ShinyAnimalCV

Open-source cloud-based web application for object detection, segmentation, and three-dimensional visualization of animals using computer vision

Table of Contents

• Introduction
• Section 1: Video demos for online usage of ShinyAnimalCV
  • 1.1: Object detection and segmentation
  • 1.2: Three-dimensional morphological feature extraction and visualization
• Section 2: Training computer vision models with custom data
  • Subsection 2.1: Image annotation
  • Subsection 2.2: Model training and evaluation
• Section 3: Local inference with trained models in ShinyAnimalCV
• Contact information and help
• References
• License

Introduction

Computer vision (CV), a non-intrusive and cost-effective technology, has facilitated the development of precision livestock farming by enabling optimized decision-making through timely and individualized animal care. However, despite the availability of various CV tools in the public domain, applying these tools to animal data can be complex, often requiring users to have programming and data analysis skills, as well as access to computing resources such as CPUs and GPUs. Moreover, the rapid expansion of precision livestock farming is creating a growing need to educate and train animal science students about CV. Thus, we developed a user-friendly online web application named ShinyAnimalCV for object detection and segmentation, three-dimensional visualization, as well as 2D and 3D morphological feature extraction using animal image data.

This documentation aims to offer users comprehensive and step-by-step instructions for utilizing ShinyAnimalCV, both online and locally. The first section provides users with video demonstrations on how to utilize ShinyAnimalCV online. Users can learn how to perform tasks such as object detection, segmentation, 3D morphological feature extraction, and visualization. For users who wish to train ShinyAnimalCV with their own data, the second section offers detailed instructions on training computer vision models using custom data. These guidelines walk users through the entire process, from data preparation to model training, ensuring that they can effectively train the models with their data. The third section focuses on incorporating the trained computer vision models into ShinyAnimalCV for local inference. By following the provided instructions, users will be able to seamlessly integrate their trained models into ShinyAnimalCV, enabling them to perform powerful and efficient inferences locally.

Section 1: Video demo for online usage of ShinyAnimalCV

We provide two video demos below to showcase the usage of two modules of ShinyAnimalCV online: object detection and segmentation, as well as 3D morphological feature extraction. Additionally, we offer example images for users to explore these modules.

1.1 Object detection and segmentation. [Video demo]

1.2 Three-dimensional morphological feature extraction and visualization. [Video demo]

Section 2: Training computer vision models with custom data

2.1 Image annotation

The first step in training computer vision models involves labeling or annotating objects in images. Here, we introduce LabelMe, which is an open-source graphical image annotation tool that allows users to draw polygons, rectangles, circles, lines and points to label the objects in images. When training the models using the data annotated from LabelMe, one of the inputs is an info.yaml file. However, it is important to note that the current version of LabelMe on GitHub does not directly return a info.yaml file. To overcome this, the following tutorials in this section will guide you through the process of downloading LabelMe from GitHub, making the necessary modifications to enable the generation of a info.yaml file, and building the modified standalone executable. For more details about building a standalone executable LabelMe, please refer to the file Install-StandAlone-labelme.txt and LabelMe GitHub Repo.

2.1.1 Create a conda environment with Python installed

# create a conda env named `labelme`
conda create --name labelme python=3.9
# activate the created `labelme`
conda activate labelme

2.1.2 Install two python packages

conda install pyqt  
conda install -c conda-forge pyside2

2.1.3 Download LabelMe from GibHub to local using git

# download the LabelMe from GitHub to local
git clone https://github.com/wkentaro/labelme.git
# change directory to the downloaed labelme/ folder 
cd labelme/

2.1.4 Modify the json_to_dataset.py file to enable the generation of a info.yaml file

# add this line at the beginning of `json_to_dataset.py`
import yaml 

# add the following codes before the line `logger.info("Saved to: {}".format(out_dir))` in `json_to_dataset.py`
logger.warning('info.yaml is being replaced by label_names.txt')
info = dict(label_names=label_names)
with open(osp.join(out_dir, 'info.yaml'), 'w') as f:
	yaml.safe_dump(info, f, default_flow_style=False)

2.1.5 Build the standalone executable LabelMe

pip install .                 
pip install pyinstaller
pyinstaller labelme.spec

2.1.6 Check the installed LabelMe version

labelme --version

2.1.7 Label images using the standalone executable LabelMe

More details about how to label images can be found on LabelMe GitHub Repo.

2.1.8 Covert the .json file returned from the previous step to dataset

labelme_json_to_dataset image_name.json -o image_name_json

The following five files will be found under the folder image_name_json/:

2.2 Model training and evaluation

This section provides a step-by-step guide to training a computer vision model for object detection and instance segmentation using a modified state-of-the-art computer vision package, Mask R-CNN. Specifically, we have made modifications to the original Mask R-CNN to ensure compatibility with tensorflow (version 2.9.0 and 2.10.0), as the current GitHub version only supports tensorflow version <= 2.5.0. Additionally, we have replaced the original axis-aligned rectangle bounding box from Mask R-CNN with a minimum rotated rectangle bounding box using OpenCV. This enhancement allows for the extraction of more precise morphological features. The modified Mask R-CNN is available under mrcnn/.

2.2.1 Dataset preparation

This subsection details the process of preparing the training, validation, and testing sets for model training and evaluation. First, users need to partition the entire dataset into training, validation, and testing sets using a specified ratio, such as 7:2:1. Additionally, the label.png file within each image_name_json folder, obtained from LabelMe, needs to be transferred to a new mask folder named cv2_mask and renamed as image_name.png. To facilitate this procedure, we have included a script called generate_cv2_mask.ipynb, which can be used iteratively to prepare the cv2_mask folder. For more comprehensive details and explanations, please refer to the generate_cv2_mask.ipynb file. Furthermore, we have provided a demo dataset, which serves as a reference for dataset preparation.

2.2.2 Train and evaluate Mask R-CNN model

In this subsection, we show how to train and evaluate the model with organized annotated custom data using our provided Jypyter Notebook (MaskRCNN.ipynb).

• Create a conda environment named maskrcnn and install all dependencies for running the MaskRCNN.ipynb. The required dependencies file of environment.yml can be downloaded here.

# create a conda environment `maskrcnn` and install all dependencies using `environment.yml`
conda env create -f environment.yml

• Open the MaskRCNN.ipynb using Jupyter Notebook and follow the detailed instructions in MaskRCNN.ipynb to train and evaluate the model.

# activate `maskrcnn` (if deactivated)
conda activate maskrcnn
# open `MaskRCNN.ipynb` 
jupyter notebook MaskRCNN.ipynb

Section 3: Local inference with trained models in ShinyAnimalCV

To enable users to perform local inference using the trained model on custom data while leveraging their local computational resources, this section provides detailed instructions on how to deploy and run ShinyAnimalCV locally with custom model weights. The source code for ShinyAnimalCV can be accessed at ShinyAnimalCV_SourceCode. The nine pre-trained models included in online ShinyAnimalCV can be accessed on Zenodo.

3.1 Download ShinyAnimalCV in local

# download this github repository in local using git
# the source code of ShinyAnimalCV is under ShinyAnimalCV/ShinyAnimalCV_SourceCode/
git clone https://github.com/uf-aiaos/ShinyAnimalCV.git

3.2 Set up running environment

ShinyAnimalCV is built through the integration of R and Python using the R package reticulate. In the server.R file, we include the command reticulate::use_condaenv(condaenv="maskrcnn", conda="auto", required=TRUE) to ensure interoperability between R and Python. If you have not created the maskrcnn environment yet, please refer to 2.2.2 Train and evaluate Mask R-CNN model for instructions on creating the maskrcnn Python environment.

3.3 Add custom trained weight into ShinyAnimalCV

To use the custom trained model weight for local inference using ShinyAnimalCV, users need to move custom weight file (.h5 file) obtained from Section 2 under ShinyAnimalCV/ShinyAnimalCV_SourceCode/logsand rename it as CustomWeights.h5.

3.4 Deploy ShinyAnimalCV and perform inferences locally using the custom trained weight

The ShinyAnimalCV application can be deployed by running the following command within RStudio. Alternatively, users can open the server.R file with RStudio and click on the "Run App" button located at the top-right corner of the RStudio interface to deploy the application.

shiny::runApp('path_to_the_folder_ShinyAnimalCV_SourceCode')

Once the ShinyAnimalCV application is successfully deployed, a web-based graphical user interface will open, allowing users to interact with the application. In the interface, users can select their custom trained weights by choosing the option CustomWeights under "Step 2: Select a pretrained machine learning model." This enables users to perform inference using their custom trained weights.

Contact information and help

• Jin Wang ([email protected])
• Haipeng Yu ([email protected])

References

If you use the materials provided in this repository, we kindly ask you to cite our paper:

• Jin Wang, Yu Hu, Lirong Xiang, Gota Morota, Samantha A. Brooks, Carissa L. Wickens, Emily K. Miller-Cushon and Haipeng Yu. Technical note: ShinyAnimalCV: open-source cloud-based web application for object detection, segmentation, and three-dimensional visualization of animals using computer vision. Journal of Animal Science. doi: 10.1093/jas/skad416

License

This project is primarily licensed under the GNU General Public License version 3 (GPLv3).