Learning Implicit Functions for Topology-Varying Dense 3D Shape Correspondence

Advances in Neural Information Processing Systems (NeurIPS 2020). Oral presentation. [Arxiv, PDF, Supp, Project]

Feng Liu, Xiaoming Liu

Department of Computer Science and Engineering, Michigan State University

The goal of this paper is to learn dense 3D shape correspondence for topology-varying objects in an unsupervised manner.

Figure 1: Given a shape S, PointNet E is used to extract the shape feature code z. Then a part embedding o is produced via a deep implicit function f. We implement dense correspondence through an inverse function mapping from o to recover the 3D shape. (b) To further make the learned part embedding consistent across all the shapes, we randomly select two shapes S_A and S_B. By swapping the part embedding vectors, a cross reconstruction loss is used to enforce the inverse function to recover to each other.

---

This code is developed with Python3 and PyTorch 1.1

Dataset

Please refer to data/README

Pretrained models

Please refer to models/README

Evaluation

Please refer to evaluation/README

Training

Our method is trained in three stages: 1) PointNet like encoder and implicit function are trained on sampled point-value pairs via occupancy loss. 2) encoder, implicit function and inverse implicit function are jointly trained via occupancy and self-reconstruction losses. 3) We jointly train encoder, implicit function and inverse implicit function with occupancy, self-reconstruction and cross-reconstruction losses.

## ----------------------  chair  ----------------------------- ##
# stage1
CUDA_VISIBLE_DEVICES=0 python train_stage1.py --cate_name chair --epoch  50 --resolution 16 --batch_size 45 
CUDA_VISIBLE_DEVICES=0 python train_stage1.py --cate_name chair --epoch 100 --resolution 32 --batch_size 25 --pretrain_model TRUE --pretrain_model_name Corr-49.pth
CUDA_VISIBLE_DEVICES=0 python train_stage1.py --cate_name chair --epoch 300 --resolution 64 --batch_size  7 --pretrain_model TRUE --pretrain_model_name Corr-99.pth

# stage2
CUDA_VISIBLE_DEVICES=0 python train_stage2.py --cate_name chair --epoch 500 --resolution 64 --batch_size  4 --pretrain_model TRUE --pretrain_model_name stage1/Corr-299.pth

# stage3
CUDA_VISIBLE_DEVICES=0 python train_stage3.py --cate_name chair --epoch 1000 --resolution 64 --batch_size  2 --pretrain_model TRUE --pretrain_model_name stage2/Corr-499.pth

---

Contact

For questions feel free to post here or directly contact the author via liufeng6@msu.edu