Li Yuan

Three-dimensional (3D) referring segmentation of auricular regions is challenging due to small targets, text–geometry misalignment, and ambiguous anatomical boundaries. We propose MLBI_AurRefNet, a text-guided network for fine-grained 3D auricular point-cloud referring segmentation. A DBSCAN-based region-aware positional prior facilitates early text–geometry alignment, while a mask cross-attention text query decoder progressively enhances multimodal interaction and target localization. In addition, contrastive region regularization and point-wise contrastive estimation improve feature discrimination across weak-boundary regions. Experiments on our self-constructed Ear Referring Segmentation Dataset (ERSD) demonstrate that MLBI_AurRefNet outperforms representative methods, including TGNN, PCMTRefer, and 3D-STMN, achieving an mIoU of 50.60%. These results demonstrate the effectiveness of MLBI_AurRefNet for natural language-guided fine-grained 3D auricular segmentation and its potential for intelligent auricular healthcare applications.

Indoor 3D referring segmentation aims to identify and segment the target object in a point cloud according to a natural language expression. Although recent advances in multimodal feature fusion have markedly improved this task, existing methods do not jointly model the structural regularities commonly observed in indoor objects, such as approximate symmetry and repetitive local patterns, in conjunction with the directional constraints conveyed by referring expressions. In this work, we formulate symmetry-aware representation as the extraction of direction-consistent structural responses from both forward and backward traversals of the same spatially ordered sequence while preserving direction-sensitive variations introduced by occlusion, point cloud incompleteness, and cluttered scene layouts. Based on this formulation, we propose PCMTRefer, a symmetry-aware text-guided Point Mamba framework for indoor 3D referring segmentation. The input point cloud is first partitioned using an octree and arranged into a spatially coherent sequence via Z-order (Morton) ordering. A bidirectional state-space encoder then aggregates complementary context from both traversal directions, yielding richer representations of regular boundaries, repetitive structures, and approximately bilateral object geometries. In parallel, an asymmetric text-to-point guidance module injects semantic cues—object categories, attributes, and spatial relationships—into point-wise features, while a Background-Relaxation Token offers an auxiliary matching channel for non-target background regions. A Gumbel-Softmax-based semantic primitive learning module further extracts discriminative cues from referring expressions and integrates language semantics with point-level geometric features through a multi-scale decoder. Experimental results on the ScanRefer benchmark show that PCMTRefer achieves an Overall Acc@0.25 of 58.56%, an Overall Acc@0.5 of 54.19%, and an mIoU of 49.97%. Multi-seed validation further confirms the statistical stability of these results, with standard deviations of less than 0.2% across three independent runs.

This paper addresses the core challenges in the 3D point cloud segmentation of the human auricle region, including the low distinguishability of weak-boundary features, the lack of anatomical priors in existing methods, the tendency of boundary features to be overwhelmed by background noise, and the weakening of supervisory signals. We propose GeoPriors-3DEarSeg, a precise geometric-priors-guided 3D point cloud segmentation network for auricle regions. The network incorporates complementary geometric features from three dimensions, normal vector orientation, local Gaussian curvature, and shape diameter function, to characterize the intrinsic geometric differences at the auricle boundary. A geometric-priors-guided QKV gated attention mechanism is designed to selectively enhance the expression of weak-boundary features. Additionally, we introduce a boundary-aware loss function, NVBLoss, which does not rely on extra annotations to strengthen the supervision of boundary features. The experimental results on a human ear dataset demonstrate that our method achieves a mean intersection over union (mIoU) of 0.9857, outperforms the comparison methods, and precisely segments of weak boundaries in the auricle region. This work provides a technical basis for accurate auricle-region point cloud segmentation and offesr methodological insights for future studies on the weak-boundary segmentation of other medical anatomical structures.

In the design of ear-worn devices, segmentation of the eight main ear regions and the small subregions within the concha is crucial for structural rationality and wearing comfort. Due to the imbalance of point distribution among different region classes, conventional point cloud segmentation algorithms often suffer from misclassification. To address this, a class-balanced attention module and a dynamic class-aware loss function are introduced based on the Point Transformer v3 segmentation model. These two category-aware strategies enhance feature representation of small-region classes and improve overall training performance. Furthermore, a centroid-guided post-processing method is proposed to enhance the precision of key small-region segmentation within the concha. By combining centroid computation and nearest neighbor search strategies, accurate recognition of small-region classes is ensured. Experiments on a dataset of 500 human ear point clouds demonstrate that the proposed method outperforms classic PointNet++ and Point Transformer models in both large- and small-region segmentation tasks, achieving clearer boundary delineation between adjacent regions and significantly higher accuracy in localization and segmentation of small regions inside the concha.

Forensic odontology research has shown that teeth are the hardest structures in the human body and can serve as reliable indicators for individual identification. In this study, we propose a Point Transformer based algorithm called MultiDentFormer, which leverages local-to-global feature fusion to achieve 3D dental biometric recognition. The global features refer to the high-level semantic representations of the full dental arch, capturing the overall structure and arrangement of the teeth, while the local features focus on the detailed geometry of individual teeth. To address the limited capability of the original Point Transformer in capturing local information at the early stages of the network, this study introduces a single-point local feature aggregation module. This module aggregates local geometric information and low-level feature representations into each local center point, thereby enhancing the expressiveness of the extracted features. Additionally, to address the issue of detail information loss during the network downsampling process, which weakens feature representation capability, this paper introduces a multi-scale feature fusion strategy. This strategy effectively integrates the rich detail features from shallow layers with the high-level semantic features extracted from deeper layers, compensating for the information lost during downsampling. Experimental evaluations on a IOS dataset comprising 104 individuals and a CBCT dataset with 51 individuals demonstrate that MultiDentFormer achieves superior recognition performance compared to common algorithms in 3D point cloud classification. The local feature aggregation strategy and multi-scale feature fusion strategy enhance the discriminative ability and classification performance of the 3D point cloud feature representations.

Remote sensing images are increasingly used in applications, such as environmental monitoring and urban planning. However, the complexity of these images poses challenges for accurate object segmentation, particularly when guided by referring expressions. In this letter, we propose RSTG-segment anything model (SAM), a novel two-stage framework for referring remote sensing image segmentation (RRSIS), which builds upon the SAM. In the first stage, a multimodal encoder and a transformer-based prompt generator produce a sparse prompt that provides coarse object localization based on fused vision-language features. In the second stage, SAM utilizes this prompt to conduct fine-grained segmentation. We also introduce a lightweight text-guided adapter module for the efficient fine-tuning of SAM and a dual-attention fusion (DAF) module to enhance multimodal alignment. Experiments on the RRSIS-D dataset demonstrate that RSTG-SAM achieves the state-of-the-art performance across multiple metrics, particularly in handling small targets.

Forensic dentistry research indicates that teeth, as one of the hardest tissues in the human body, can be well-preserved even in disasters and crime scenes, providing a reliable means of human identification. Dental identification based on 3D tooth point clouds has made some progress recently. But the whole-dentition-based identification mode may fail in complex forensic application scenarios such as tooth loss, changes in arrangement or scattering. Therefore, we propose a dental identification deep neural network based on single tooth feature concatenation in this paper, offering more flexibility and accuracy as it can adapt to various tooth arrangements and different tooth numbers. To address the challenges of difficult samples with fewer teeth, a progressive random tooth dropout training strategy is devised to aid learning. For inadequate discrimination of individual tooth features, PointResNet is introduced as a feature extraction module. To handle significant differences in samples with varying tooth numbers and arrangements from the same individual, the additive angular margin loss is adopted. To verify the effectiveness of our method in various dental conditions, we apply random spatial transformations to the teeth of the same individual and simulate different cases of missing teeth on our self-constructed dataset. The results from comparative and ablation experiments confirmed the effectiveness of this approach.