Reproduce “Neuromorphic Monocular Depth Estimation with Uncertainty Modeling” 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
Viktor Bergkvist, Felix Rydell, Per-Erik Forssén, David Gustafsson, Johan Rideg
Research from the Swedish Defence Research Agency introduces methods for per-pixel uncertainty quantification in monocular depth estimation using event cameras, comparing Gaussian, log-normal, and evidential learning models with various event representations. The proposed models, particularly log-normal with a 10-bin voxel grid, demonstrated that uncertainty estimates can identify highly reliable depth predictions, achieving a 4.8-fold RMSE reduction for the most confident 10% of pixels.
Key insights
- Uncertainty estimates effectively identify reliable depth predictions, leading to substantial error reductions when focusing on highly confident pixels.
- The choice of event representation has a modest impact on raw depth accuracy, with intermediate temporal resolutions (e.g., 10-bin voxel grid) often outperforming higher resolutions due to representational complexity.
- Different uncertainty models exhibit distinct behaviors in correlating uncertainty with error; log-normal achieved the lowest AUSE, while evidential learning provided more reliable ranking in high-uncertainty regions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.