Accepted in ICRAS 2026 Object Detection • Pipe Inspection

Substrate-Driven Localization Gaps in Pipe Crack Detection

Mark Kenneth Biasbas¹ • Faustino Miguel Flores¹ • Carl Christian Gatchalian¹ • Lorin Angela Velasco¹ • Dulce Maria Yadao¹ • Ma. Madecheen Pangaliman¹ • Anthony James Bautista²

¹Department of Electronics Engineering, ²Department of Mechanical Engineering
University of Santo Tomas, Manila, Philippines

🥇 Champion • FUSERS 2026 Video Presentation 🏅 People's Choice • ECE Open Lab Exhibit 🎤 Presenter • S&T Undergraduate Research Symposium 📌 Participant • 2026 NRCP Annual Scientific Conference & 93rd GMA Poster Exhibit

📄 Paper Coming soon 💻 Code 🗂️ Dataset 📋 BibTeX

📄 Abstract

Automated visual inspection of piping systems is essential for structural integrity monitoring, yet the effect of pipe surface properties on detection accuracy remains underexplored. This paper examines whether the visual complexity of the pipe surface degrades crack detection accuracy, even under class-balanced training conditions. Three crack types were created on surfaces ranging from smooth PVC to textured tissue paper, with 400 labels per class. Two object detection models, YOLO-World XL and YOLOv8x, were each trained five times with different random seeds on an NVIDIA DGX A100. YOLO-World XL achieved an overall mAP (IoU 0.50–0.95) of 0.863, but per-class performance varied widely: 94.2 percent for cracks on smooth PVC versus only 73.9 percent on textured tissue, a gap of over 20 percentage points present in both models. An analysis across IoU thresholds revealed that the model can detect cracks on textured surfaces (97.1 percent AP at IoU 0.50) but fails to produce precise bounding boxes, dropping 23.1 percentage points at stricter thresholds. Three data augmentation strategies (CutMix, multi-scale training, and their combination) were tested but none significantly reduced the gap (all p > 0.05 after Bonferroni correction). These results indicate that surface texture, rather than data quantity or model choice, is the primary factor limiting detection accuracy on challenging pipe materials.

YOLO-World crack detection pipe inspection surface texture localization data augmentation

🗂️ Dataset

The dataset is purpose-built as a controlled texture gradient: three crack types are placed on surfaces that progress from smooth, low-texture PVC to porous, high-texture tissue paper. Because every class carries an identical number of labels (400 each), the dataset isolates substrate visual complexity as the variable under study, holding data quantity constant so that any per-class performance gap can be attributed to surface texture rather than label imbalance.

Class Composition

Class	Substrate	Surface texture	Labels	AP_50:95
Dummy crack	Smooth PVC sheet	Low	400	94.2%
PVC pipe crack	PVC pipe (real fractures)	Low / curved	400	90.8%
Paper crack	Tissue paper	High / porous	400	73.9%

AP is YOLO-World XL, mean of 5 seeds. Detection quality tracks the texture gradient even though label counts are equal.

Splits

	Train	Val	Test	Total
Images	2,292	217	108	2,617
Classes	3	3	3	3

Specifications

TaskObject detection (bounding boxes)
Annotation formatYOLO (axis-aligned boxes)
Training resolution640 × 640 (960 for multi-scale)
Annotations1,200 total (400 per class, balanced)
SourceRoboflow Universe (workspace gazxard, v1)
LicenseCC BY 4.0

Full images are hosted on Roboflow Universe (not bundled in the repository); the repo ships the data.yaml configuration and class definitions for reproducible training.

🗂️ Open dataset on Roboflow 📜 CC BY 4.0 License

🔍 Qualitative Results

Localization Gap Across IoU Thresholds

Dataset Samples

Qualitative Detections

Qualitative detection outputs — Predicted boxes are tight on smooth surfaces and visibly looser on textured tissue.

Substrate gap under each mitigation strategy — The substrate gap survives CutMix, multi-scale, and combined augmentation.

Training convergence curves — Both models train stably to comparable overall mAP.

📊 Key Results

Mean of 5 random seeds. All classes have identical training-label counts (400 per class).

0.863

mAP_50:95
YOLO-World XL

94.2%

Best class AP
smooth PVC

73.9%

Worst class AP
textured tissue

20.2 pp

Substrate gap
persists under all aug.

Substrate-Dependent AP Gap

Class	AP₅₀	AP_50:95	Loc. gap
Dummy crack (smooth)	99.5%	95.6%	3.9 pp
PVC pipe crack	98.6%	91.2%	7.4 pp
Paper crack (textured)	97.1%	74.0%	23.1 pp

YOLO-World XL, best run (seed 42), test split. Detection is reliable at IoU 0.50; the gap is a localization failure that grows at stricter thresholds. Across 5 seeds the per-class AP_50:95 means are 94.2 / 90.8 / 73.9.

Model Comparison (n = 5)

Metric	YOLO-World XL	YOLOv8x	Cohen's d
mAP_50:95	0.863	0.855	1.35
mAP₅₀	0.986	0.981	1.31
Recall	0.980	0.969	1.40
F1	0.983	0.974	1.44

Paired t-test. Cohen's d > 0.8 = large effect, but the substrate gap appears in both.

Mitigation Strategies: None Close the Gap

Strategy	mAP_50:95	Substrate gap	p (gap)
Baseline	0.863	20.2 pp	—
CutMix	0.840	21.2 pp	0.562
Multi-scale	0.839	18.1 pp	0.334
Combined	0.805	18.4 pp	0.102

None significant after Bonferroni correction (α = 0.0167). The gap is a fundamental, substrate-driven localization challenge.

📝 BibTeX

If you find this work useful in your research, please consider citing:

@inproceedings{biasbas2026substrate, title = {Substrate-Driven Localization Gaps in Pipe Crack Detection}, author = {Biasbas, Mark Kenneth and Flores, Faustino Miguel and Gatchalian, Carl Christian and Velasco, Lorin Angela and Yadao, Dulce Maria and Pangaliman, Ma. Madecheen and Bautista, Anthony James}, booktitle = {Proc. International Conference on Robotics and Automation Sciences (ICRAS)}, year = {2026} }

🙏 Acknowledgment

This research was conducted at the Department of Electronics Engineering, Faculty of Engineering, University of Santo Tomas. We gratefully acknowledge and sincerely thank the SafeGuard project of Dr. Anthony James Bautista for funding this research and making this work possible.