Skip to content

Latest commit

 

History

History
104 lines (79 loc) · 3.71 KB

README.md

File metadata and controls

104 lines (79 loc) · 3.71 KB
Raw

PatchRD

Code for ECCV 2022 paper PatchRD: Detail-Preserving Shape Completion by Learning Patch Retrieval and Deformation. (PDF)

Installation

Install Pytorch. It is required that you have access to GPUs. The code is tested with Ubuntu 18.04, Pytorch v1.9.0 and CUDA 10.2.

Install ChamferDistancePytorch by following the instructions in their github repo.

Install psbody for mesh reading and writing.

Install the following Python dependencies (with pip install):

numpy
scipy
h5py
sklearn
skimage
opencv-python
kornia
pytorch=1.9.0

Data Download

You can download the dataset here. It's a voxelized dataset for 8 classes in ShapeNet. Train/Test spliting is in ./splits.
To evaluate the chamfer distance, we convert our output to point cloud and use the point cloud with 16384 points from Completion3D dataset as the ground truth. You can either download it from the offcial website or directly here.

Demo with Pre-trained Models

You can download the pre-trained models for the chair category here.

To run the completion pipeline, first uncomment --dump_deform, --mode test, --small_dataset in patch_train.sh and deform_train.sh, then uncomment --small_dataset in joint_test.sh.

Then run:

./patch_train.sh

./deform_train.sh

./joint_test.sh

You will get the input, output meshes for a small set of samples in the samples_joint folder.

Training

There are four steps to train our framework. The default settings train model for the chair category. You can change the arguments --data_content and --data_dir in each bash script to train on other categories.

Step 1: Coarse Completion

This stage takes a partial shape as input, and output the coarse full shape (4x downsampled shape of the detailed full shape).

./complete_train.sh

Step 2: Retrieval Learning

First train the patch encoder to learn the feature embeddings for the coarse and detailed patch pair.

./patch_train.sh

Then dump the intermediate retrieval results to train the deformation. You have to add --dump_deform in patch_train.sh, then run

./patch_train.sh

Step 3: Initial Deformation Learning

This step learns the initial deformation for the jointly learning stage. Run

./deform_train.sh

Then dump the intermediate initial deformation results to train the deformation and blending stage. You have to add --dump_deform in deform_train.sh, then run

./deform_train.sh

Step 4: Deformation and Blending

This step joinly learn the deformation and blending. Run

./joint_train.sh

Evaluation

After training all four networks, you can run the following scripts to run the completion results from a randomly cropped shape. You can find the input and output mesh in ./samples_joint.

./joint_test.sh

Citation

If you find our work useful in your research, please consider citing:

@inproceedings{sun2022patchrd,
    author = {Bo, Sun and Kim, Vladimir(Vova) and Huang, Qixing and Aigerman, Noam and Chaudhuri, Siddhartha},
    title = {PatchRD: Detail-Preserving Shape Completion by Learning Patch Retrieval and Deformation},
    booktitle = {Proceedings of the IEEE European Conference on Computer Vision},
    year = {2022}
}