InfiniDepth: Arbitrary-Resolution and Fine-Grained Depth Estimation with Neural Implicit Fields

2 min read

InfiniDepth represents depth as a neural implicit field: a local implicit decoder can be queried at any continuous 2D coordinate, so the output resolution is not tied to the input grid. The authors also curate a 4K synthetic benchmark from five video games to measure fine detail. The README lists three uses: single-image depth at any resolution, single-image Gaussian splats for view synthesis, and a “depth sensor augmentation” mode that takes RGB plus sensor depth and returns dense metric depth.

Findings

  • The abstract reports the best results on synthetic and real benchmarks for both relative and metric depth, with the largest gains in fine-detail regions, and fewer holes when rendering new views under large viewpoint shifts. It gives no headline number.
  • The README shows it uses a MoGe-2 checkpoint to recover metric scale when exporting point clouds, so its metric output leans on MoGe-2.

Limits

  • I read the abstract and README only.
  • The fine-detail benchmark is rendered game footage. Nothing in what I read is underwater or in scattering media.

Relevance

Thin structures (cables, branches, railings) are where grid-based depth models smear, and obstacle avoidance needs them. The sensor-augmentation mode is the same idea as Pow3R-SLAM: treat sparse measured depth as a prompt. See foundation models for robotics.