Semantic segmentation of large-scale point clouds is an indispensable component of outdoor scene perception, providing essential 3D semantic insights for applications in scene reconstruction, urban planning, autonomous driving, and more. However, the discriminative capability of point clouds features declines with increasing distance from the sensor, causing current methods to usually perform poorly in segmenting distant objects. To overcome this challenge and improve the differentiation between classes with similar geometric features, we propose the Cross-modal Semantic Focus Network (CSFNet). Firstly, we design a Multi-scale feature Dynamic Fusion module (MDF) to leverage multi-scale image features, thereby enriching the feature representation of point clouds with additional images color and texture information. Then, in order to extract the distinguishing features of distant and different categories of objects more efficiently, we propose a Semantic Focus Module (SFM) that employs a multi-class contrastive learning strategy to enhance feature discrimination. Finally, we introduce cross-modal knowledge distillation to augment the model’s comprehension of point clouds. Extensive experiments conducted on the SemanticKITTI and nuScenes datasets demonstrate the effectiveness of our method. Notably, our method achieves superior segmentation accuracy across multiple classes at various distances compared to current methods.

3D Visual Grounding (3D VG) has become increasingly important in applications such as robotics and human-computer interaction. However, accurately localizing target objects in complex 3D scenes based on natural language descriptions remains a significant challenge. While current methods based on DETR have shown notable improvements in the field of 3D VG, they struggle to achieve precise localization in the presence of multiple objects. To address these challenges, we propose Progressive Proposal Detection Transformer for 3D VG (PP-DETR). We introduce a progressive proposal generation strategy into the DETR-based 3D VG model which is named Progressive Category-Constrained Sampler(PCCS). After thoroughly integrating point clouds features with natural language features, our method generates object queries through a stepwise sampling of point clouds features. These queries are then decoded to produce the final bounding boxes for the target objects. PCCS reduces the interference from irrelevant proposals, enhancing the model's focus on relevant objects and thereby improving overall localization accuracy. Extensive experiments conducted on the ScanRefer dataset demonstrate that our model achieves an overall accuracy of 54.06% on the overall0.25 metric, surpassing the baseline EDA by 0.23%. In more complex multi-object scenarios, our approach yields improvements of 0.45% and 0.49% in the multi0.25 and multi0.5 metrics, respectively.

3D visual grounding is a fundamental task for human–machine interaction, aiming to localize specific objects in complex 3D point clouds based on natural language descriptions. Despite recent advancements, existing Transformer-based architectures often rely on absolute position embeddings and heuristic query initialization, which lack the capacity to capture fine-grained relative spatial dependencies and fail to effectively filter out scene clutter. In this paper, we propose SESQ, a novel framework that synergizes Spatially Aware Encoding and Semantically Guided Querying for 3D grounding. Our approach introduces two key innovations. First, we propose the Rotary Spatially Aware Encoder (RSAE), which incorporates Rotary Position Embeddings (RoPE) into the self-attention layers. By transforming 3D coordinates into a rotary representation, RSAE enables the model to inherently capture relative spatial distances and maintains geometric consistency throughout the encoding stage. Second, a Semantic Query Initialization (SQI) module is designed to initialize object queries by explicitly computing the cross-modal similarity between textual embeddings and visual point cloud features. By replacing traditional heuristic-based selection with semantic-aware alignment, SQI ensures that the decoding process originates from contextually relevant object candidates, significantly reducing the impact of task-irrelevant distractors. Extensive experiments on ScanRefer and ReferIt3D (Nr3D/Sr3D) benchmarks demonstrate the effectiveness of our framework. Compared to the baseline EDA, our method achieves a significant performance gain of 2.68% in overall Acc@0.5 on ScanRefer, a 4.9% improvement on the challenging Nr3D “Hard” subset, and a 1.1% increase in overall Acc@0.25 on Sr3D.

Accurate forest resource assessment is critical but challenging due to forest structural complexity. Unmanned aerial vehicle laser scanning provides detailed three-dimensional data, but foliage occlusion and severe class imbalance between woody and non-woody component hinder segmentation. Conventional methods often misclassify uncertain points as leaves due to imbalance, while existing approaches overemphasizing woody features risk structural bias. To address this, we propose a novel Geometric-Semantic Enhancement Network with a dual-branch architecture. The first branch employs a dynamic mask generation mechanism simulating natural forest occlusion patterns and contextual modeling of occluded regions, enhancing cross-class feature learning and mitigating imbalance. The second branch integrates a cross-scale attention feature fusion mechanism within an encoder-decoder framework, effectively preserving multi-resolution geometric features and minimizing information loss. Extensive experiments on SOUL and For-Instance datasets demonstrate the superiority of proposed network. On the SOUL dataset, it improves Specificity index by 8.3% and G-mean by 1.0% over baselines. Notably, the proposed network achieves high-quality segmentation on For-Instance without fine-tuning, highlighting its strong generalization and practical utility for diverse forests.