zmbhou / MLPL

MLPL

Code and Datasets for paper ["Multi-task Learning and Prototype Learning for Weakly Supervised Semantic Segmentation"]

1. You can download the Localization Maps from the following links:

CAM: Link SeeNet: Link OAA+: Link PLLM: Link MLLM: Link MLPL:Link

2. PR curves:

By running the code: PRcurve/drawPR.m, you can reproduce the PR curves in Figure 9.

3. Generating proxy annotations on attention maps:

By running the code GTgeneration/CRFrefinementforGT.m, you can generate the proxy annotations on MLPL.

4. Generating proxy annotations with saliency maps:

Run gen_gtwith_saliency maps.py. The saliency maps can be obtained from Link

5. Retraining the network with proxy annotations:

By running Training code in the 2nd stage/main.py, the segmentation network will be optimized by the proxy annotations.