Conical Window-Attentive Point Cloud Segmentation Framework

The sparse and uneven nature of point cloud data leads to suboptimal performance of traditional segmentation approaches at greater distances, resulting in a degradation of overall accuracy. To tackle this challenge, this study employs a cone window technique to effectively segment point clouds, thereby aggregating long-range information. Additionally, exponential positional encoding is introduced to enhance the capability of positional encoding. Ultimately, an adaptive feature selection module is utilized to yield more effective features. Experimental results show that the proposed algorithm in this paper effectively improves the model’s segmentation accuracy and outperforms other solutions proposed in the field.