一句话：提出 SegFaultLoc，将频谱故障定位迁移到语义分割标注组件，无需训练分割模型即可排序可疑标注错误。
数据标注故障会损害深度学习模型可靠性，而语义分割需逐像素标注，标注错误密集，现有方法多面向分类/检测且依赖含噪模型。论文提出 SegFaultLoc：将程序频谱故障定位迁移到分割标注组件，把掩码分解为连通分量，用冻结预训练编码器提取与噪声无关特征，构建语义谱（分类/漏标）和几何谱（定位/碎片化），再学习式聚合可疑度并排序。在三个基准上与七种基线相比，APFD 最高提升 9.2%，Precision@10% 提升 21.9%，nDCG@10% 提升 22.4%；在 COCO-Stuff 真实标注错误上验证，修正错误可持续提升分割性能。
Data annotation fault localization is a critical software quality assurance task in deep learning pipelines, as annotation faults can severely compromise model reliability and system dependability in safety-critical applications. Semantic segmentation requires assigning a class label to every pixel in the image, making it one of the most fault-prone annotation tasks, with faults ranging from incorrect class labels to imprecise boundaries. However, existing annotation fault localization methods target simpler annotation types such as classification and object detection, and rely on models trained on potentially noisy data, making them inadequate for dense annotations such as semantic segmentation. We adapt spectrum-based fault localization from program elements to annotation components in semantic segmentation, and propose SegFaultLoc, an annotation fault localization framework that identifies and ranks faulty annotation components by suspiciousness without segmentation model training. SegFaultLoc decomposes annotation masks into connected components and extracts noise-independent features from frozen pretrained encoders to construct dual fault spectra: a semantic spectrum for classification and missing faults, and a geometric spectrum for localization and fragmentation faults. Learning-based suspiciousness aggregation integrates these signals into unified suspiciousness scores for prioritized inspection. Systematic experiments on three benchmarks against seven data-quality baselines demonstrate that SegFaultLoc achieves up to 9.2% improvement in APFD, 21.9% in Precision@10%, and 22.4% in nDCG@10% over the best baseline. Validation on naturally occurring annotation faults in the real-world COCO-Stuff dataset further confirms that correcting detected faults consistently improves segmentation performance.