Unsupervised Domain Adaptation for 3D Keypoint Estimation via View Consistency

This repository is the PyTorch implementation for the network presented in:

Xingyi Zhou, Arjun Karpur, Chuang Gan, Linjie Luo, Qixing Huang, Unsupervised Domain Adaptation for 3D Keypoint Estimation via View Consistency ECCV 2018(arXiv:1712.05765)

Contact: [email protected]

Requirements

• cudnn
• PyTorch
• Python with h5py, opencv and progress
• Optional: tensorboard

Data

• The following datasets are used in this repo. If you use the data provided, please also consider citing them:
  • ModelNet
  • ShapeNet and keypoint annotation provided in SyncSpecCNN.
  • Redwood Dataset
• Download the pre-processing data and annotations here, and un-zip them on data.

Testing

• Download our pre-trained model on Redwood Depth dataset and move it to models.
• Run the test.

 python main.py -expID demo -loadModel ../models/Redwood.pth.tar -test

• Visualize the results.

python tools/vis.py ../exp/Chair/demo/img_valTarget ../exp/Chair/demo/valTarget.txt

Training

• Stage1: Train the source model.

python main.py -expID Source -epochs 120 -dropLR 90

Our results of this stage is provided here.

• Stage2: Adapt to the target domain with shape consistency loss.

python main.py -expID Redwood -targetDataset Redwood -targetRatio 1 -shapeWeight 1 -loadModel ../models/ModelNet120.tar -LR 0.01