Who Left the Dogs Out?

Evaluation and demo code for our ECCV 2020 paper: Who Left the Dogs Out? 3D Animal Reconstruction with Expectation Maximization in the Loop.

Install

Clone the repository with submodules:

git clone --recurse-submodules https://github.com/benjiebob/WLDO

For segmentation decoding, install pycocotools python -m pip install "git+https://github.com/philferriere/cocoapi.git#subdirectory=PythonAPI"

Datasets

To use the StanfordExtra dataset, you will need to download the .json file via the repository.

Please ensure you have StanfordExtra_v12 installed, which we released 1 Feb 2021.

You may also wish to evaluate the Animal Pose Dataset. If so, download all of the dog images into data/animal_pose/images. For example, an image path should look like: data/animal_pose/images/2007_000063.jpg. We have reformatted the annotation file and enclose it in this repository data/animal_pose/animal_pose_data.json.

Splits

The train/validation/test splits used for our ECCV 2020 submission are contained in the data/StanfordExtra_v12 repository and under the data/animal_pose folder.

Pretrained model

Please download our pretrained model and place underneath data/pretrained/3501_00034_betas_v4.pth.

Quickstart

Eval

To evaluate the performance of the model on the StanfordExtra dataset, run eval.py:

cd wldo_regressor
python eval.py --dataset stanford

You can also run on the animal_pose dataset

python eval.py --dataset animal_pose

Results

| Dataset | IOU | PCK @ 0.15 |

Avg Legs Tail Ears Face
StanfordExtra 74.2 78.8 76.4 63.9 78.1 92.1
Animal Pose 67.5 67.6 60.4 62.7 86.0 86.7

Note that we have recently updated the tables in the arxiv version of our paper to account for some fixed dataset annotations and to use an improved version of the PCK metric. More details can be found in the paper.

Demo

To run the model on a series of images, place the images in a directory, and call the script demo.py. To see an example of this working, run demo.py and it will use the images in example_imgs:

cd wldo_regressor
python demo.py

Related Work

This repository owes a great deal to the following works and authors:

  • SMALify; Biggs et al. provided an energy minimization framework for fitting to animal video/images. A version of this was used as a baseline in this paper.
  • SMAL; Zuffi et al. designed the SMAL deformable quadruped template model and have provided me with wonderful advice/guidance throughout my PhD journey.
  • SMALST; Zuffi et al. provided PyTorch implementations of the SMAL skinning functions which have been used here.
  • SMPLify; Bogo et al. provided the basis for our original ChumPY implementation.

Acknowledgements

If you make use of this code, please cite the following paper:

@inproceedings{biggs2020wldo,
  title={{W}ho left the dogs out?: {3D} animal reconstruction with expectation maximization in the loop},
  author={Biggs, Benjamin and Boyne, Oliver and Charles, James and Fitzgibbon, Andrew and Cipolla, Roberto},
  booktitle={ECCV},
  year={2020}
}

GitHub

https://github.com/benjiebob/WLDO