh and run:

sh train_seg_UR.sh

the model will be saved in ./runs/

second-stage training to refine the model with soft self supervision loss

Please modify the model(obtained by first-stage training) file path in train_seg_SS.sh and run:

sh train_seg_SS.sh

evaluate

Please modify the model(obtained by second-stage training) file path and save path in evaluate.sh and run:

sh evaluate.sh

All the computations are carried out on NVIDIA TITAN RTX GPUs.