PosE-Enhanced Point Transformer with Local Surface Features (LSF) for Wood–Leaf Separation

Forests, Journal Year: 2024, Volume and Issue: 15(12), P. 2244 - 2244

Published: Dec. 20, 2024

Wood–leaf separation from forest LiDAR point clouds is a challenging task due to the complex and irregular structures of tree canopies. Traditional machine vision deep learning methods often struggle accurately distinguish between fine branches leaves. This challenge arises primarily lack suitable features limitations existing position encodings in capturing unique intricate characteristics clouds. In this work, we propose an innovative approach that integrates Local Surface Features (LSF) Position Encoding (PosE) module within Point Transformer (PT) network address these challenges. We began by preprocessing applying technique, supplemented manual correction, create wood–leaf-separated datasets for training. Next, introduced Feature Histogram (PFH) construct LSF each input, while utilizing Fast PFH (FPFH) enhance computational efficiency. Subsequently, designed PosE PT, leveraging trigonometric dimensionality expansion Random Fourier Feature-based Transformation (RFFT) nuanced feature analysis. design significantly enhances representational richness precision Afterward, segmented branch cloud was used model skeletons automatically, leaves were incorporated complete digital twin. Our enhanced network, tested on three different types forests, achieved up 96.23% accuracy 91.51% mean intersection over union (mIoU) wood–leaf separation, outperforming original PT approximately 5%. study not only expands limits research but also demonstrates significant improvements reconstruction results, particularly twigs, which paves way more accurate resource surveys advanced twin construction.

Language: Английский

Advancing a Vision Foundation Model for Ming-Style Furniture Image Segmentation: A New Dataset and Method

Sensors, Journal Year: 2024, Volume and Issue: 25(1), P. 96 - 96

Published: Dec. 27, 2024

This paper tackles the challenge of accurately segmenting images Ming-style furniture, an important aspect China’s cultural heritage, to aid in its preservation and analysis. Existing vision foundation models, like segment anything model (SAM), struggle with complex structures Ming furniture due need for manual prompts imprecise segmentation outputs. To address these limitations, we introduce two key innovations: material attribute prompter (MAP), which automatically generates based on furniture’s properties, structure refinement module (SRM), enhances by combining high- low-level features. Additionally, present MF2K dataset, includes 2073 annotated pixel-level masks across eight materials environments. Our experiments demonstrate that proposed method significantly improves accuracy, outperforming state-of-the-art models terms mean intersection over union (mIoU). Ablation studies highlight contributions MAP SRM both performance computational efficiency. work offers a powerful automated solution intricate structures, facilitating digital in-depth analysis furniture.

Language: Английский