Scholarlink

<p>Dynamic point clouds (DPCs) represent realistic 3D scenes in motion and have a broad range of applications. Efficient compression of point clouds is essential for the storage and transmission of such data. Video-based Point Cloud Compression (V-PCC), developed by the Moving Picture Experts Group (MPEG), has demonstrated outstanding performance in compressing DPCs. In this manuscript, we present an enhancement to our previously proposed 3D-based post-processing framework for the V-PCC decoder. The framework comprises two key components: (1) a systematically optimized lightweight BaseNet stream, which employs depth and width reduction strategies to correct geometric distortions in existing points, and (2) an InterpolateNet stream, which complements missing points. This dual-stream approach enables efficient and simultaneous handling of both types of degradation. Furthermore, we propose a comprehensive two-stage training strategy based on the 3D Minkowski UNet architecture. In the first stage, reliable ground truth is established by aligning compressed point clouds with their original counterparts through V-PCC matching. In the second stage, dynamic prediction refinement is applied, where the model’s intermediate outputs are trained against the original ground truth to achieve optimal enhancement performance. Experimental results demonstrate that the proposed framework significantly improves the restoration of reduced coordinates and overall point cloud quality, as validated by both objective metrics and subjective evaluations.</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

Video-based point cloud compression quality enhancement using deep learning approach

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Siu-Kei Au Yeung,
  • Haoyu Dong,
  • Zhitan Wang

摘要

Dynamic point clouds (DPCs) represent realistic 3D scenes in motion and have a broad range of applications. Efficient compression of point clouds is essential for the storage and transmission of such data. Video-based Point Cloud Compression (V-PCC), developed by the Moving Picture Experts Group (MPEG), has demonstrated outstanding performance in compressing DPCs. In this manuscript, we present an enhancement to our previously proposed 3D-based post-processing framework for the V-PCC decoder. The framework comprises two key components: (1) a systematically optimized lightweight BaseNet stream, which employs depth and width reduction strategies to correct geometric distortions in existing points, and (2) an InterpolateNet stream, which complements missing points. This dual-stream approach enables efficient and simultaneous handling of both types of degradation. Furthermore, we propose a comprehensive two-stage training strategy based on the 3D Minkowski UNet architecture. In the first stage, reliable ground truth is established by aligning compressed point clouds with their original counterparts through V-PCC matching. In the second stage, dynamic prediction refinement is applied, where the model’s intermediate outputs are trained against the original ground truth to achieve optimal enhancement performance. Experimental results demonstrate that the proposed framework significantly improves the restoration of reduced coordinates and overall point cloud quality, as validated by both objective metrics and subjective evaluations.