Khang Vi Ly

Although point cloud segmentation has a principal role in 3D understanding, annotating fully large-scale scenes for this task can be costly and time-consuming. To resolve this issue, we propose Point Central Transformer (PointCT), a novel end-to-end trainable transformer network for weaklysupervised point cloud semantic segmentation. Divergent from prior approaches, our method addresses limited point annotation challenges exclusively based on 3D points through central-based attention. By employing two embedding processes, our attention mechanism integrates global features across neighborhoods, thereby effectively enhancing unlabeled point representations. Simultaneously, the interconnections between central points and their distinct neighborhoods are bidirectional cohered. Position encoding is further applied to enforce geometric features and improve overall performance. Notably, PointCT achieves outstanding performance under various labeled point settings without additional supervision. Extensive experiments on public datasets S3DIS, ScanNet-V2, and STPLS3D demonstrate the superiority of our proposed approach over other state-of-the-art methods.

Exploiting global factors and embedding them directly into local graphs in point clouds are challenging due to dense points and irregular structure. To accomplish this goal, we propose a novel end-to-end trainable graph attention network that extracts global features in terms of local graphs. Our network presents general local graph, which obtains the most fundamental features based on point order positions in different neighborhoods. Central point attention is introduced to share weights with neighboring points to reinforce central point impacts. As a result, one point in the specific local graph can obtain global features from corresponding ones in other neighborhoods but still focus on its area by encoded central point weight sharing. Experimental results on benchmark point cloud segmentation datasets demonstrate our proposed network’s competitive performance.

Due to the massive number of points and irregular structure, there is challenging to determine local graph relationships in large-scale point clouds. To address this problem, we propose a novel graph attention network that can embed global factors into neighboring points in specific local graphs. In this network, we introduce general local graph, which integrates all points sharing same order positions with encoded central point weights. In other words, it represents the most fundamental features of total local graphs. Besides, local graphs are regenerated to embed global factors by combining general local graph and central point attention. As a result, one point can obtain additional features from corresponding ones with same order positions in different neighborhoods. Experiment on large-scale point cloud segmentation dataset proves our network’s competitive performance.