Reproduce “MAMVI: 3D Test-Time Adaptation via Masked Multi-View Point Clouds” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Inseok Kong, Geunyoung Jung, Jiyoung Jung

Researchers at the University of Seoul developed MAMVI, an efficient Test-Time Adaptation framework for 3D point cloud classification under distribution shifts. This method utilizes masked multi-view consensus and unified loss aggregation to achieve robust performance with significantly faster inference speeds compared to prior multi-view TTA approaches.

Key insights

  • Aggregating losses from multiple masked views into a single backward pass dramatically improves inference efficiency for 3D TTA, addressing the latency issues of sequential multi-view adaptation.
  • A hybrid masking strategy, which combines fixed and stochastic masking ratios, effectively generates diverse yet stable partial views crucial for robust adaptation.
  • Dynamically adjusting the learning rate based on the model's multi-view consensus confidence improves adaptation stability and prevents over-adaptation to uncertain individual samples.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.