Deployable 3D Semantic Segmentation for AI-Assisted Bridge Inspection: Benchmarking Modern Point-Cloud Models Under an 8 GB GPU Constraint
Year: 2026
Type: article
Abstract: AI-powered damage identification and condition assessment of civil infrastructure increasingly rely on accurate semantic understanding of three-dimensional sensing data. In bridge inspection and post-disaster assessment, component-level semantic segmentation of point clouds provides essential structural context for localizing and interpreting potential damage. However, modern 3D segmentation models are often evaluated mainly by accuracy under high-end computing environments, whereas field-oriented damage assessment may require local processing with limited computational resources. This paper studies deployable 3D semantic segmentation for real-world bridge point clouds under an explicit 8 GB dedicated-VRAM single-GPU constraint. We benchmark representative architectures on five heterogeneous bridge datasets, reporting not only mIoU and overall accuracy but also training time, peak GPU memory, and cross-dataset generalization to a post-disaster bridge dataset. The results show that deployability is not determined by model capacity, memory footprint, or single-dataset accuracy alone; instead, practical deployment for AI-assisted infrastructure inspection and damage-assessment workflows requires a balance of accuracy, runtime, memory use, and robustness to domain variation.
Source: Infrastructures
Institution: The University of Tokyo
Language: English
Subfield: Civil and Structural Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,172
OpenAlex ID: W7235464968
