Skip to content


Folders and files

Last commit message
Last commit date

Latest commit



13 Commits

Repository files navigation


This repository contains the TensorFlow implementation for the following paper

Pixel2Mesh: Generating 3D Mesh Models from Single RGB Images (ECCV2018)

Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, Yu-Gang Jiang.


If you use this code for your research, please consider citing:

  title={Pixel2Mesh: Generating 3D Mesh Models from Single RGB Images},
  author={Nanyang Wang and Yinda Zhang and Zhuwen Li and Yanwei Fu and Wei Liu and Yu-Gang Jiang},

Try it on Colab

Installing all the dependencies might be tricky and you need a computer with a CUDA enabled GPU. To get started fast you can just try this demo developed by Mathias Gatti using Google Colab.

Open In Colab

Project Page

The project page is available at



Our code has been tested with Python 2.7, TensorFlow 1.3.0, TFLearn 0.3.2, CUDA 8.0 on Ubuntu 14.04.


  • Nov. 8, we update the script for generate auxiliary data.

Running the demo

git clone
cd Data/

Download the pre-trained model and unzip to the Data/ folder.


Reconstructing shapes

python --image Data/examples/plane.png

Run the demo code and the output mesh file is saved in Data/examples/plane.obj

Input image, output mesh


If you use CD and EMD for training or evaluation, we have included the cuda implementations of Fan et. al. in external/

cd Pixel2Mesh/external/

Modify the first 3 lines of the makefile to point to your nvcc, cudalib and tensorflow library.



We used the ShapeNet dataset for 3D models, and rendered views from 3D-R2N2:
When using the provided data make sure to respect the shapenet license.

Below is the complete set of training data. Download it into the Data/ folder.

The training/testing split can be found in Data/train_list.txt and Data/test_list.txt

Each .dat file in the provided data contain:

  • The sampled point cloud (with vertex normal) from ShapeNet. We transformed it to corresponding coordinates in camera coordinate based on camera parameters from the Rendering Dataset.

Input image, ground truth point cloud.



You can change the training data, learning rate and other parameters by editing

The total number of training epoch is 30; the learning rate is initialized as 3e-5 and drops to 1e-5 after 25 epochs.


The evaluation code was released, please refer to for more details.

Notice that the 3D shape are downscaled by a factor of 0.57 to generate rendering. As result, all the numbers shown in experiments used 0.57xRaw Shape for evaluation. This scale may be related to the render proccess, we used the rendering data from 3DR2N2 paper, and this scale was there since then for reason that we don't know.


This software is for research purpose only.
Please contact us for the licence of commercial purposes. All rights are preserved.


Nanyang Wang (nywang16 AT

Yinda Zhang (yindaz AT

Zhuwen Li (lzhuwen AT

Yanwei Fu (yanweifu AT

Yu-Gang Jiang (ygj AT


Apache License version 2.0