Shape and Viewpoints without Keypoints

Shubham Goel, Angjoo Kanazawa, Jitendra Malik

University of California, Berkeley In ECCV, 2020

Project Page

Requirements

• Python 3.7
• Pytorch 1.1.0
• Pymesh
• SoftRas
• NMR

Installation

Please use this Dockerfile to build your environment. For convenience, we provide a pre-built docker image at shubhamgoel/birds. If interested in a non-docker build, please follow docs/installation.md

Training

Please see docs/training.md

Demo

1. From the ucmr directory, download the pretrained models (cub_train_cam4_withcam.tar.gz) from this gdrive folder and extract it:

tar -vzxf cub_train_cam4_withcam.tar.gz

You should see cachedir/snapshots/cam/e400_cub_train_cam4

2. Run the demo:

python -m src.demo \
    --pred_pose \
    --pretrained_network_path=cachedir/snapshots/cam/e400_cub_train_cam4/pred_net_600.pth \
    --shape_path=cachedir/template_shape/bird_template.npy\
    --img_path demo_data/birdie1.png

Evaluation

To evaluate camera poses errors on the entire test dataset, first download the CUB dataset and annotation files as instructed in docs/training.md. Then run

python -m src.experiments.benchmark \
        --pred_pose \
        --pretrained_network_path=cachedir/snapshots/cam/e400_cub_train_cam4/pred_net_600.pth \
        --shape_path=cachedir/template_shape/bird_template.npy \
        --nodataloader_computeMaskDt \
        --split=test

Citation

If you use this code for your research, please consider citing:

@inProceedings{ucmrGoel20,
  title={Shape and Viewpoints without Keypoints},
  author = {Shubham Goel and
  Angjoo Kanazawa and
  and Jitendra Malik},
  booktitle={ECCV},
  year={2020}
}

Acknowledgements

Parts of this code were borrowed from CMR and CSM.