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

Arxiv 2025

Hao Yu^1,2* Haotong Lin^1* Jiawei Wang^1* Jiaxin Li¹ Yida Wang² Xueyang Zhang² Yue Wang¹

Xiaowei Zhou¹ Ruizhen Hu³ Sida Peng^1†

¹Zhejiang University ²Li Auto ³Shenzhen University

^*Equal contribution ^†Corresponding author

Paper

Code

TL;DR

This paper introduces InfiniDepth, a novel monocular depth estimation method with neural implicit fields. It leverages continuous coordinate queries to enable arbitrary resolutions and fine-grained depth estimation.
We design a depth query strategy along with a Gaussian decoder to improve novel view synthesis quality, especially under large viewpoint shifts.
We curate Synth4K, a high-quality 4K benchmark for evaluating depth estimation methods at high resolution and fine geometric details.

Abstract

Existing depth estimation methods are fundamentally limited to predicting depth on discrete image grids. Such representations restrict their scalability to arbitrary output resolutions and hinder the geometric detail recovery. This paper introduces InfiniDepth, which represents depth as neural implicit fields. Through a simple yet effective local implicit decoder, we can query depth at continuous 2D coordinates, enabling arbitrary-resolution and fine-grained depth estimation. To better assess our method's capabilities, we curate a high-quality 4K synthetic benchmark from five different games, spanning diverse scenes with rich geometric and appearance details. Experiments demonstrate that InfiniDepth achieves SOTA performance on both synthetic and real-world benchmarks across relative and metric depth estimation tasks, particularly excelling in fine-detail regions. It also benefits the task of novel view synthesis under large viewpoint shifts, producing high-quality results with fewer holes and artifacts.

Method

Pipeline of InfiniDepth:

Feature Query: given an input image and a continuous query 2D coordinate, we extract feature tokens from multiple layers of the ViT encoder, and query local features for the coordinate at each scale through bilinear interpolation.
Depth Decoding: given the multi-scale local features queried at the continuous coordinate, we hierarchically fuse features from high spatial resolution to low spatial resolution, and decode the fused feature to the depth value through a MLP head.

Qualitative Visualization

Interactive Depth Map Visulization: Hover over the RGB image (left-side) to explore details in the 8K-resolution depth map (right-side). Use the mouse scroll wheel to zoom in/out on a specific patch.

Loading scene...

Zoom: 1.0x

Point Cloud Visualization: Point clouds from predicted depth maps. Use the mouse to zoom and rotate and use "Alt + mouse" to pan.

Novel View Synthesis (NVS): Single-view NVS results. The rendered video demonstrates the view synthesis quality under large viewpoint shifts.

Qualitative Comparison

Predicted depth maps from different methods on the same input. Blue and pink boxes highlight regions with fine-grained geometric details.

Predicted point clouds from different methods on the same input. Orange boxes highlight regions with fine-grained geometric details.

Single-View NVS results under large viewpoint changes, e.g. Bird's-eye (BEV) views.

Acknowledgement

We thank Yuanhong Yu, Gangwei Xu and Haoyu Guo for their insightful discussions and valuable suggestions, and Zhen Xu for his dedicated efforts in curating the synthetic data.

Citation

@article{yu2026infinidepth,
    title={InfiniDepth: Arbitrary-Resolution and Fine-Grained Depth Estimation with Neural Implicit Fields},
    author={Hao Yu, Haotong Lin, Jiawei Wang, Jiaxin Li, Yida Wang, Xueyang Zhang, Yue Wang, Xiaowei Zhou, Ruizhen Hu and Sida Peng},
    booktitle={arXiv preprint},
    year={2026}
}