Scholarlink

<p>In 3D point cloud generation tasks, traditional geometric optimization methods struggle to capture complex structures in high-dimensional spaces. Deep learning-based attention mechanisms, Transformers, and graph convolutional networks have significantly improved point cloud representation and generation quality in recent years. However, existing approaches often fail to simultaneously capture multi-scale geometric features while effectively coupling local and global structures. We propose PointAttn-TransGraphNet, an integrated framework for multi-scale attentive feature fusion and point-wise local–global coupling. On the encoder side, PointAttention adaptively aggregates point-wise cues into a global token that is refined by a lightweight Transformer at very low cost. On the decoder side, a GraphSAGE-based hierarchy performs mean aggregation message passing with residual connections on k-nearest neighbor graphs across stages, yielding explicit multi-scale representations; at each upsampling stage, a fusion MLP injects the encoder’s global token into per-point local features, achieving point-wise local–global feature fusion while increasing resolution. We evaluate on two complementary tracks. On ShapeNet, our method achieves consistent and significant advantages across the vast majority of metrics. On ModelNet40, we follow an unsupervised representation learning protocol and report linear evaluation accuracy, where our features outperform strong baselines. Meanwhile, the model attains smaller parameter counts and lower inference time while maintaining comparable FLOPs, validating the effectiveness of the lightweight design. Ablations show that, although compute is decoder-dominated, deepening the lightweight encoder improves global semantics with negligible overhead; module ablations further confirm their complementary roles in multi-scale decoding. The code and data supporting this study are available at <a href="https://github.com/zhangyunyun1/PointAttn-TransGraphNet">https://github.com/zhangyunyun1/PointAttn-TransGraphNet</a> (DOI: <a href="https://doi.org/10.5281/zenodo.15303929">https://doi.org/10.5281/zenodo.15303929</a>).</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

High-fidelity point cloud generation via multi-scale attentive feature fusion with graphSAGE and transformer

  • 期刊论文

  • 发表日期：

    2026年2月28日

查看全文
  • Yunyun Zhang,
  • Zhongrong Zhang,
  • Yulin Shen,
  • Aizhu Li,
  • Kang Lin,
  • Yuyuan Pan

摘要

In 3D point cloud generation tasks, traditional geometric optimization methods struggle to capture complex structures in high-dimensional spaces. Deep learning-based attention mechanisms, Transformers, and graph convolutional networks have significantly improved point cloud representation and generation quality in recent years. However, existing approaches often fail to simultaneously capture multi-scale geometric features while effectively coupling local and global structures. We propose PointAttn-TransGraphNet, an integrated framework for multi-scale attentive feature fusion and point-wise local–global coupling. On the encoder side, PointAttention adaptively aggregates point-wise cues into a global token that is refined by a lightweight Transformer at very low cost. On the decoder side, a GraphSAGE-based hierarchy performs mean aggregation message passing with residual connections on k-nearest neighbor graphs across stages, yielding explicit multi-scale representations; at each upsampling stage, a fusion MLP injects the encoder’s global token into per-point local features, achieving point-wise local–global feature fusion while increasing resolution. We evaluate on two complementary tracks. On ShapeNet, our method achieves consistent and significant advantages across the vast majority of metrics. On ModelNet40, we follow an unsupervised representation learning protocol and report linear evaluation accuracy, where our features outperform strong baselines. Meanwhile, the model attains smaller parameter counts and lower inference time while maintaining comparable FLOPs, validating the effectiveness of the lightweight design. Ablations show that, although compute is decoder-dominated, deepening the lightweight encoder improves global semantics with negligible overhead; module ablations further confirm their complementary roles in multi-scale decoding. The code and data supporting this study are available at https://github.com/zhangyunyun1/PointAttn-TransGraphNet (DOI: https://doi.org/10.5281/zenodo.15303929).