Masks can identify regions to ignore, but cannot recover a background that was never observed. Verify mask use separately in camera estimation and training, including polarity and resizing. Prefer unoccluded observations and validate the pipeline on a small subset.
Masking features and masking training are separate
SfM may need to avoid feature tracks on moving people. Training needs a separate decision about how masked pixels affect comparisons and losses. Painting a person black and treating the result as an ordinary photograph is not equivalent to ignoring the region.
Tools and versions differ. A file called a mask does not prove that every stage uses it. Verify behavior with a small recognizable example.
Review segmentation quality
Tools such as SAM 2 can assist image or video segmentation, but inspect hair, carried objects and changing occlusions. Shadows and mirror reflections may need separate treatment.
Preserve image-to-mask correspondence and verify size, crop, undistortion and paths. Check whether black means ignore or retain in the destination pipeline; conventions are not universal.
Ignored pixels do not supply background evidence
Look for other times or angles that reveal the background. If none exists, consider recapture, a restricted viewing range or disclosed generative completion.
Evaluate residual people, background holes and novel-view floaters, not just frontal disappearance. This article does not claim an existing one-click Visuloma service that removes every moving object.
Sources and scope
Research synthesis and a proposed diagnostic workflow. Not an unprovided Visuloma case study or benchmark.
Sources checked on 16 October 2026. Capture conditions, methods and versions affect applicability.