PA-DFNet: Polarity-Aware Attention Network with Feature Dynamic Fusion for Point Cloud Classification and Semantic...

Point cloud segmentation constitutes a core task in 3D computer vision. However, prevailing models suffer from inherent limitations, including the absence of polarity correlation (i.e., spatial attribute-containing features derived from the separation and calculation of positive/negative correlations within point cloud query-key pairs), inefficient feature fusion, loss of fine-grained geometric details, and excessive computational complexity in self-attention mechanisms. These deficiencies constrain both the performance and practical deployment of such models. To address these challenges, the Polarity-Aware Attention and Feature Dynamic Fusion Network (PA-DFNet) is proposed in this paper. Built upon the PointNet++ framework, PA-DFNet replaces the original Multilayer Perceptron (MLP) with a Polarity-Aware Network (PAN). The PAN enhances key semantic interactions by explicitly separating positive and negative correlations from point cloud query-key pairs, generates adaptive neighborhood weights via integration with a linear attention mechanism, and introduces a learnable power function to perform nonlinear scaling of attention, thereby improving the model's structural perception capability. Additionally, a Point Cloud Feature Dynamic Fusion (PFF) module is proposed to enable adaptive fusion of encoder-decoder features, preserving rich geometric details. Experimental results demonstrate that, on the ModelNet40 classification task, the overall accuracy (OA) and mean accuracy (mAcc) of PA-DFNet are improved by 2.4% and 2.2%, respectively, compared with PointNet++. On the S3DIS semantic segmentation task, PA-DFNet achieves an mAcc of 72.8% and a mean Intersection over Union (mIoU) of 66.2%, while exhibiting a shorter training time than Point Transformer. In summary, PA-DFNet achieves an optimal balance between segmentation performance and efficiency by effectively controlling the number of model parameters and computational complexity.