Efficient Foreground Reconstruction via Mask-Guided Adaptive Multi-View Stereo

초록

Reconstructing a point cloud of a single static foreground object, such as a stone pagoda, from low-resolution multi-view RGB images is challenging for conventional multi-view stereo (MVS): computation is wasted on background regions, and fixed-resolution processing allocates effort uniformly rather than where it is most useful. This work presents a COLMAP-based pipeline that combines learned foreground masking with confidence-driven adaptive dense reconstruction. Foreground masks from a salient-object segmentation network are integrated directly into COLMAP's PatchMatch dense stage so that background pixels are skipped, yielding a foreground point cloud without a separate post-processing step. In addition, a confidence-driven adaptive sampling scheme allocates more depth hypotheses to low-confidence foreground regions - such as object boundaries and low-texture areas - and fewer to already-confident ones, improving the quality-cost trade-off. On an 84-view stone-pagoda dataset at one-quarter resolution, the method reduces dense reconstruction runtime by 44.6% (from 29.1 to 16.1 minutes) while mask-guided background skipping alone also lowers peak GPU memory by 11.8%. The reconstruction yields a denser foreground object with far fewer background outliers than the standard pipeline. Because no ground-truth geometry is available, we report intrinsic relative metrics. The approach avoids generative models and offers a practical route to efficient foreground point-cloud reconstruction within COLMAP. © 2026 IEEE.

키워드