Scholarlink

<p>The restoration of the Terracotta Warriors faces challenges due to the lack of large-scale, high-quality annotated datasets. We present PointDecoupler, a novel contrastive learning framework that explicitly disentangles augmentation-invariant representations (AIR) and augmentation-variant representations (AVR) to improve efficiency and generalization. Our method includes two core components: (1) a novel decoupling architecture with an adaptive loss function that systematically disentangles and utilizes AVR information to optimize downstream adaptability; (2) a cross-layer contrastive mechanism inspired by self-distillation, enabling intermediate layers to acquire discriminative features from the final layer. This dual strategy improves feature quality and supports early-exit subnetworks that reduce computational cost without sacrificing performance. Experiments on standard point cloud benchmarks demonstrate consistent gains in classification and segmentation. We further applied PointDecoupler to the Terracotta Warriors dataset, achieving promising results and demonstrating its potential in cultural heritage restoration.</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

Self-supervised learning via disentangled representation and self-distillation for 3D terracotta warriors

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月10日

查看全文
  • Xin Cao,
  • Linrong Ye,
  • Xinxin Han,
  • Jia Zhang,
  • Xingxing Hao,
  • Linzhi Su,
  • Kang Li

摘要

The restoration of the Terracotta Warriors faces challenges due to the lack of large-scale, high-quality annotated datasets. We present PointDecoupler, a novel contrastive learning framework that explicitly disentangles augmentation-invariant representations (AIR) and augmentation-variant representations (AVR) to improve efficiency and generalization. Our method includes two core components: (1) a novel decoupling architecture with an adaptive loss function that systematically disentangles and utilizes AVR information to optimize downstream adaptability; (2) a cross-layer contrastive mechanism inspired by self-distillation, enabling intermediate layers to acquire discriminative features from the final layer. This dual strategy improves feature quality and supports early-exit subnetworks that reduce computational cost without sacrificing performance. Experiments on standard point cloud benchmarks demonstrate consistent gains in classification and segmentation. We further applied PointDecoupler to the Terracotta Warriors dataset, achieving promising results and demonstrating its potential in cultural heritage restoration.