Conceptual
Browse Instruct Curate

Lightweight Tile-Based High-Resolution Metric Depth Estimation with Guided Denoising

A fast, parameter-efficient approach to estimating metric depth from a single high-resolution image: a small encoder refines patch-level predictions fused with a downsampled global prediction, and guided denoising units use coarse depth features to clean the lightweight encoder's noisy high-resolution features, trained with a gradient-matching loss for synthetic-to-real transfer.