MCTNet: Multiscale Cross-Attention-Based Transformer Network for Semantic Segmentation of Large-Scale Point Cloud

In this work, we implement a hybrid method to utilize sufficient information by aggregating both fine-grained and globally contextual features for point cloud semantic segmentation with a hierarchical network. By surpassing the defects of convolution operation mainly for extracting low-level features, we combine higher level cross-attention-based transformers to investigate the importance of long-range relations together with position embedding for multiscale feature representation. Specifically, by adding a learnable token to the feature sequence of a layer, a transformer encoder is first implemented with limited scope to embed these features. Furthermore, instead of performing all-to-all attention, we merely fuse tokens spanning various scales. To improve efficiency, we propose a simple yet efficient token-fusing architecture based on cross-attention, in which the computation of attention maps can be restricted within linear time by only using a token to calculate the query. The cross-attention module can be efficiently aggregated in a multiscale network to further enlarge the scope of the receptive field for attention. Experiments show that our multiscale cross-attention-based transformer network (MCTNet) achieves promising results on the three largest point cloud datasets, DALES, DublinCity, and S3DIS datasets. For the DALES benchmark dataset, MCTNet improves the mean intersection-over-union (mIoU) to 83.3% and the overall accuracy (OA) to 98.3%, which outperforms other existing baselines. We also perform abundant ablation studies on various attention and normalization modules and discuss the effect of parameters to validate the descriptive power of cross-attention modules and provide an understanding of how long-range dependency can be used to learn fair and unbiased features.