Enhancing Performance of Point Cloud Completion Networks with Consistency Loss

Christofel Rio Goenawan,Kevin Tirta Wijaya,Seung-Hyun Kong

2024-10-15

Abstract:Point cloud completion networks are conventionally trained to minimize the disparities between the completed point cloud and the ground-truth counterpart. However, an incomplete object-level point cloud can have multiple valid completion solutions when it is examined in isolation. This one-to-many mapping issue can cause contradictory supervision signals to the network because the loss function may produce different values for identical input-output pairs of the network. In many cases, this issue could adversely affect the network optimization process. In this work, we propose to enhance the conventional learning objective using a novel completion consistency loss to mitigate the one-to-many mapping problem. Specifically, the proposed consistency loss ensure that a point cloud completion network generates a coherent completion solution for incomplete objects originating from the same source point cloud. Experimental results across multiple well-established datasets and benchmarks demonstrated the proposed completion consistency loss have excellent capability to enhance the completion performance of various existing networks without any modification to the design of the networks. The proposed consistency loss enhances the performance of the point completion network without affecting the inference speed, thereby increasing the accuracy of point cloud completion. Notably, a state-of-the-art point completion network trained with the proposed consistency loss can achieve state-of-the-art accuracy on the challenging new MVP dataset. The code and result of experiment various point completion models using proposed consistency loss will be available at: <a class="link-external link-https" href="https://github.com/kaist-avelab/ConsistencyLoss" rel="external noopener nofollow">this https URL</a> .

Computer Vision and Pattern Recognition,Artificial Intelligence

What problem does this paper attempt to address?

This paper attempts to solve a key problem in Point Cloud Completion Networks (PCCNs): the **one - to - many mapping issue**. Specifically, when dealing with incomplete 3D point - cloud data, the same incomplete point cloud may have multiple reasonable completion results. In this case, traditional loss functions may generate contradictory supervision signals, affecting the network optimization process and thus the final completion performance.

### Detailed Explanation

1. **Problem Background**:
- Point cloud completion is part of 3D reconstruction tasks, aiming to recover the complete 3D shape from incomplete or occluded point - cloud data.
- Traditional methods are usually trained by minimizing the difference between the predicted complete point cloud and the ground truth, commonly using the Chamfer Distance (CD) metric.
- However, for an incomplete point cloud, there may be multiple valid completion schemes. This will lead to different outputs corresponding to the same input during the training process, and then generate contradictory supervision signals, making it difficult for the network to converge to the optimal solution.

2. **Proposed Problem and Solution**:
- **Consistency Loss**: To address this challenge, the author proposes a new loss function - consistency loss. The core idea of this loss function is to consider multiple incomplete views from the same source object simultaneously instead of independently examining each view.
- **Specific Implementation**:
- **Self - guided Consistency**: Generate multiple incomplete point clouds of the same object and ensure that the completion results of these point clouds are consistent.
- **Target - guided Consistency**: Use the original ground truth to guide the consistency loss, ensuring that the completion results are as close as possible to the ground truth.

3. **Experimental Verification**:
- The author conducted experiments on multiple existing PCCN models, including PCN, AxFormNet, AdaPoinTr, etc., to verify the effectiveness of the consistency loss.
- The experimental results show that after adding the consistency loss, the completion performance of these models on multiple benchmark datasets has been significantly improved, especially when dealing with objects with diverse geometric shapes.

4. **Conclusion**:
- The consistency loss can effectively alleviate the one - to - many mapping problem, improve the performance and generalization ability of point cloud completion networks, and does not require modification of the network architecture.
- This method also performs well on the latest MVP dataset, further demonstrating its potential in practical applications.

### Formula Summary

- **Chamfer Distance (CD)**:
\[
CD(A, B)=\frac{1}{|A|} \sum_{a \in A} \min_{b \in B}\|a - b\|_{2}^{2}+\frac{1}{|B|} \sum_{b \in B} \min_{a \in A}\|b - a\|_{2}^{2}
\]

- **Self - guided Consistency Loss (Lc - sg^k)**:
\[
L_{c - sg}^{k}=\frac{2}{n(n - 1)} \sum_{i = 1}^{n - 1} \sum_{j = i + 1}^{n} CD(\hat{P}_{com}^{k, i}, \hat{P}_{com}^{k, j})
\]

- **Target - guided Consistency Loss (Lc - tg^k)**:
\[
L_{c - tg}^{k}=\frac{1}{n} \sum_{i = 1}^{n} CD(\hat{P}_{com}^{k, i}, P_{com}^{k})
\]

- **Total Loss Fu