TL;DR
This paper replaces the Stem layer in FRNet with the proposed FD-Stem, which improves feature representation while reducing computational complexity, and introduces long-range modeling capability with limited additional parameters, enabling effective learning of both spatial and channel-wise representations.
LiDAR semantic segmentation is significant in applications such as autonomous driving and robot navigation, as it greatly improves scene perception and object detection. However, the existing methods face the challenges of achieving high segmentation accuracy while maintaining low computational cost and complexity. In this paper, we propose a new, to our knowledge, efficient and accurate semantic segmentation network for LiDAR, called Range-FDSeg. To reduce the risk of information compression and loss when projecting 3D point cloud data onto 2D range images, we design a multi-channel fusion interactive learning (FIL) module. This module effectively integrates multimodal channels, such as coordinates, depth, and reflectivity, for interactive learning. As a result, FIL module can reduce the noise interference inherent in individual channels and capture the underlying relationships between different physical quantities. To further improve the performance, we introduce a lightweight and dynamic upsampler, called Dysample-S+. It effectively resolves the inherent challenges of traditional sampling methods through its adaptive weighting mechanism, which dynamically adjusts to local geometric patterns and density variations in raw point clouds. Extensive evaluations on publicly available benchmark datasets, including SemanticKITTI, SemanticPOSS, and NuScenes, demonstrate that the proposed Range-FDSeg outperforms most existing state-of-the-art methods.
Light Detection and Ranging (LiDAR)-based semantic segmentation is significant in advanced autonomous driving systems. However, it is challenging to achieve accurate and efficient semantic segmentation because of the sparse and uneven distribution of LiDAR point cloud data. To address the above problem, we propose a LiDAR point cloud semantic segmentation network based on point-wise distance encoding and pointed-seg head, named as PDES-Net. The point-wise distance encoding mechanism introduces normalized continuous depth information as a supplementary feature into the network, enhances the perception of both near and distant points, and reduce geometric information loss. The pointed-seg head module adaptively integrates multilevel features through learnable weight coefficients, and enhances the expressive ability of point-wise prediction. The performance of the proposed PDES-Net is evaluated on the publicly available benchmarks, SemanticKITTI and nuScenes, achieving mIoU of 68.9% and 78.9%, respectively. The proposed PDES-Net enhances segmentation performance while maintaining advantages in model parameters and inference speed. Overall, the network achieves a well-balanced trade-off between accuracy and computational efficiency.
The proposed Dilated Context Attention Network (DCA-Net), which consists of a dilated local geometric encoding module, a channel attention pooling module, and a category-boundary sampling strategy, alleviates boundary confusion in point cloud segmentation with long-tail categories.
Abstract. As a key carrier of 3D spatial information, the semantic segmentation of urban LiDAR point clouds directly impacts the reliability of applications such as autonomous driving and digital twins. However, existing methods face two core bottlenecks: firstly, insufficient adaptation to scene-specific semantics, and secondly, an inference gap between LiDAR structured semantics and segmentation instructions, which makes it difficult to effectively combine the reasoning ability of large language models with LiDAR geometric semantics. To address these issues, this paper proposes a contextual cue framework of "LiDAR semantics-large language model-retrieval-enhanced generation". The framework first designs a lightweight semantic mapping module to convert the structured information inherent to LiDAR into natural language cues that can be understood by LLMs. Secondly, it constructs a LiDAR semantics-text vector library, utilizing the RAG mechanism to retrieve fine-grained knowledge of similar scenes in real-time, generating precise segmentation cues that include geometric features and contextual relationships. Finally, through a three-stage progressive training strategy, it guides LLMs to gradually learn the mapping relationship from semantic understanding to segmentation instruction generation.Ablation experiments verify the effectiveness of each module, and the inference efficiency meets the real-time processing requirements of large-scale urban data. This study provides a new technical path for high-precision, fine-grained semantic segmentation of urban LiDAR point clouds and also offers a theoretical reference for promoting the in-depth application of large language models in 3D spatial intelligence.
3D point clouds, with their compact structural representation and rich geometric information, have become fundamental data sources for visual understanding tasks in computer vision, robotics, and intelligent systems.Despite extensive progress, many existing methods still place strong emphasis on local geometric modeling while exhibiting limited ability to capture global long-range contextual dependencies. Moreover, the increasing architectural complexity of modern models often leads to high computational cost and memory consumption. In this paper, we propose Point BiLSTM, an efficient and lightweight framework for 3D point cloud classification and segmentation. The core of the proposed method is a bidirectional long short-term memory (BiLSTM)-based sequencer module, which models long-range contextual dependencies among points with linear computational complexity, enabling effective global feature learning at a low cost. Considering the unordered nature of point clouds, we further propose a Mixed Sequence Soft Cross-Entropy Loss that jointly supervises fixed-order and randomly permuted point sequences during training. This design explicitly enhances robustness to permutation ambiguity and improves training stability. Extensive experiments conducted on three widely used benchmarks—ModelNet40, ScanObjectNN, and ShapeNet Part—demonstrate that Point BiLSTM achieves highly competitive performance. In particular, the proposed method attains the fastest inference speed on both idealized and real-world datasets, outperforming current state-of-the-art methods by 30.2% and 54.2%, respectively. In addition, Point BiLSTM significantly reduces computational complexity and memory consumption, providing an effective solution for efficient point cloud learning. Our code will be available at https://github.com/wendaodao04/PointBiLSTM.