Point clouds exhibit non-uniform spatial distributions and varying scale characteristics. Under such conditions, single-scale feature representations are insufficient to capture both local geometric details and high-level semantic information. As a result, multi-scale feature aggregation has become essential. Existing methods generally follow two strategies for multi-scale processing. One category of methods introduces multi-scale receptive fields within local neighborhoods. This is commonly achieved through mechanisms such as multi-scale grouping or parallel feature extraction to enhance local geometric feature representations, as explored in PointNet++ [3], KPConv [12] and PointConv [13]. Recent studies have further explored adaptive multi-scale feature aggregation by leveraging voxel-based context and point-level representations, enabling more flexible receptive field adaptation for point cloud semantic segmentation [14,15]. Another category focuses on cross-level feature fusion within encoder–decoder architectures, where features at different resolutions are integrated through skip connections or feature concatenation. Representative models following this strategy include RandLA-Net [16], PointNeXt [8] and U-Next [17].

Although the above methods alleviate the loss of fine-grained information caused by downsampling to some extent, most of them rely on fixed feature fusion strategies. Such designs make it difficult to balance feature contributions across scales under varying scene structures. Moreover, some multi-scale approaches depend on parallel branches or introduce additional parameters, leading to increased model complexity. Although segmentation accuracy is improved, this increase in complexity limits their suitability for direct integration into efficiency-oriented baseline networks. A key challenge, therefore, lies in enabling more flexible multi-scale feature aggregation while maintaining model efficiency.

To enhance the capture of long-range dependencies and global contextual information, attention mechanisms have been increasingly applied to point cloud semantic segmentation. Self-attention–based models explicitly capture global relationships among points and enable more flexible feature interactions. Representative approaches following this direction include Point Transformer [18], Point Cloud Transformer (PCT) [19], Stratified Transformer [20], and Swin3D [21,22], which have demonstrated competitive performance in semantic segmentation tasks.

However, global context integration based purely on self-attention often incurs high computational and memory costs [23]. This issue becomes more pronounced when processing large scale point cloud data, which limits scalability to some extent. Motivated by this issue, recent studies have explored more efficient and scalable attention designs to reduce the computational overhead of global context integration in point cloud semantic segmentation [24,25]. To further alleviate this problem, recent studies have explored combining attention mechanisms with local operators such as convolutions or MLPs. These approaches introduce non-local contextual information while preserving local spatial structures, aiming to balance performance and efficiency. Representative examples include Superpoint Transformer [26], which introduces an efficient superpoint representation to model contextual relationships in large scale point clouds, and Cross-Fusion Self-Attention Network (CFSA-Net) [27], which employs a cross-fusion self-attention mechanism to jointly capture local structures and long-range contextual dependencies. Although such hybrid methods reduce computational complexity, they still struggle to effectively coordinate local geometric features with global semantic information.

Although PointNeXt achieves a good balance between efficiency and performance, its single-scale grouping strategy in the encoder inevitably leads to the loss of fine-grained geometric information. This design also limits effective feature fusion across different receptive fields. These limitations become more apparent in indoor scenes with non-uniform point distributions or complex structures, where segmentation accuracy around object boundaries and thin structures is often degraded.

To alleviate these issues, we introduce a Multi-Scale Feature Enhancement (MSFE) module to replace and enhance the conventional cross-level feature propagation mechanism used in PointNeXt. As illustrated in Fig. 3, the MSFE module incorporates multi-scale feature aggregation and feature refinement in the decoding stage. By more effectively leveraging features from different encoding levels, the module supports the recovery of geometric details lost during downsampling and produces more informative representations for semantic segmentation.

The MSFE module consists of two cascaded sub-units: an Adaptive Multi-Scale Aggregation Block (AMSAB) and an Activation-Free Feature Refinement Block (AFFRB). The AMSAB adaptively fuses features from different scales, while the AFFRB further refines the fused features to achieve stable feature representations.

Local neighborhood aggregation alone constrains the effective receptive field and limits the capture of non-local semantic dependencies. In complex indoor environments, structurally or semantically correlated regions may not be spatially adjacent, which necessitates explicit non-local interaction. To this end, a Convolution-Attention Mixing (CA-Mix) module is introduced, in which the attention branch explicitly captures non-local contextual dependencies, while the convolution branch preserves local geometric structures. As the non-local component of the proposed non-local multi-scale aggregation framework, CA-Mix complements multi-scale feature enhancement and improves semantic improves segmentation performance in complex scenes.

PointNMSA is evaluated on the S3DIS dataset [9], which is a widely used benchmark for indoor point cloud semantic segmentation. The dataset contains 272 room scans collected from six indoor areas (Area 1–Area 6) and is annotated with 13 semantic categories. Following the standard evaluation protocol, Area 5 is used as the test set, while the remaining areas are used for training. All experimental results are reported on Area 5.

In addition, experiments are conducted on the ScanNet dataset [10], another large scale benchmark for indoor scene understanding. ScanNet consists of richly annotated RGB-D reconstructions of real indoor environments and contains 1513 scanned scenes with 20 semantic categories commonly used for evaluation. Following the official data split, the standard training and validation sets are adopted for performance comparison.

During data preprocessing, point cloud coordinates are first normalized, and voxel-based downsampling is applied to reduce computational cost, with the voxel size set to 0.04. This setting provides a reasonable balance between preserving geometric details and controlling computational complexity, and is consistent with the preprocessing strategy used in the PointNeXt baseline. Instead of block-based cropping, entire rooms are adopted as network inputs to retain global spatial structure. For fair comparison, all baseline methods follow the same preprocessing pipeline and input configuration.

For performance evaluation, widely accepted metrics in 3D point cloud semantic segmentation are adopted to provide a comprehensive and fair assessment. These metrics include mean Intersection over Union (mIoU), Overall Accuracy (OA), and mean Class Accuracy (mAcc). Specifically, mIoU is computed by averaging the intersection-over-union scores over all semantic categories and reflects overall segmentation quality. OA measures the proportion of correctly classified points, while mAcc computes the average accuracy across all classes, which helps mitigate the impact of class imbalance.

All experiments are conducted on a workstation equipped with an NVIDIA RTX 3090 GPU (24 GB memory), running Ubuntu 20.04. Model training and inference are implemented using PyTorch and accelerated with Compute Unified Device Architecture (CUDA).

During training, the cross-entropy loss is used as the optimization objective, and a label smoothing strategy is applied to improve generalization, with the smoothing factor set to 0.2. The model is trained for 100 epochs with a batch size of 16. The AdamW optimizer is adopted with a weight decay of 0.05, and gradient norm clipping is applied with a maximum value of 28 to stabilize training. The initial learning rate is set to 0.006 and is dynamically adjusted using a cosine annealing schedule to ensure stable convergence throughout training.

For the main ablation results in overall module ablation results on S3DIS Area 5, each configuration was evaluated over three independent runs with different random seeds. The reported results are averaged across these runs, and the mean with standard deviation of mIoU is additionally provided to reflect performance stability, as mIoU is the primary evaluation metric for semantic segmentation. For the remaining ablation and comparison experiments, all results are obtained from a single run with a fixed random seed to ensure fair and consistent comparisons across different methods and configurations.

To evaluate the effectiveness of the proposed Multi-Scale Feature Enhancement (MSFE) and Convolution-Attention Mixing (CA-Mix) modules, systematic ablation studies are conducted on the PointNeXt baseline [8]. All experiments are carried out under the same data split, training strategy, and evaluation settings to ensure fair comparison.

As shown in Table 1, progressively introducing MSFE and CA-Mix into the PointNeXt baseline leads to consistent improvements in segmentation performance. When MSFE is added alone, the mIoU increases from 63.45% to 64.50%, while both OA and mAcc also improve slightly. This improvement indicates that multi-scale feature enhancement helps recover geometric details lost during hierarchical downsampling. When CA-Mix is introduced independently, the mIoU increases to 64.82%, suggesting that incorporating non-local contextual interaction improves semantic discrimination in complex indoor scenes. Compared with MSFE, CA-Mix yields a larger performance gain, indicating that it plays a more dominant role in driving the overall improvement, while MSFE mainly provides complementary enhancement.

When both modules are integrated, the proposed PointNMSA achieves the best performance with 65.05% mIoU, 88.78% OA, and 71.30% mAcc. However, the combined improvement is smaller than the sum of their individual gains, indicating that the two modules are not strictly additive. This can be attributed to their partially overlapping effects in structurally complex regions, which limit the accumulated gain. Such overlap is more likely to occur in categories characterized by ambiguous boundaries or fine-grained structures (e.g., window and board), where both global context and local geometric details are simultaneously required, leading to partially redundant improvements. Similarly, for thin or sparsely represented structures (e.g., column), both modules may respond to the same structural cues, further reducing the independence of their contributions. Therefore, the advantage of combining the two modules lies not in a simple superposition of gains, but in achieving a better balance between geometric detail preservation and contextual semantic discrimination.

The relatively small standard deviations reported in Table 1 (ranging from ±0.14 to ±0.18) indicate that the performance variations across independent runs remain limited, suggesting that the observed improvements are consistently achieved rather than caused by random fluctuations.

Fig. 5 further illustrates the performance-efficiency trade-off under different module configurations. As shown in Fig. 5a, the proposed method achieves consistently improved segmentation performance across different module combinations, indicating the effectiveness of the introduced components. Fig. 5b,c further reveal how this performance improvement is associated with increases in parameter count and inference latency, respectively. According to Table 1, incorporating MSFE increases the parameter count slightly from 3.16 to 3.35M, while the latency rises from 42.24 to 43.41 ms, indicating that the additional overhead in both model size and inference time remains limited. In contrast, introducing CA-Mix results in a more significant increase in parameters from 3.16 to 8.48M, while the latency grows from 42.24 to 44.62 ms. Notably, the increase in inference latency is much smaller than the increase in parameter count. Although the parameter growth is noticeable, it mainly comes from the multi-branch design and additional projection and channel mixing layers introduced in the contextual interaction module, which increase model capacity but do not proportionally increase computational intensity during inference. As a result, the growth in actual inference time remains moderate despite the larger model size. When both modules are combined, the parameter count reaches 8.67M and the latency increases to 45.18 ms, representing a moderate overall increase compared to the baseline. Taken together, these results indicate that the proposed design maintains a controlled increase in inference cost while achieving improved segmentation performance, demonstrating a favorable trade-off between model complexity and practical efficiency.

To further analyze the internal design of the MSFE module, additional ablation experiments are conducted, as reported in Table 2. Starting from the baseline model, introducing the Adaptive Multi-Scale Aggregation Block (AMSAB) alone improves the mIoU from 63.51% to 64.09%, indicating that adaptive multi-scale feature fusion helps capture complementary geometric and semantic information from different resolutions. This first step already brings a gain of 0.58 mIoU with only a 0.07M increase in parameters, suggesting that AMSAB provides a favorable balance between accuracy and computational complexity. Therefore, AMSAB can be regarded as the primary contributor within the MSFE module. When the Spatial-Channel Attention (SCA) mechanism is added (A2), the mIoU further increases to 64.31%, suggesting that feature recalibration enhances the discriminative representation of important features. Compared with A1, A2 yields an additional gain of 0.22 mIoU with another 0.07M parameters, indicating that SCA mainly enhances the discriminative ability of the fused features rather than introducing new information. Finally, incorporating the FFN refinement unit (A3) leads to the full MSFE module and achieves 64.56% mIoU. Relative to A2, the FFN refinement contributes a further 0.25 mIoU improvement with only 0.05M additional parameters, showing that the refinement stage further improves the feature representation. Compared with AMSAB, the contributions of SCA and FFN are more incremental, mainly providing complementary refinement rather than dominating the performance improvement. AMSAB contributes cross scale feature aggregation, SCA enhances feature selection after fusion, and FFN further refines the aggregated representation. The full MSFE module therefore outperforms each partial variant because these components operate in a complementary manner rather than repeating similar functions.

Table 3 further investigates the contributions of different components in the CA-Mix module. When only the global attention branch is enabled (B1), the mIoU improves to 64.07%, indicating that non-local contextual interaction helps capture long-range semantic dependencies. The parameter count increases to 5.30M, which can be attributed to the additional projection and interaction operations introduced by the attention mechanism. Using only the local convolution branch (B2) also leads to a slight improvement (63.96% mIoU) with a relatively smaller parameter increase (4.91M), since convolution mainly performs localized feature aggregation without introducing extensive pairwise interactions. When both branches are combined (B3), the mIoU further increases to 64.54%, confirming that global and local feature aggregation provide complementary information, while the parameter count rises to 6.59M due to the joint modeling of local structures and long-range dependencies. This indicates that the integration of global and local branches provides a more balanced representation, but neither branch alone dominates the overall performance. In addition, introducing the Multi-Dimensional Attention Fusion Block (MDAFB) alone (B4) improves the mIoU to 64.18% with 5.26M parameters, suggesting that feature recalibration enhances discriminative responses with moderate structural overhead. Compared with the global and local branches, MDAFB mainly refines the fused features and plays a complementary role, rather than being the primary driver of performance improvement. When all components are integrated (B5), the full CA-Mix module achieves the best performance of 64.88% mIoU, while the parameter count increases to 8.48M. This increase mainly comes from the combined effect of attention-based interaction, convolutional aggregation, and feature fusion operations. From a performance and model complexity perspective, B1 and B2 provide relatively lightweight improvements, B3 offers a more balanced configuration by integrating global and local representations, and B5 achieves the highest accuracy with increased structural complexity. Overall, the performance gain of CA-Mix is primarily driven by the global attention branch, while the local convolution branch and MDAFB provide complementary and refinement effects. This suggests that the improvement stems not merely from increased model capacity, but from the coordinated interaction among global attention, local convolution, and feature fusion mechanisms.

To further evaluate the effectiveness, PointNMSA is compared with several representative point cloud semantic segmentation approaches on the S3DIS Area 5 benchmark. The quantitative comparison results are summarized in Table 4.

Compared with earlier point-based networks such as PointNet and PointCNN, PointNMSA achieves better segmentation performance. The improvement mainly benefits from the enhanced feature representation introduced by the proposed modules. When compared with several representative methods such as MinkowskiNet and PCM-Tiny, PointNMSA also achieves competitive performance. Although some methods, such as KPConv and Point Transformer, report higher mIoU values on this benchmark, they adopt more complex convolution operators or transformer-based architectures for feature interaction. In contrast, PointNMSA focuses on enhancing the PointNeXt backbone by incorporating multi-scale feature enhancement and non-local context interaction modules. The consistent improvements over the baseline demonstrate that PointNMSA effectively improves feature representation while maintaining the original architectural framework.

From the per-class Intersection over Union (IoU) results reported in Table 5, the improvements brought by PointNMSA are mainly concentrated on several fine-grained and structurally ambiguous categories. In particular, the IoU gains for window and board reach +8.30 and +8.98, respectively, which contribute most to the overall performance improvement. In addition, categories such as door (+2.23), bookcase (+1.64), and clutter (+2.24) also show consistent positive gains. These categories are often characterized by thin structures, sparse point distributions, or semantic similarity with surrounding objects in indoor scenes. The improvements suggest that the multi-scale feature enhancement and non-local context integration in PointNMSA help improve prediction consistency in regions with ambiguous boundaries and strong semantic confusion. It should also be noted that the IoU of the column category decreases slightly (22.79% → 18.12%). This reduction may be related to the relatively small number of column instances in the dataset and the sensitivity of slender structures to voxel-based downsampling [30]. Despite this localized decline, PointNMSA still achieves improvements in several challenging categories and leads to a higher overall mIoU of 65.10%, compared with 63.51% for the baseline.

To further evaluate the generalization ability of PointNMSA, additional experiments are conducted on the ScanNet dataset. The same model configuration is directly applied to ScanNet without modifying the network architecture. The results are reported in Table 6.

As shown in Table 6, PointNMSA consistently improves segmentation performance over the PointNeXt baseline on the ScanNet dataset. The mIoU increases from 65.25% to 67.34%, while OA and mAcc also show noticeable improvements. Similar to the results on S3DIS, introducing either MSFE or CA-Mix individually improves performance, and their combination achieves the best results. These results indicate that PointNMSA maintains stable performance across different indoor datasets and demonstrates good cross-dataset generalization ability.

To further examine the segmentation behavior of PointNMSA in complex indoor scenes, qualitative visual comparisons are conducted on two representative rooms from the S3DIS Area 5 test set, as shown in Figs. 6 and 7. The figure presents the ground-truth labels, the predictions of the PointNeXt baseline, and PointNMSA predictions for comparison.

As shown in Fig. 6a–c, this scene focuses on the segmentation performance for the Board category. Since boards are typically adjacent to walls and exhibit relatively flat geometries with weak boundary cues, the PointNeXt baseline in Fig. 6b tends to produce blurred boundaries and confusion with the Wall category, resulting in partial misclassification of board regions. In contrast, PointNMSA in Fig. 6c more accurately recovers the overall board structure within the highlighted regions, yielding clearer boundaries and predictions that are more consistent with the ground truth shown in Fig. 6a. These results demonstrate the advantage of the proposed approach in capturing fine-grained planar structures.

As illustrated in Fig. 7a–c, this scene highlights the segmentation results for the Window category. Windows usually exhibit thin structures, sparse point distributions, and geometric similarity to surrounding background regions. Under these conditions, the PointNeXt baseline in Fig. 7b often misclassifies window points as Clutter, leading to noticeable semantic noise and fragmented predictions that degrade structural continuity. By comparison, PointNMSA in Fig. 7c significantly reduces misclassified Clutter points within the window regions, enabling a more complete recovery of window structures in the highlighted areas. This improvement can be attributed to the incorporation of multi-scale feature enhancement and non-local contextual integration, which helps enforce prediction consistency when local geometric cues are limited.