Learning Fine-Grained Geometry for Sparse-View Splatting via Cascade Depth Loss

TL;DR

This paper introduces a hierarchical depth supervision framework with a cascade loss to improve 3D reconstruction quality from sparse views, enhancing geometric accuracy and detail fidelity.

Contribution

The paper proposes HDGS, a novel multi-scale depth supervision method with a cascade Pearson correlation loss, to better leverage depth priors in sparse-view neural rendering.

Findings

Achieves state-of-the-art results on LLFF and DTU datasets under sparse-view conditions.

Improves structural fidelity and detail preservation compared to existing methods.

Effectively enforces multi-scale depth consistency to enhance geometric accuracy.

Abstract

Novel view synthesis is a fundamental task in 3D computer vision that aims to reconstruct photorealistic images from novel viewpoints given a set of posed images. However, reconstruction quality degrades sharply under sparse-view conditions due to insufficient geometric cues. Existing methods, including Neural Radiance Fields (NeRF) and more recent 3D Gaussian Splatting (3DGS), often exhibit blurred details and structural artifacts when trained from sparse observations. Recent works have identified rendered depth quality as a key factor in mitigating these artifacts, as it directly affects geometric accuracy and view consistency. However, effectively leveraging depth under sparse views remains challenging. Depth priors can be noisy or misaligned with rendered geometry, and single-scale supervision often fails to capture both global structure and fine details. To address these challenges, we introduce Hierarchical Depth-Guided Splatting (HDGS), a depth supervision framework that progressively refines geometry from coarse to fine levels. Central to HDGS is our novel Cascade Pearson Correlation Loss (CPCL), which enforces consistency between rendered and estimated depth priors across multiple spatial scales. By enforcing multi-scale depth consistency, our method improves structural fidelity in sparse-view reconstruction. Experiments on LLFF and DTU demonstrate state-of-the-art performance under sparse-view settings.