Contrastive Adversarial Loss for Point Cloud Reconstruction

Tianxin Huang, Zhonggan Ding, Jiangning Zhang, Ying Tai, Zhenyu Zhang, Mingang Chen, Chengjie Wang, Yong Liu

Keywords: Point clouds, reconstruction loss, learning-based

TL;DR: Learn a point cloud reconstruction loss by contrastive constraint and adversarial training

Abstract: For point cloud reconstruction-related tasks, the reconstruction losses to evaluate the shape differences between reconstructed results and the ground truths are typically used to train the task networks. The Chamfer Distance (CD) and Earth Mover's Distance (EMD) are two widely-used reconstruction losses, which firstly use predefined strategies to match points in two point clouds and then apply the average distances from points to their matched neighbors as differentiable measurements of shape differences. However, the predefined matching rules may deviate from the real shape differences and cause defective reconstructed results. To solve the above problem, we propose a learning-based Contrastive adversarial Loss (CALoss) to train a reconstruction-related task network without the predefined matching rules. CALoss learns to evaluate shape differences by combining the contrastive constraint with the adversarial strategy. Specifically, we use the contrastive constraint to help CALoss learn shape similarity, while we introduce the adversarial strategy to help CALoss mine differences between reconstructed results and ground truths. According to experiments on reconstruction-related tasks, CALoss can help task networks improve reconstruction performances and learn more representative representations.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning