论文阅读 ASFM-Net: Asymmetrical Siamese Feature Matching Network for Point Completion

单位：西电，慕尼黑工业大学，Inceptio
论文链接：https://dl.acm.org/doi/10.1145/3474085.3475348
代码链接：https://github.com/Yan-Xia/ASFM-Net

简介

为解决Point Completion问题，提出了一种Asymmetrical Siamese Feature Matching strategy，包含两个部分：

The Siamese auto-encoder neural network: map the partial and complete input point cloud into a shared latent space, which can capture detailed shape prior.

The iterative refinement unit: generate complete shapes with fine-grained details by integrating prior information.

• 第一和第三行分别是完整点云和残缺点云的自编码过程，其中G1和G3分别代表完整点云和残缺点云的全局特征
• 经过特征matching后，G1和G3的空间分布变得一致（G1固定，G3去匹配G1），然后match后的G2通过refine的操作获得补全点云输出

方法

3.1 Asymmetrical Siamese Auto-encoder

AE1先训练，然后固定，与AE2进行feature mathcing，实验证明这样比只用AE2效果好

3.2 Refinement Unit

• we concatenate the partial inputs with the 𝑃𝑐𝑜𝑎𝑟𝑠𝑒 to form a synthetic point cloud 𝑃𝑠𝑦𝑛𝑡ℎ𝑒𝑡𝑖𝑐 using the farthest points sampling (FPS) algorithm and mirror operations.
• We explore various symmetry operations, including plane symmetry, projective symmetry, and affine transformation operations. Experiments confirm that the XY-plane symmetry achieves the best performance.
• Inspired by FoldingNet [57], we utilize a 2D grid generator and concatenates these 2D grids with each point coordinate to increase the variability of each point.
• In order to narrow the distribution difference between the partial and the complete point cloud, the refinement unit concatenates the global feature with the coordinate of each point in 𝑃𝑠𝑦𝑛𝑡ℎ𝑒𝑡𝑖𝑐.
• Due to the superiority of neural networks in residuals prediction [47], the refinement unit predicts the coordinate offset for every point between the point set 𝑃𝑠𝑦𝑛𝑡ℎ𝑒𝑡𝑖𝑐 and the ground truth point cloud. we pass the 𝑃𝑠𝑦𝑛𝑡ℎ𝑒𝑡𝑖𝑐 through a series of bottom-up and top-down structural styles of MLPs.
  
  where 𝑅(·) denotes the function of predicting the coordinate residuals for the 𝑃𝑠𝑦𝑛𝑡ℎ𝑒𝑡𝑖𝑐 .

3.3 Loss Function

Feature matching loss.

Reconstruction loss.

Overall loss.

实验

PCN数据集4096个点对比

Completion3D数据集2048个点

KTTI数据集

讨论

5.1 Ablation Study

Asymmetrical Siamese autoencoder: We use the pre-trained asymmetrical Siamese autoencoder instead of the encoder modules of PCN and TopNet but keep their decoder modules respectively, referred as SA-PCN Decoder and SA-TopNet Decoder.

Refine unit: We choose the TopNet [45] and PCN [60] as the baseline. The refinement unit is first just integrated into TopNet and PCN, referred as TopNet-Refine and PCN-Refine.

5.2 Robustness Test

5.3 Completion on novel categories