Large Depth Completion Model from Sparse Observations

Zhu Yu^1,* Zhengyi Zhao² Runmin Zhang¹ Lingteng Qiu² Kejie Qiu²

Yisheng He² Siyu Zhu³ Zilong Dong^2,‡ Si-Yuan Cao^4,5,6,‡ Hui-Liang Shen¹

¹Zhejiang University ²Tongyi Lab, Alibaba Group ³Fudan University ⁴Ningbo Innovation Center, Zhejiang University ⁵NingboTech University ⁶Jinhua Institute of Zhejiang University

^*Internship at Tongyi Lab ^‡Corresponding author

ICLR 2026

Paper Code

LDCM teaser showing input images, sparse depth priors, completed depth maps, and point maps. — LDCM is a simple and effective model for depth completion. Without complex architectural designs, LDCM achieves state-of-the-art performance in zero-shot depth completion and metric point map estimation.

Abstract

This work presents the Large Depth Completion Model (LDCM), a simple, effective, and robust framework for single-view metric depth estimation with sparse observations. Without relying on complex architectural designs, LDCM generates metric-accurate dense depth maps using a transformer. It outperforms existing approaches across diverse datasets and sparse observations.

LDCM leverages existing monocular foundation models to improve sparse depth inputs and reformulates training objectives to better capture geometric structure and metric consistency. A Poisson-based depth initialization strategy generates a uniform coarse dense depth map from diverse sparse observations, while a point map head regresses per-pixel 3D coordinates in camera space. This design enables direct learning of 3D scene structure, eliminates the need for camera intrinsic parameters, and produces metric-scaled 3D point maps.

BibTeX

@inproceedings{yu2026ldcm,
  title={Large Depth Completion Model from Sparse Observations},
  author={Yu, Zhu and Zhao, Zhengyi and Zhang, Runmin and Qiu, Lingteng and Qiu, Kejie and
          He, Yisheng and Zhu, Siyu and Dong, Zilong and Cao, Si-Yuan and Shen, Hui-Liang},
  booktitle={International Conference on Learning Representations},
  year={2026}
}