百科.dev
全部条目在线工具AI 编程趋势榜开源项目技术资讯提交条目
登录
< 返回工具列表
D

Dens3R

> 编程语言
开源

[ICLR2026] "Dens3R: 一种用于 3D 几何体预测的基础模型"的官方实现

395 stars0 点赞0 次浏览
访问官网GitHub

工具介绍

[ICLR2026] "Dens3R: 一种用于 3D 几何体预测的基础模型"的官方实现

Dens3R

[Paper] [Project Page] [Taobao3D]

Dens3R: A Foundation Model for 3D Geometry Prediction,
Xianze Fang1*, Jingnan Gao2*, Zhe Wang1, Zhuo Chen2, Xingyu Ren2, Jiangjing Lyu1†, Qiaomu Ren1, Zhonglei Yang1, Xiaokang Yang2, Yichao Yan2‡, Chengfei Lyu

1 Alibaba Group 2 Shanghai Jiao Tong University

*Equal contribution † Project leader ‡ Corresponding author

Official implementation of "Dens3R: A Foundation Model for 3D Geometry Prediction".

This repository is now a branch of the Taobao3D repository. Please check the Taobao3D repo for further releases. We will also release frontier 3D research projects there, so feel free to check it out and leave a star!

TODO

  • Checkpoint release
  • AutoRecon Application release
  • Evaluation Code release
  • Inference Code release

Installation

1. Clone & Install Dependencies of Dens3R

1.1 Prepare Python Environment

Note: We use python 3.10.13 and CUDA 11.8. Other versions of cuda and python should be compatible. If your CUDA version is not 11.8, please replace the lines below with proper versions of torch and torchvision in before installing. We do not require specific torch and torchvision versions so other torch versions should be compatible.

…

1.2 Install COLMAP[Required]

We require COLMAP for reconstruction. Please follow the official guidance to install COLMAP.

1.3 Install GLOMAP[Optional]

GLOMAP is optional in our project. GLOMAP can speedup the reconstruction process compared with using COLMAP. Please follow the official Repo to install GLOMAP. Note that COLMAP is required for installing GLOMAP.

2. [Optional] Install Dependencies of AutoRecon

Note: We modified the official AutoRecon to support automated SDF reconstruction of an object. The modified version of AutoRecon is contained in our repo. We delete all the submodule of the official AutoRecon so the '--recursive' command is not neccessary.

2.1 Install Dependencies of AutoRecon

Install tiny-cuda-nn

bash
pip install git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch

Install faiss following the official guide

bash
# for example
conda install -c conda-forge faiss-gpu

Install AutoRecon

bash
cd AutoReconForDens3R
pip install -e .

Usage

Infer normal/depth of an image sequence

bash
# example
python infer/infer_normal_pts3d.py \
  --model_path path/to/model/checkpoint \
  --img_dir path/to/images/folder \
  --save_dir path/to/save/folder \
  --pair_mode self \
  --size 512

[Arguments Explanation - click to expand]

  • model_path: The path to the Dens3R model checkpoint.
  • --img_dir: The path to the images folder.
  • --save_dir: The path to the folder to save the results.
  • --pair_mode: Choose in [self, near, sequence]. self: an image paired with itself. near: an image paired with its nearby image. sequence: an images paired with more than 3 images in order to get averaged results.

Reconstruct pose and pointcloud of an image sequence

bash
# example
python infer/dens3r_recon.py \
  --img_dir path/to/images/folder \
  --output_dir path/to/save/results \
  --model_path path/to/model/checkpoint

[Arguments Explanation - click to expand]

  • --img_dir: path to images folder.
  • --output_dir: path to save results.
  • -model_path: path to Dens3R model checkpoint.
  • --triangulate_from_pose: path to the pre-computed COLMAP-format pose of the images. Default is None. If set, the dens3r_recon.py will not estimate pose and use the pre-computed pose to triangulate instead.
  • --matching_mode: {exhaustive,sequential} exhaustive or sequential pairwise matching mode. Default is exhaustive.
  • --window_size: If matching_mode is sequential, window_size is the window size for sequential matching. Default is 30.
  • --size: Infer resolution. Default 512. Set to 1024 or higher for better normal and depth estimation.
  • --use_mapper: {colmap,glomap}.
  • --camera_model: Camera model define in COLMAP, default is SIMPLE_RADIAL.
  • --conf_thr: Confidence threshold for matching, default is 1.001. Suggested unchanged.
  • --pixel_tol: Pixel tolerance for matching, default is 5. Suggested unchanged.
  • --subsample: Pixel subsample factor for matching, default is 8. Suggested 4 or 8.

Automated 3D Reconstruction of an image sequence Using AutoRecon

Please refer to the shell script in AutoRecon_Dens3R/exps/code-release/run_dens3r_recon.sh.

The process contains 3 steps:

  1. Reconstruct the poses and the pointcloud of the image sequence. Also predict the normalmaps of the images for step 2 usage.
  2. Reconstruct the SDF using AutoRecon (Based on NeuS).
  3. Extract the mesh from SDF using Marching Cubes.

Acknowledgements

Our work builds upon several fantastic open-source projects. We'd like to express our gratitude to the authors of:

  • DUSt3R
  • MASt3R

Citation

bibtex
@inproceedings{dens3r,
    title={Dens3R: A Foundation Model for 3D Geometry Prediction}, 
    author={Xianze Fang and Jingnan Gao and Zhe Wang and Zhuo Chen and Xingyu Ren and Jiangjing Lyu and Qiaomu Ren and Zhonglei Yang and Xiaokang Yang and Yichao Yan and Chengfei Lyu},
    booktitle={The Fourteenth International Conference on Learning Representations},
    year={2026}
}

Issues· 4 开放

查看全部 Issues在 GitHub 打开

暂无开放 Issues,或尚未同步最近议题。

> 标签

Python

暂无评论,来聊聊你的看法吧

> 工具信息

发布日期2026年8月1日
最后更新2026年9月17日
分类编程语言
定价开源

> 相关工具

T
TypeScript
JavaScript 的超集,为前端与全栈提供静态类型
P
Python
通用编程语言,广泛用于 Web、数据与 AI
G
Go
Google 推出的简洁高效系统语言