Image Segmentation Potentials: Original vs Thresholded, Potts, and Topology Models #1689538 (License: Personal Use)

The image displays four heatmaps representing pixel-wise label assignments across a 10×10 grid. The first panel shows the clean ground-truth segmentation with three vertical bands (blue, green, red). The second panel illustrates thresholded unary potentials, resulting in a noisy, checkerboard-like pattern due to local decisions without spatial regularization. The third panel applies Potts potentials, enforcing smoothness and yielding large homogeneous regions with sharp but globally consistent boundaries. The fourth panel uses 1D topology potentials, preserving structural continuity along one axis while allowing controlled boundary irregularities. These visualizations help illustrate trade-offs between local fidelity and global coherence in Markov Random Field and CRF-based segmentation.

Used in academic papers, tutorials, or documentation on graphical models for computer vision-especially when explaining energy minimization, CRFs, or structured prediction. Targets researchers and engineers implementing or evaluating segmentation algorithms.