Reproduce “LaS-Comp: Zero-shot 3D Completion with Latent-Spatial Consistency” 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

Weilong Yan, Haipeng Li, Hao Xu, Nianjin Ye, Yihao Ai, Shuaicheng Liu, Jingyu Hu

The LaS-Comp framework introduces a zero-shot, training-free approach for 3D shape completion that addresses the latent-spatial consistency challenge in pre-trained 3D foundation models. It achieves high-fidelity and geometrically coherent completions across diverse object categories and partiality patterns, demonstrating improvements of up to 49.6% in Chamfer Distance on a new challenging benchmark.

Key insights

  • Successfully bridging the 'latent-spatial gap' in 3D foundation models requires a combined approach of explicit spatial geometry injection and implicit latent space alignment to effectively utilize their powerful generative priors.
  • A partial-aware noise schedule, which differentiates between known and unknown regions, is crucial for maintaining fidelity in observed areas while enabling diverse and plausible generation in missing parts.
  • Iterative refinement of latent features, guided by spatial consistency losses, enables high-quality, category-agnostic 3D completion without requiring any task-specific training data.

Summary by alphaXiv.

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