hippogriff / parsenet-codebase Goto Github PK

View Code? Open in Web Editor NEW

88.0 10.0 19.0 332 KB

Code base of ParSeNet: ECCV 2020.

Shell 0.12% Python 99.88%

3d-deep-learning 3d-reconstruction shape-parsing eccv-2020

parsenet-codebase's Introduction

ParSeNet: A Parametric Surface Fitting Network for 3D Point Clouds

Authors: Gopal Sharma, Difan Liu, Evangelos Kalogerakis, Subhransu Maji, Siddhartha Chaudhuri, Radomír Měch

This repository contains codebase for the ParSeNet paper published at ECCV-2020.

Paper | Project Page

Installation

To install conda environment:

conda env create --force environment.yml -n parsenet
source activate parsenet

Dataset

To dowload the dataset, run:

bash download_dataset.sh

For data organization, please see readme_data.md.

Experiments

Experiments are done on Nvidia 1080ti gpus.

SplineNet

To train open SplineNet (with 2 gpus):

python train_open_splines.py configs/config_open_splines.yml

To test open SplineNet:

python test_open_splines.py configs/config_test_open_splines.yml

To train closed SplineNet (with 2 gpus):

python train_closed_control_points.py configs/config_closed_splines.yml

To test closed SplineNet:

python test_closed_control_points.py configs/config_test_closed_splines.yml

ParSeNet

To train ParseNet with only points as input (with 4 gpus):

python train_parsenet.py configs/config_parsenet.yml

To train ParseNet with points and normals as input (with 4 gpus):

python train_parsenet.py configs/config_parsenet_normals.yml

To train ParseNet in an end to end manner (note that you need to first pretrain the above models), then specify the path to the trained model in configs/config_parsenet_e2e.yml (with 2 gpus). Further note that, this part of the training requires dynamic amount of gpu memory because a shape can have variable number of segment and corresponding number of fitting module. Training is done using Nvidia m40 (24 Gb gpu).

python train_parsenet_e2e.py  configs/config_parsenet_e2e.yml

Testing can be done using test.py

python test.py 0 3998

Acknowledgements

This project takes inspiration of designing network architecture from the code base provided by Wang et.al.: https://github.com/WangYueFt/dgcnn
We also thank Sebastian for timely release and advice on ABC dataset: https://deep-geometry.github.io/abc-dataset/

Citation

@misc{sharma2020parsenet,
    title={ParSeNet: A Parametric Surface Fitting Network for 3D Point Clouds},
    author={Gopal Sharma and Difan Liu and Evangelos Kalogerakis and Subhransu Maji and Siddhartha Chaudhuri and Radomír Měch},
    year={2020},
    eprint={2003.12181},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

parsenet-codebase's People

Contributors

Stargazers

Watchers

Forkers

aoe-khkhan hyprivate star-cold trendingtechnology braveryang 4ndr3ar luciuspennyworth kapil-agrawal18 prabindh lavender-lee wenyusun whuhxb dalbenziog mstoreide skorzhovskyi kane-lau igormaurell anantk2006

parsenet-codebase's Issues

How to test ParSeNet in my own dataset?

Hi @Hippogriff , thanks for sharing your work and code.
Now, I 'm trying to test ParSeNet in my own dataset. I found that your network need normal, primitive and cluster_id which is directly get from h5 files. So, If I want to train and test your network in my own dataset, how can I get those data(normal, primitive and cluster_id).
Thanks in advance.

Pre-trained model

Hello,

Thanks for sharing the code.

Can you also provide the pre-trained model?

Best.
Mulin

Incorrect conda env create statement in readme

The installation guide in the readme says that to install the conda environment, the following code should be run:
conda env create --force environment.yml -n parsenet
But there is to my knowledge no "--force" flag in the conda env create command, and from the documentation of that command and from trying it out myself, --file would make much more sense. Running the code as above gave me the following error: SpecNotFound: Invalid name 'environment.yml', try the format: user/package, however running it as
conda env create --file environment.yml -n parsenet works perfectly.

No configobj

please upload the file "configobj.py".
thanks.

Issue on running test.py

When I download all the dataset and pretrained models and run 'python test.py 0 3999', the following error occur:

Expected tensor for argument #1 'input' to have the same device as tensor for argument #2 'weight'; but device 0 does not equal 1 (while checking arguments for cudnn_convolution)

Could you please give me some suggestions on how to fix this?

Generate HDF5

Thanks for making the code available. I am training a parsenet with the provided dataset. When this is ready I will test with my own data, I would like to know how you transformed the dataset from CAD format to an HDF5 representation, some specific library was used? Or some other code public available? Thanks.

Can't found the file "predictions.h5"

hello, when i run "bash download_dataset.sh" according to readme, i can't found the file "predictions.h5", I also can't download the file on the website. Please, now how do I get the dataset, thanks！

No test_parsenet.py

Thanks for your work!
It is said that Testing can be done using test_parsenet.py in Readme. But it seems that there is no test_parsenet.py

Visualization of parsenet

Hi Gopal, the meshes of parsenet results in your paper looks pretty and clean. But when I tried to visualize the 'newer_pred_mesh' in 'test.py', the mesh with trimmed boundary does not look so good, see figure below. I guess I haven't figure out the right way to extract the correct predicted mesh. Can you give me some suggestions?