MODEL CARD:

Model Details

• SuperAnimal-Quadruped model(s) developed by the M.W.Mathis Lab in 2023, trained to predict quadruped pose from images. Please see Shaokai Ye et al. 2023 for details.

• The there are three main models (and several auxiliary model checkpoints):

• pose_model.pth is an HRNet-w32 compatable for DLC3.0+ Pytorch code, trained on our Quadruped-80K dataset.
• detector.pt is a Faster R-CNN that can be used as a detector for top-down detection.
• hrnet_w32_quadruped80k.pth is an HRNet-w32 trained with mmpose on our Quadruped-80K dataset.

• Full training details can be found in Ye et al. 2023. You can use the pose_model and detector simply with our light-weight loading package called DLCLibrary. Here is an example useage:

from pathlib import Path
from dlclibrary import download_huggingface_model

# Creates a folder and downloads the model to it
model_dir = Path("./superanimal_quadruped_model_pytorch")
model_dir.mkdir()
download_huggingface_model("superanimal_quadruped_pytorch", model_dir)

Intended Use

• Intended to be used for pose estimation of quadruped images taken from side-view. The model serves a better starting point than ImageNet weights in downstream datasets such as AP-10K.

• Intended for academic and research professionals working in fields related to animal behavior, such as neuroscience and ecology.

• Not suitable as a zeros-shot model for applications that require high keypiont precision, but can be fine-tuned with minimal data to reach human-level accuracy. Also not suitable for videos that look dramatically different from those we show in the paper.

Factors

• Based on the known robustness issues of neural networks, the relevant factors include the lighting, contrast and resolution of the video frames. The present of objects might also cause false detections and erroneous keypoints. When two or more animals are extremely close, it could cause the top-down detectors to only detect only one animal, if used without further fine-tuning or with a method such as BUCTD (Zhou et al. 2023 ICCV).

Metrics

• Mean Average Precision (mAP)

Evaluation Data

• In the paper we benchmark on AP-10K, AnimalPose, Horse-10, and iRodent using a leave-one-out strategy. Here, we provide the model that has been trained on all datasets (see below), therefore it should be considered “fine-tuned" on all animal training data listed below. This model is meant for production and evaluation in downstream scientific applications.

Training Data:

It consists of being trained together on the following datasets:

• AwA-Pose Quadruped dataset, see full details at (1).
• AnimalPose See full details at (2).
• AcinoSet See full details at (3).
• Horse-30 Horse-30 dataset, benchmark task is called Horse-10; See full details at (4).
• StanfordDogs See full details at (5, 6).
• AP-10K See full details at (7).
• iRodent We utilized the iNaturalist API functions for scraping observations with the taxon ID of Suborder Myomorpha (8). The functions allowed us to filter the large amount of observations down to the ones with photos under the CC BY-NC creative license. The most common types of rodents from the collected observations are Muskrat (Ondatra zibethicus), Brown Rat (Rattus norvegicus), House Mouse (Mus musculus), Black Rat (Rattus rattus), Hispid Cotton Rat (Sigmodon hispidus), Meadow Vole (Microtus pennsylvanicus), Bank Vole (Clethrionomys glareolus), Deer Mouse (Peromyscus maniculatus), White-footed Mouse (Peromyscus leucopus), Striped Field Mouse (Apodemus agrarius). We then generated segmentation masks over target animals in the data by processing the media through an algorithm we designed that uses a Mask Region Based Convolutional Neural Networks(Mask R-CNN) (9) model with a ResNet-50-FPN backbone (10), pretrained on the COCO datasets (11). The processed 443 images were then manually labeled with both pose annotations and segmentation masks. iRodent data is banked at https://zenodo.org/record/8250392. APT-36K See full details at (12).

Here is an image with the keypoint guide:

Please note that each dataset was labeled by separate labs & separate individuals, therefore while we map names to a unified pose vocabulary (found here: https://github.com/AdaptiveMotorControlLab/modelzoo-figures), there will be annotator bias in keypoint placement (See the Supplementary Note on annotator bias). You will also note the dataset is highly diverse across species, but collectively has more representation of domesticated animals like dogs, cats, horses, and cattle. We recommend if performance is not as good as you need it to be, first try video adaptation (see Ye et al. 2023), or fine-tune these weights with your own labeling.

Ethical Considerations

• No experimental data was collected for this model; all datasets used are cited.

Caveats and Recommendations

• The model may have reduced accuracy in scenarios with extremely varied lighting conditions or atypical animal characteristics not well-represented in the training data.

• Please note that each dataest was labeled by separate labs & separate individuals, therefore while we map names to a unified pose vocabulary, there will be annotator bias in keypoint placement (See Ye et al. 2023 for our Supplementary Note on annotator bias). You will also note the dataset is highly diverse across species, but collectively has more representation of domesticated animals like dogs, cats, horses, and cattle. We recommend if performance is not as good as you need it to be, first try video adaptation (see Ye et al. 2023), or fine-tune these weights with your own labeling.

License

Modified MIT.

Copyright 2023 by Mackenzie Mathis, Shaokai Ye, and contributors.

Permission is hereby granted to you (hereafter "LICENSEE") a fully-paid, non-exclusive, and non-transferable license for academic, non-commercial purposes only (hereafter “LICENSE”) to use the "MODEL" weights (hereafter "MODEL"), subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software:

This software may not be used to harm any animal deliberately.

LICENSEE acknowledges that the MODEL is a research tool. THE MODEL IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE MODEL OR THE USE OR OTHER DEALINGS IN THE MODEL.

If this license is not appropriate for your application, please contact Prof. Mackenzie W. Mathis (mackenzie@post.harvard.edu) and/or the TTO office at EPFL (tto@epfl.ch) for a commercial use license.

Please cite Ye et al if you use this model in your work https://arxiv.org/abs/2203.07436v2.

References

1. Prianka Banik, Lin Li, and Xishuang Dong. A novel dataset for keypoint detection of quadruped animals from images. ArXiv, abs/2108.13958, 2021
2. Jinkun Cao, Hongyang Tang, Haoshu Fang, Xiaoyong Shen, Cewu Lu, and Yu-Wing Tai. Cross-domain adaptation for animal pose estimation. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9497–9506, 2019.
3. Daniel Joska, Liam Clark, Naoya Muramatsu, Ricardo Jericevich, Fred Nicolls, Alexander Mathis, Mackenzie W. Mathis, and Amir Patel. Acinoset: A 3d pose estimation dataset and baseline models for cheetahs in the wild. 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13901–13908, 2021.
4. Alexander Mathis, Thomas Biasi, Steffen Schneider, Mert Yuksekgonul, Byron Rogers, Matthias Bethge, and Mackenzie W Mathis. Pretraining boosts out-of-domain robustness for pose estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1859–1868, 2021.
5. Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition, Colorado Springs, CO, June 2011.
6. Benjamin Biggs, Thomas Roddick, Andrew Fitzgibbon, and Roberto Cipolla. Creatures great and smal: Recovering the shape and motion of animals from video. In Asian Conference on Computer Vision, pages 3–19. Springer, 2018.
7. Hang Yu, Yufei Xu, Jing Zhang, Wei Zhao, Ziyu Guan, and Dacheng Tao. Ap-10k: A benchmark for animal pose estimation in the wild. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021.
8. iNaturalist. OGBIF Occurrence Download. https://doi.org/10.15468/dl.p7nbxt. iNaturalist, July 2020
9. Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961–2969, 2017.
10. Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection, 2016.
11. Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll’ar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. CoRR, abs/1405.0312, 2014
12. Yuxiang Yang, Junjie Yang, Yufei Xu, Jing Zhang, Long Lan, and Dacheng Tao. Apt-36k: A large-scale benchmark for animal pose estimation and tracking. Advances in Neural Information Processing Systems, 35:17301–17313, 2022