Transfer Learning Between MLS Point Clouds for Vegetation Segmentation

Vegetation segmentation stands as a critical task in environmental monitoring and management. Mobile laser scanning (MLS) datasets combined with deep learning (DL) techniques have achieved remarkable progress toward precise vegetation segmentation. However, most studies have confined their evaluations to individual MLS datasets, limiting the understanding of model adaptability and generalization across diverse geographical regions. Moreover, achieving generalization across MLS datasets remains challenging due to the inherent differences in vegetation distribution and diversity, as well as scene complexity. Limited availability of labeled point clouds and constraints in computational resources further highlight the importance of research employing multiple MLS datasets. In this study, we investigate fine-tune-based transfer learning (FTL) as an approach to enhance the generalization capability of vegetation segmentation across diverse MLS datasets. With GreenSegNet as the backbone architecture, three MLS datasets, Toronto3D, Chandigarh, and Kerala, have been employed. Direct feature transfer resulted in substandard results, requiring fine-tuning on the target dataset. Notably, fine-tuning gradually improves the performance, even surpassing baselines on simpler datasets. The effectiveness of FTL is strongly influenced by the alignment of feature characteristics between the source and target datasets. Overlapping feature distributions between the source and target datasets demonstrate significant improvements in performance with minimal training data. Overall, the study underscores the potential of FTL as a viable approach for achieving generalization across MLS datasets for vegetation segmentation.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.