Does Depth Improve Pothole Instance Segmentation? A Controlled Comparison of Early and Late RGB-D Fusion in YOLOv8n-seg

Pothole instance segmentation is a perception task for pavement inspection and intelligent transportation systems. Depth sensing provides geometric cues that may complement RGB appearance, but it is unclear whether simple RGB plus depth (RGB-D) fusion improves mask accuracy when a strong RGB baseline is available. We present a controlled study of early and late image-depth fusion using a fixed YOLOv8n-seg architecture on the PothRGBD dataset (993 valid of 1,000 released RGB-depth samples; 80/10/10 split). Four matched configurations are evaluated: RGB-only, depth-only, early fusion (4-channel RGB-D stem), and late fusion (decision-level RGB and depth ensemble). On the held-out test set, RGB-only achieves the highest precision (0.980), mean average precision (mAP@50, 0.975), and mAP@50-95 (0.668), while early fusion has the highest recall (0.962). Early fusion is tied on mAP@50 (0.974) but lower on mAP@50-95 (0.654). Late fusion yields fused-mask mAP@50/mAP@50-95 of 0.952/0.642 at approximately double the compute (24.22 vs. 12.14 giga floating-point operations, GFLOPs). Stem-channel L1 analysis shows that depth is used by the early-fusion model but remains secondary to RGB. These results indicate that, under this protocol, adding depth inside YOLOv8n-seg does not automatically improve segmentation over a strong RGB baseline; depth remains more clearly motivated for post-segmentation geometric measurement.