Aashish Dhawan, Pankaj Bodani, Vishal Garg
arXiv 2025 · orig. INDIACom 2019 · 2025
Conditional Random Fields (CRFs) are a common post-processing step for sharpening the boundaries of image segmentation maps. We investigate how different CRF formulations affect segmentation clarity, evaluating them on two contrasting datasets: lower-quality satellite imagery and high-quality aerial photographs. By comparing performance across image-quality regimes, we identify which CRF variant works best in each setting and characterize the strengths and limitations of each approach.
Segmentation models often produce maps with ragged or uncertain boundaries. CRFs refine these maps by encouraging label agreement between similar, nearby pixels. We evaluate several CRF formulations — linear, grid, and fully-connected (dense) — and ask a practical question: which one to reach for given the quality of the input imagery?
Testing across satellite and aerial datasets — deliberately different in resolution and noise — surfaces clear trade-offs and shows that the best-performing CRF variant is not the same across image-quality regimes.
Sweeping the fully-connected CRF's negative-probability parameter progressively cleans up the segmentation: clutter is removed and object boundaries — the bus, the car, the tree canopy — sharpen as the setting increases from 80% to 99%.
Back to Aashish Dhawan's site