chenchao15/lpi

[ECCV 2022] Latent Partition Implicit with Surface Codes for 3D Representation

This repository contains the code to reproduce the results from the paper "Latent Partition Implicit with Surface Codes for 3D Representation".

Homepage |Paper | Supp

If you find our code or paper useful, please consider citing:

@inproceedings{LPI,
    title = {Latent Partition Implicit with Surface Codes for 3D Representation},
    author = {Chao, Chen and Yu-shen, Liu and Zhizhong, Han},
    booktitle = {European Conference on Computer Vision (ECCV)},
    year = {2022}
}

demo

Installation

We support python3, you can first create an virture environment called LPI_venv :

python -m venv LPI_venv
source LPI_venv/bin/activate

Then, to install the dependencies, run:

pip install -r requirements.txt

Next, for evaluation of the models, complie the extension modules, which are provided by Occupancy Networks. run:

python setup.py build_ext --inplace

To compile the dmc extension, you have to have a cuda enabled device set up. If you experience any errors, you can simply comment out the dmc_* dependencies in setup.py. You should then also comment out the dmc imports in im2mesh/config.py.

Finally, for calculating chamfer distance faster during training, we use the Customized TF Operator nn_distance, run:

cd nn_distance
./tf_nndistance_compile.sh

If you encounter any errors, please refer to the relevant instructions of PU-net and modify the compile scripts slightly.

Dataset

You can download our preprocessed ShapeNet dataset with the access code bycc.

• samples_*: preprocessing data of ShapeNet (sample points and ground truth points).
• centers_100_*: 100 centers sampled by fps algorithm for each object.
• class_list: list of objects in each category of ShapNet.

Put all samples_* folders, centers_100_* folders or any of the categories you want to try, and class_list in data.

Training

Training and evaluating single 3d object:

./run.sh

Training and evaluating all 3d objects of a class:

./multi_run.sh

You can change the class_name in the script to another category of ShapeNet. And note that modify the number of for loops to the number of objects of the current category.

Evaluation

Evaluating single 3d object:

./test.sh

Evaluating all 3d objects of a class:

./multi_test.sh

lpi's Issues

When will training data be released?

Hi~ Thanks for your great work!
May I know when will you release the training data and deployment instructions?
Thank you!

How can I use geodesic distance to compute the blending weights?

I find that there is only Euclidean distance is used in LPI.py. So could you please tell me how I can use geodesic distance to compute the blending weights?