lucasb-eyer/pydensecrf

Implemented as a Cython wrapper around the original C++ library, it provides both high-level 2D convenience methods (DenseCRF2D) for standard image segmentation tasks and lower-level APIs for custom pairwise potentials and label compatibilities. Supports configurable unary potentials from network outputs or hard labels, spatial and bilateral Gaussian pairwise terms with learnable parameters, and mean-field inference via iterative optimization. Targets computer vision pipelines that need structured prediction post-processing, commonly following deep learning segmentation models.

2,013 stars. No commits in the last 6 months.