Mask Extent and Texture Content: A Controlled Digital-Compositing Study of Person Detectors

This preprint presents research conducted at U-Rec Inc. on how mask extent and texture content influence person detection under controlled digital compositing. The study compares synthetic camouflage, smoothed palette-quantized noise, solid fills, and optimized textures across six nested proxy person-region masks using Faster R-CNN, DETR, and YOLO26n. The primary analysis uses the same 32 development scenes across mask levels, with exploratory evaluation on an expanded 132-scene sample at one selected mask setting. Within the tested compositor and optimization budget, optimized textures do not consistently suppress detection more than synthetic camouflage. The expanded-sample analysis does not confirm the small YOLO-family optimization advantage observed at the selected setting in the initial sample. These findings are specific to the evaluated digital-compositing conditions. Mask geometry, location, and covered body parts vary with extent, and texture conditions are not fully color-distribution matched. The study does not establish a causal effect of garment coverage or demonstrate effectiveness in physical environments. Research code: https://github.com/u-rec-inc/mask-extent-texture-study