SSR refines monocular geometry with sparse volumetric updates and sparse 3D U-Nets

zhenjun_zhao · x · 2026-07-21

## Method overview The paper proposes **Self-Guided Sparse Volumetric Refinement (SSR)** for fine-detail monocular geometry estimation. ## How it works - A base model first predicts an initial point map and 2D features. - The SSR module iteratively refines geometry by discretizing current estimates into sparse voxel shells. - A Sparse 3D U-Net then uses multi-scale sparse 3D convolutions plus injected 2D features to predict log-depth residuals. ## Reported outcome The figures claim improved zero-shot performance across several benchmarks, with particularly strong gains in local, boundary, and global depth/point metrics, and better recovery of thin structures and complex regions.

Original post →

More from Research

Research channel →