Reproduce “SAND: Spatially Adaptive Network Depth for Fast Sampling of Neural Implicit Surfaces” 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
Chuanxiang Yang, Junhui Hou, Yuan Liu, Siyu Ren, Guangshun Wei, Taku Komura, Yuanfeng Zhou, Wenping Wang
The SAND framework introduces a spatially adaptive network depth mechanism for neural implicit surfaces, dynamically adjusting computational effort based on a query point's proximity to the surface and local geometric complexity. This approach substantially accelerates inference speed, with SAND-SIREN achieving total inference times of 0.310s (Stanford) and 0.344s (Thingi10K) compared to SIREN's ~12s, while preserving or enhancing geometric reconstruction accuracy.
Key insights
- Computational effort for neural implicit surfaces can be optimized by adaptively varying network depth based on a query point's spatial location relative to the surface and its local geometric complexity.
- A Multi-Layer Perceptron can be structured with cumulative residual outputs at each layer, enabling early termination and focused learning of progressively finer details.
- An octree-based depth map provides an efficient mechanism to pre-determine and store the necessary network evaluation depth for different spatial regions, avoiding dynamic runtime overhead.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.