Reproduce “WHU-Infra3D: A Full-stack Multi-modal Dataset and Benchmark for 3D Roadside Infrastructure Inventory” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chong Liu, Luxuan Fu, Xuyu Feng, Zhen Dong, Bisheng Yang
WHU-Infra3D is presented as a full-stack multi-modal dataset for 3D roadside infrastructure inventory, integrating panoramic images and LiDAR with fine-grained 2D-3D association, cross-frame tracking, and attribute/status annotations. This dataset provides a benchmark for five core tasks, revealing limitations of current models in cross-city generalization and detailed attribute recognition for urban asset management.
Key insights
- The dataset uniquely supports the development of AI algorithms for comprehensive "cognitive diagnosis" of urban assets, going beyond mere visual detection to enable automated maintenance and lifecycle management.
- Current AI models exhibit substantial performance degradation when faced with cross-city domain shifts and struggle with recognizing rare, long-tailed defective statuses of infrastructure.
- Effective solutions require a combination of domain-specific adaptation (e.g., fine-tuned open-set models) and explicit integration of geometric cues or knowledge retrieval (e.g., RAG) to achieve robust localization and fine-grained attribute recognition.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.