TOTMSeg: A Texture-Aware Octree-Based Transformer-Mamba Framework for Large-Scale Urban Mesh Semantic Segmentation.
Saved in:
| Title: | TOTMSeg: A Texture-Aware Octree-Based Transformer-Mamba Framework for Large-Scale Urban Mesh Semantic Segmentation. |
|---|---|
| Authors: | Zhang, Ruiming1 (AUTHOR), Ma, Mengyu1 (AUTHOR), Du, Chun1 (AUTHOR), Li, Jun1 (AUTHOR), Chen, Hao1 (AUTHOR), He, Shitian1 (AUTHOR) heshitian19@nudt.edu.cn |
| Source: | Remote Sensing. Sep2026, Vol. 18 Issue 18, p3198. 27p. |
| Subjects: | Texture analysis (Image processing), Time complexity |
| Abstract: | Highlights: What are the main findings? We propose TOTMSeg, a texture-aware octree-based Transformer-Mamba framework that achieves state-of-the-art mIoU of 83.44% on the SUM dataset and 81.67% on the seMet dataset, outperforming all compared SSM-based and non-SSM baselines. The proposed texture sampling and lightweight Texture Triangular Convolution (TTC) module significantly improve segmentation accuracy on fine-grained structures and small objects, with IoU gains of up to 24.70 percentage points for the boat category over the best SSM competitor. What are the implications of the main findings? This work demonstrates that explicitly exploiting intrinsic texture information is critical for 3D mesh semantic segmentation, overcoming the discriminative deficiency of conventional geometry-only or single-color point cloud methods. The framework maintains strict linear computational complexity and achieves competitive inference than the SOTA SSM baseline, making it practically deployable for low-latency, large-scale urban scene understanding with high computational efficiency. Semantic segmentation of 3D urban meshes is a fundamental task for intelligent urban scene understanding. However, conventional methods face severe computational efficiency bottlenecks in large-scale scenarios. Meanwhile, state space models (SSMs) with inherent linear time complexity offer an efficient solution, but existing SSM-based approaches fail to fully exploit the intrinsic textural properties of 3D meshes, resulting in degraded feature discriminability and inferior performance on fine-grained structures and small urban objects. To address these issues, we propose TOTMSeg, a texture-aware octree-based Transformer-Mamba framework for large-scale urban mesh semantic segmentation. It follows a coherent pipeline of texture feature extraction, octree-guided local feature aggregation, and global semantic feature modeling. Specifically, we design a Point-template-based Texture Sampling (PTS) strategy and integrate a lightweight Texture Triangular Convolution (TTC) module to extract fine-grained texture details, effectively alleviating the texture information deficiency in conventional point cloud-based methods. Extensive experiments on two large-scale public datasets demonstrate that TOTMSeg outperforms state-of-the-art (SOTA) baselines on main evaluation metrics, achieves notable performance gains on fine-grained structures and small-object categories, and preserves the high computational efficiency of SSM-based architectures. [ABSTRACT FROM AUTHOR] |
| Copyright of Remote Sensing is the property of MDPI and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.) | |
| Database: | Engineering Source |
|
Full text is not displayed to guests.
Login for full access.
|
|
| Abstract: | Highlights: What are the main findings? We propose TOTMSeg, a texture-aware octree-based Transformer-Mamba framework that achieves state-of-the-art mIoU of 83.44% on the SUM dataset and 81.67% on the seMet dataset, outperforming all compared SSM-based and non-SSM baselines. The proposed texture sampling and lightweight Texture Triangular Convolution (TTC) module significantly improve segmentation accuracy on fine-grained structures and small objects, with IoU gains of up to 24.70 percentage points for the boat category over the best SSM competitor. What are the implications of the main findings? This work demonstrates that explicitly exploiting intrinsic texture information is critical for 3D mesh semantic segmentation, overcoming the discriminative deficiency of conventional geometry-only or single-color point cloud methods. The framework maintains strict linear computational complexity and achieves competitive inference than the SOTA SSM baseline, making it practically deployable for low-latency, large-scale urban scene understanding with high computational efficiency. Semantic segmentation of 3D urban meshes is a fundamental task for intelligent urban scene understanding. However, conventional methods face severe computational efficiency bottlenecks in large-scale scenarios. Meanwhile, state space models (SSMs) with inherent linear time complexity offer an efficient solution, but existing SSM-based approaches fail to fully exploit the intrinsic textural properties of 3D meshes, resulting in degraded feature discriminability and inferior performance on fine-grained structures and small urban objects. To address these issues, we propose TOTMSeg, a texture-aware octree-based Transformer-Mamba framework for large-scale urban mesh semantic segmentation. It follows a coherent pipeline of texture feature extraction, octree-guided local feature aggregation, and global semantic feature modeling. Specifically, we design a Point-template-based Texture Sampling (PTS) strategy and integrate a lightweight Texture Triangular Convolution (TTC) module to extract fine-grained texture details, effectively alleviating the texture information deficiency in conventional point cloud-based methods. Extensive experiments on two large-scale public datasets demonstrate that TOTMSeg outperforms state-of-the-art (SOTA) baselines on main evaluation metrics, achieves notable performance gains on fine-grained structures and small-object categories, and preserves the high computational efficiency of SSM-based architectures. [ABSTRACT FROM AUTHOR] |
|---|---|
| ISSN: | 20724292 |
| DOI: | 10.3390/rs18183198 |
