Abstract
Techniques are described for temporally refining per-frame portrait segmentation masks in video. A raw mask is processed to form an eroded foreground confidence core and a three-zone partition including core, boundary, and background regions. Motion is obtained from the core region, for example using confidence-aware block matching that applies a motion-magnitude bias and produces a confidence derived from matching cost. Motion vectors are propagated from the core into boundary pixels using confidence- and distance-weighted aggregation. A prior refined mask is warped using the resulting motion field and fused with the raw mask using confidence-gated temporal blending with zone-specific rules, including forcing background pixels to zero and falling back to the raw mask when confidence is low. The framework can accept motion from block matching, a neural motion estimator, or codec motion metadata.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Temporal Mask Refinement for Real-Time Portrait Matting via Motion-Guided Spatial Smoothing with Confidence-Core Motion Propagation", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10611