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
P

pytorch-3d-point-cloud-generation

> 编程语言
开源

Pytorch 代码,用于从单个 RGB 图像构建 3D 点云模型。

392 stars0 点赞0 次浏览
访问官网GitHub

工具介绍

Pytorch 代码,用于从单个 RGB 图像构建 3D 点云模型。

Pytorch: Learning Efficient Point Cloud Generation for Dense 3D Object Reconstruction

A Pytorch implementation of the paper: Learning Efficient Point Cloud Generation for Dense 3D Object Reconstruction

  • Article: https://medium.com/@lkhphuc/create-3d-model-from-a-single-2d-image-in-pytorch-917aca00bb07
  • Original Website: https://chenhsuanlin.bitbucket.io/3D-point-cloud-generation
  • Original TF implementation: https://github.com/chenhsuanlin/3D-point-cloud-generation

Training/evaluating the network

Prerequisites

This code is developed with Python3 (python3). Pytorch 0.4+ is required.

Dataset

(Provided in TF's repo) The dataset (8.8GB) can be downloaded by running the command

bash
wget https://cmu.box.com/shared/static/s4lkm5ej7sh4px72vesr17b1gxam4hgy.gz

This file includes:

  • Train/test split files (from Perspective Transformer Nets)
  • Input RGB images (from Perspective Transformer Nets)
  • Pre-rendered depth images for training
  • Ground-truth point clouds of the test split (densified to 100K points)

After downloading, run tar -zxf s4lkm5ej7sh4px72vesr17b1gxam4hgy.gz under the main directory. The files will be extracted to the data directory. (Please also cite the relevant papers if you plan to use this dataset package.)

Running the code

The following scripts gives examples for running the code.

  • Pretraining the network: scripts/train-stg1.sh
  • Fine-tuning with joint 2D optimization: scripts/train-stg2.sh
  • Evaluating on the test set: scripts/evaluate.sh
  • Computing the error metrics: scripts/evaluate_dist.sh

Checkpoints are stored in models/${EXPERIMENTATION}, summaries are stored in runs/_${EXPERIMENTATION}, and evaluated point clouds are stored in results_${GROUP}. The list of optional arguments can be found by executing python3 train-stg1.py --help.


Rendering ground-truth depth images

(Provided in TF's repo) We provide the code to render depth images for supervision.

Prerequisites

This code requires the following:

  • Blender as the rendering engine. This code was developed with Blender 2.78. After installation, please make sure the command blender is callable (use which blender to check installation).
  • The OpenEXR Python binding for .exr to .mat file conversion.

Dataset

The raw ShapeNet dataset can be downloaded here. This rendering code was developed to use ShapeNetCore v2. (The provided depth images were rendered from ShapeNetCore v1.)

Running the code

Under render, run ./run.sh 03001627 8 to render depth images for fixed and arbitrary viewpoints, and convert them to .mat files. This will convert all objects in the ShapeNet chair category (03001627) with 8 fixed viewpoints. The rendered files will be stored in the output directory.


Creating densified point clouds of CAD models for evaluation

(Provided in TF's repo) We also provide the code to densify the vertices of CAD models to a specified number. This code can be run independently; only the ShapeNet dataset is required. It repeats the process of adding a vertex to the center of the longest edge of the triangular mesh and subsequently re-triangulating the mesh. This will create (generally) uniformly densified CAD models.

Running the code

Under densify, run ./run.sh 03001627 to run densification. The densified CAD models will be stored in the output directory.

Issues· 0 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

Python3d-deep-learningdeep-learning

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