Dynamic-focus transformer for point cloud segmentation

Transformer-based methods have significantly advanced 3D point cloud segmentation by effectively capturing long-range dependencies. However, the global or fixed-window self-attention mechanisms they often employ suffer from computational redundancy and overfitting due to processing excessive, potentially irrelevant key-value pairs for each query. To address this, we propose the Dynamic-Focus Transformer, a novel architecture that introduces a data-dependent adaptive attention mechanism. Through learned soft point masks, we selectively sparsify keys and values to focus on semantically critical regions. Our method enables flexible, input-adaptive receptive fields without the heavy memory overhead associated with per-point offset learning in deformable designs. Furthermore, when integrated into a U-Net-style encoder-decoder, our method attains a highly efficient balance between modeling capability and computational cost. Extensive experiments on S3DIS and ScanNetv2 benchmarks demonstrate that our method achieves state-of-the-art performance with notably improved efficiency, validating its effectiveness for large-scale point cloud understanding.