Sympo

Sympo: Points | Point Cloud Refinement

Increase density

Table of contents

Upsampling

DL-based

PU Net

Problems:

  1. Upsampling task: Increase density and Increase details of a point cloud.

  2. 3DGS needs to increase density in some cases.

    • In the application of stream transmission, 3DGS needs to be compressed. That means reducing the number of Gaussians before transmission, and restoring density after receiving the low-resolution 3DGS model.

References:

  1. PU-Net: Point Cloud Upsampling Network
  2. 【官方】2025小迈步之使用 AI 求解偏微分方程：探索 PINN 和 NO 的应用 -bilibili - MATLAB中国
  3. 【GAMES Webinar-175期:Bijective Projection in a Shell-Zhongshi Jiang-几何处理专题 - GAMES-Webinar
    • This video was played automatically after the newest episode in the list loop mode.

Addressing:

(2025-04-15T01:01:07)

  1. Patch-wise feature extraction

    • Supports:

      1. Context information

  1. Loss function encourages points on underlying surface

    • Supports:

      1. This paper doesn’t explicitly reconstruct surface. In contrast, MLS does compute a surface with polynomials.

      2. Splats rendering doesn’t require a mesh, but the underlying surface is useful for further process. For example, the surface can serve as a reference when the resolution of a set of 3D Gaussians is required to be adaptive to present different LOD.

        Therefore, determining the surface is an important task.

        (2025-04-17T06:34)

        • Splats 在应用中可能确实不需要 mesh 参与，但是 mesh 可以作为处理过程中的中间态（类比于“潜空间”）， 比如构建出 mesh 之后，可以通过做 remesh 来实现先从稀疏到稠密，再从稠密到稀疏的映射r3-Jiang（这让我想起了神经网络的维度变化：先升维再降维）， 从而实现上采样。

          不过，这种上采样方式还是依赖 mesh 的质量。

      (2025-04-15T23:51:43)

      1. PINN embeds physics information into network through loss functionsr2-小迈步

        • Another way is using neural operator

    • Actions:

Representation

GaussianPU

Problems:

  1. I suspected the gaps between Gaussians will be revealed when zooming in.

References:

  1. GaussianPU: A Hybrid 2D-3D Upsampling Framework for Enhancing Color Point Clouds via 3D Gaussian Splatting

Flow Matching

EF

References:

  1. Efficient Point Clouds Upsampling via Flow Matching

Registration

SDFReg

References:

  1. SDFReg: Learning Signed Distance Functions for Point Cloud Registration

Completion

DL

(2025-04-14T14:32:48)

  • These four papers are for point cloud completion tasks.

PointAttN

Problems:

  1. “摆脱对 KNN 的依赖”

References:

  1. 秘塔 | 今天学点啥
  2. 元宝 | 笔记整理

Notes:

  1. What are their innovations?

    • Reasons:

      1. Figuring out their novelty to compare it with mine method.

    • Actions:

      1. existing methods use explicit local region partitions like kNNs, which makes them sensitive to density distribution and limits receptive fields.

        • I do used kNN.
        • Reviewer 2 also said “sensitivity”

(2025-04-30T01:03)

  1. How do they generate new points?

    • Supports:

      1. A point cloud learn itself through a downsampled version r1-秘塔, r2-元宝

      (2025-04-30T23:38)

      1. New points are Seeds, which are feature vectors Line #206

DMF-Net

References:

  1. DMF-Net

PINN

Problems:

(2025-04-15T19:17:00)

  1. I believe a point cloud is a physical field, which is supposed to be approximated/regressed by a PINN.

    Specifically, the observation (image or sampled points) are training data, from which the physics field is integrated into a neural network.

Addressing:

(2025-04-16T18:33:00)

  1. PINN: 把物理先验知识加入神经网络的训练中r2-小迈步

    • Supports:

      1. NERF 也是pinn 吧，其中的物理先验是渲染方程。渲染方程是光传输的近似，所以也可以直接应用光传输或者计算成像的物理方程作为损失函数

Projection

EAR

Problems:

  1. Enhance edges in a point cloud

References:

  1. Edge-Aware Point Set Resampling - VCC
  2. 秘塔-今天学点啥 | Edge-Aware Point Set Resampling

Notes:

(2025-05-03T23:21)

  1. 利用平滑区域的法向量指导填充边缘的法向量r2-秘塔

    • Supports:

      1. EAR 依据平滑区域的法向量渐进式 Progressively 推测尖锐区域的法向量

      2. Steps:

        1. Ensure the confidence of normal vectors of the flat areas: Denoising + 重投影

        2. Completing edges: Progressively upsampling by expanding flat arear

        flowchart LR A["EAR"] --> Goal["尖锐区域的法向量"] A --> Base["利用平滑区域"] -- "逐步填充
        双侧投影(新点)" --> Goal Denoise["双边滤波
        去除噪声
        测试不同 level 噪声"] Refine["各向异性重投影
        异构投影"] subgraph Prep direction BT Denoise --> Refine end Refine --> Base

      3. Usages of normal vectors

        1. EAR：法向量 -> “检测”/突出边缘 -> 投影填充，增强边缘

        2. 可微渲染 (3DGS)：像素在多视图中的一致性 -> 点的空间位置 -> 几何 -> 采样 -> 渲染

      4. 辐射度缩放？

        • 这也许也说明了：辐射度与几何之间存在关联，可以利用辐射度推断几何

Meso-Skeleton

D-Points

References:

  1. Deep Points Consolidation - VCC
  2. 元宝 | 计算机图形学笔记整理

Notes:

(2025-05-04T02:07)

  1. 以“中尺度骨架”作为全局信息r2-元宝

    • Supports:

      1. D-point 是把 Meso-skeleton 的点与外层点用 直线段相连，那我是否可以用 多项式曲线 连接一个内点和一个外点呢？

      2. D-Point 的叙事也是：结合基础层次的 全局信息和外延的局部信息

      3. D-Point 可以保留 动物的翅膀，这个任务 不常见。 以及对细微物体的重建也是一个值得探索的课题。

      4. Repulsion 的目的是鼓励形成曲面？

        ⋯ ⋯

      5. 对点云做搜索（retrieveal）也很重要啊！

        • 比如一个场景被扫描成点云之后（点云不包含语义信息），我想搜索：“房间里有几把椅子？”。

        • 所以这个问题是：如何对点云做物体识别？（PointNet 是做什么的？）

        • 直接对点云做操作处理，计算量太大，而这种“深点”可以找到点云的拓扑吗？Meso-skeleton 是点云的拓扑吗？

          应该不是，因为输入点云的形状不一定完整，提取出来的拓扑结构不是物体本身的拓扑。

        • 不过把点云的拓扑结构作为神经网络的训练数据，应该比较高效

Denoising

References:

  1. NoiseTrans: Point Cloud Denoising with Transformers
© 2020 - 2025 Zichen Wang
Built with Hugo
Theme Stack designed by Jimmy