search

dongzhang89 / CONTA

Implementation for paper: Causal Intervention for Weakly-Supervised Semantic Segmentation
MIT License
185 stars 27 forks source link

In make_sem_seg_labels.py, how to apply crf postprocessing at final result pseudo mask image? #7

Closed TyroneLi closed 4 years ago

TyroneLi commented 4 years ago

No description provided.

The fully-connected CRF [27] was used to refine CAM, pseudo-mask, and segmentation mask with the default parameters given in the original code. The hard threshold parameter for CAM was set to 16 by default and changed to 4 and 24 to amplify and weaken the background activation, respectively. The radius γ that limits the search space of pairs was set to 10 when training, and reduced to 5 at inference (conservative propagation in inference). The number of random walk iterations t was fixed to 256. The hyper-parameter β in the Hadamard power of the affinity matrix was set to 10. For this description in your paper, but I cannot refer to this implementation in your repro? Could you point that out?

dongzhang89 commented 4 years ago

@TyroneLi Thanks for your queation. 1) The CRF code has been released in pseudo_mask/tools/CRF.py, you can use it to refine the corresponding masks. 2) all the training hyper-parameters for IRNet are in run_samle.py.