Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?


Failed to load latest commit information.
Latest commit message
Commit time
February 6, 2018 16:41
February 13, 2018 18:34
February 7, 2018 14:18
February 6, 2018 16:03
February 6, 2018 15:55
March 3, 2018 13:31
March 19, 2019 13:47
February 6, 2018 15:55
February 6, 2018 15:55
February 6, 2018 15:55
February 6, 2018 15:55
February 6, 2018 15:55
February 6, 2018 15:55
February 6, 2018 15:55
February 6, 2018 15:55

Learning Efficient Point Cloud Generation for Dense 3D Object Reconstruction

Chen-Hsuan Lin, Chen Kong, and Simon Lucey
AAAI Conference on Artificial Intelligence (AAAI), 2018

Project page:
arXiv preprint:

We provide TensorFlow code for the single-category experiment (for ShapeNet chairs).

Training/evaluating the network


This code is developed with Python3 (python3). TensorFlow r1.0+ is required. The dependencies can install by running

pip3 install --upgrade numpy scipy termcolor tensorflow-gpu

If you don't have sudo access, add the --user flag.


The dataset (8.8GB) can be downloaded by running the command


This file includes:

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/
  • Fine-tuning with joint 2D optimization: scripts/
  • Evaluating on the test set: scripts/
  • Computing the error metrics: scripts/

Checkpoints are stored in models_${GROUP}, summaries are stored in summary_${GROUP}, and evaluated point clouds are stored in results_${GROUP}.
The list of optional arguments can be found by executing python3 --help. The default training settings in this released code is slightly different from the paper but optimizes the networks better.
We provide two different network architectures: (1) originally from the paper (2) deeper with residual blocks. Reference performances on the test set is as follows (note that different runs will result in slightly different performances):

pred→GT GT→pred
original 1.7342 1.8371
ResNet 1.6723 1.8169

Rendering ground-truth depth images

We provide the code to render depth images for supervision.


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.


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 ./ 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

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

If you find our code useful for your research, please cite

  title={Learning Efficient Point Cloud Generation for Dense 3D Object Reconstruction},
  author={Lin, Chen-Hsuan and Kong, Chen and Lucey, Simon},
  booktitle={AAAI Conference on Artificial Intelligence ({AAAI})},

Please contact me ( if you have any questions!


Learning Efficient Point Cloud Generation for Dense 3D Object Reconstruction 🍇 (AAAI 2018)







No releases published


No packages published