Samenvatting

Sewer pipe defects pose a serious challenge for municipalities, as delayed detection leads to costly repairs and service disruptions. Timely inspection is therefore essential, yet current practice relies on manual video review, which is labor-intensive and subjective. Computer vision offers automation, but models are data-intensive and require fully annotated datasets that are expensive and often incomplete.

This study addresses incomplete labels and class imbalance in sewer inspection data by evaluating semi-supervised object detection on 360° images using YOLOv11. We evaluate focal loss, class weighting, and consistency-based pseudo-labeling to improve the detection of underrepresented defects such as fissures. While the model achieves over 90\% mAP on frequent and visually consistent classes such as inlets and displaced joints, rare defects remain challenging with less than 30\% mAP. Object tracking with DeepSORT and ByteTrack was applied to assess temporal consistency, but results were limited by annotation variability. Our findings show that without industry-wide harmonized protocols for bounding box definitions, even advanced methods such as semi-supervised learning cannot compensate for poor annotation quality