Details

Autor(en) / Beteiligte

• Flotzinger, Johannes
• Rosch, Philipp J.
• Benz, Christian
• Ahmad, Muneer
• Cankaya, Murat
• Mayer, Helmut
• Rodehorst, Volker
• Oswald, Norbert
• Braml, Thomas

Titel

dacl-challenge: Semantic Segmentation during Visual Bridge Inspections

Ist Teil von

• 2024 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), 2024, p.716-725

Ort / Verlag

IEEE

Erscheinungsjahr

2024

Link zum Volltext

Beschreibungen/Notizen

• Civil engineering structures - such as bridges - form an essential component of the transportation infrastructure. A failure of an individual structure can result in enormous damage and costs. The economic costs caused by the clo-sure of a bridge due to congestion can be many times the costs of the bridge itself and its maintenance. Thus, it is mandatory to keep these structures in a safe and operational state. In order to ensure this, they are frequently inspected. However, the current inspection process is error-prone and lengthy. Especially the damage documentation using a hand-drawn sketch causes inconsistencies in the building assessment. On the other hand, recent advancements in hardware enable the deployment of computer vision mod-els for increasing the quality, traceability, and efficiency of structural inspections. Such models are the key element of digitized structural inspections and the basis for automated damage classification, measurement and localization on a pixel-level. Current datasets available for this task suffer from limitations in both size and diversity of classes, raising concerns about their applicability in real-world con-texts and their effectiveness as benchmarks. Addressing this problem, we introduced "dacl10k" (damage classification), a diverse dataset designed for multi-label semantic segmen-tation. Comprising 9,920 images extracted from real-world bridge inspections, "dacl10k" stands out by its comprehen-sive coverage. It includes 13 damage classes and 6 crucial bridge components pivotal in assessing structures and guiding decisions on restoration, traffic restrictions, and bridge closures. To accelerate progress in baseline development, we organized the "dacl-challenge", inviting enthusiasts in damage recognition to vie for training the best performing model on the "dacl10k" dataset. The competition is at the core of the "1st Workshop on Vision-Based Structural In-spections in Civil Engineering", hosted at WACV 2024. In total, 23 participants registered for the challenge, with eight achieving a performance superior to our baseline. The best result shows a mean intersection-over-union of 51%. This paper delineates the challenge's structure, introduces the dataset utilized, presents the achieved outcomes, and outlines prospective avenues for further exploration in this domain.