Reproduce “FineRecon: Depth-aware Feed-forward Network for Detailed 3D Reconstruction” 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

Noah Stier, Anurag Ranjan, Alex Colburn, Yajie Yan, Liang Yang, Fangchang Ma, Baptiste Angles

FineRecon enhances the fidelity and coherence of feed-forward 3D reconstructions from posed images through resolution-agnostic supervision, depth guidance, and point back-projection inference. The method achieves superior geometric and depth metrics on the ScanNet dataset, showing recovery of finer details and improved surface consistency compared to previous techniques.

Key insights

  • Accurate, interpolation-free ground-truth supervision is critical for training deep networks to reconstruct fine geometric details, providing the largest individual performance gain in ablation studies.
  • Integrating explicit geometric priors, such as multi-view depth estimates fused via TSDF, effectively guides volumetric networks to localize surfaces more precisely and complements implicit feature learning.
  • Sub-voxel geometric detail can be achieved by conditioning final predictions on high-resolution 2D image features sampled at query points, circumventing the computational limits of increasing dense 3D CNN resolution.

Summary by alphaXiv.

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