Benchmarking PointNet++ and RandLA-Net for Urban Furniture Segmentation with Consumer-Grade LiDAR: A Pilot Study

Satama-Bermeo G., Caballero-Martin D., Affou H., KURT E., Lopez-Guede J. M.

11th International Work-Conference on the Interplay Between Natural and Artificial Computation, IWINAC 2026, Canary Islands, İspanya, 26 - 29 Mayıs 2026, cilt.16575 LNCS, ss.107-117, (Tam Metin Bildiri) identifier identifier

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Cilt numarası: 16575 LNCS
  • Doi Numarası: 10.1007/978-3-032-27317-8_11
  • Basıldığı Şehir: Canary Islands
  • Basıldığı Ülke: İspanya
  • Sayfa Sayıları: ss.107-117
  • Anahtar Kelimeler: Class Imbalance, Consumer-Grade LiDAR, PointNet++, RandLA-Net, Urban Semantic Segmentation
  • Gazi Üniversitesi Adresli: Evet

Özet

This study establishes an experimental pilot for urban digitalization through manual in-situ acquisition with consumer-grade LiDAR sensors (iPhone 15 Pro Max, iPad Pro M4). PointNet++ is implemented as comparative baseline, evidencing geometric difficulties that hierarchical architectures face with portable device noise. Contrasting against RandLA-Net, it is demonstrated that local feature learning capacity is determinant for overcoming such limitations. The evaluation on 759 objects from Vitoria-Gasteiz, Spain (124:1 class imbalance), reveals RandLA-Net achieves 90.4% overall accuracy and 61.5% Intersection over Union (IoU) mean (+12.2% over baseline), with strong per-class performance on majority classes (89.5% Benches, 83.3% Bins, 92.9% Sewer Caps). Results confirm technical superiority and operational viability of this low-cost solution, identifying critical thresholds (<50 instances) requiring specialized strategies.