Evaluation of alternative blending methods for compositing artifacts

Compare mask-based copy-paste, gradient-domain Poisson blending, and learned image harmonization within the Semantically-Guided Domain Randomization pipeline to determine their effects on compositing artifacts and detection performance.

Background

S-GDR re-inserts rendered target objects onto diffusion-generated backgrounds using segmentation masks. This can create boundary discontinuities involving lighting, shadows, color tone, scale, and perspective. Although Poisson blending and learned harmonization are identified as possible alternatives, neither has been tested in the pipeline, leaving their relative effectiveness unresolved.