Reproduce “PointMAC: Meta-Learned Adaptation for Robust Test-Time Point Cloud Completion” on OpenResearch

About this paper

Linlian Jiang, Rui Ma, Li Gu, Ziqiang Wang, Xinxin Zuo, Yang Wang

Researchers at Concordia University and Mila developed PointMAC, a framework that employs meta-learned test-time adaptation (TTA) to refine point cloud completions dynamically for individual test samples. This approach consistently achieved state-of-the-art results across synthetic, simulated, and real-world datasets, reducing Chamfer Distance (CD-l1) to 6.33 (x1000) on the PCN dataset.

Key insights

• Applying meta-learned test-time adaptation for the first time to point cloud completion effectively shifts from static, generic inference to dynamic, sample-specific refinement.
• Self-supervised auxiliary tasks, when harmonized through meta-learning and adaptive weighting, can robustly guide model adaptation without requiring ground truth labels during inference.
• Dynamically calibrating auxiliary loss weights during training prevents task interference and ensures stable, performance-enhancing gradient contributions for the primary task.

Summary by alphaXiv.