Minhyuk Sung, Hao Su, Vladimir G. Kim, Siddhartha Chaudhuri, and Leonidas Guibas
Siggraph Asia 2017
[Project] | [arXiv]
@article{Sung:2017,
author = {Sung, Minhyuk and Su, Hao, and Kim, Vladimir G. and Chaudhuri, Siddhartha
and Guibas, Leonidas},
title = {Complement{Me}: Weakly-Supervised Component Suggestions for 3D Modeling},
Journal = {ACM Transactions on Graphics (Proc. of SIGGRAPH Asia)},
year = {2017}
}
ComplementMe is a neural network framework for suggesting complementary components and their placement for an incomplete 3D part assembly. The component retrieval is performed by two neural networks called embedding and retrieval networks; the first indexes parts by mapping them to a low-dimensional feature space, and the second maps partial assemblies to appropriate complements. These two networks are jointly trained on unlabeled data obtained from public online repositories without relying on consistent part segmentations or labels. The retrieval network predicts a probability distribution over the space of part embeddings to deal with ambiguities of the multiple complementary components. The placement is performed by a separate network called placement network, which predicts a coordinates of the newly added component.
The ShapeNet model component and semantic part data are available on our project website.
- Numpy (tested with ver. 1.13.1)
- TensorFlow (tested with ver. 1.0.1)
The files in utils are directly brought from the PointNet.
This code is released under the MIT License. Refer to LICENSE for details.
- Script files reproducing results in the paper.