• an extensible proof-of-concept
• a free, open source cow weight estimation iOS app which tries to help small time farmers.
• our research found that current automated cattle weight prediction implementations are either:
  • expensive and subscription based
  • free and malfunctioning/not of acceptable quality / don't use lidar and real life distances
  • government funded but obsolete/abandoned/not publicly available
    • note: most research on this topic often involves complex algorithms and data pipelines
    • the current scope of the projected and our limited skillset has been a limiting factor in taking advantage of this, and our predictions are not as accurate as state of the art approaches, but can be improved over time

CowDar requires an IOS device equipped with LiDAR capability, such as:

• iPhone 12 Pro / Pro Max
• iPhone 13 Pro / Pro Max
• iPhone 14 Pro / Pro Max
• iPhone 15 Pro / Pro Max
• iPhone 16 Pro / Pro Max
• iPhone 17 Pro / Pro Max
• iPad Pro (2020+)

• simulatenously extracts a 3D LiDAR point cloud, a photo image, and phone intrinsic data (helps to calculate real life distances) from compatible iOS devices with LIDAR hardware
• sends that data to a fastapi python server
  • the photo image is passed into a custom trained yolov8m-pose model with 100 epochs using a dataset of labelled cow poses created by Sorin Workspace of 1042 total images (729 training, 209 validation, 104 testing) to analyse points of interest on the cow (approx 30 min on Nvidia T4)
    • this model outputs a set of normalised coordinates (normalised refers to a relative x,y value range of 0-1, as opposed to pixel measurements)
    • special note: point prediction is particularly difficult on angus cows due to their extremely dark skin colour, and we are currently experimenting with the model training and image pre processing to improve this
  • the 3D LiDAR point cloud is mathematically projected onto a 2D image plane
  • the server matches the normalised coordinates to the closest projected 3D LiDAR points to determine (estimate) the physical distance between points of interest on the cow
  • 3D euclidean distances are extracted between these coordinates (e.g., body length and radius)
  • weight is predicted using a variant of Schaeffer's Formula

• POST /api/tim/dist (with debug logs and debug image i/o)
  • req: multi form, image jpg file, ply ply file
  • res: json: json { "success" : boolean, "predictedWeight" : number, "distPoint2To10" : number, "distMidpointTo3" : number }

• Dataset used to train the cow pose model
  • The YOLO cow pose model was trained on a relatively small dataset (1042 labelled images). A larger dataset would likely result in more accurate and consistent keypoint detection.
• The formula used to estimate the mass of the cow (Schaeffer's formula) is a simplified model which does not take into account various factors such as breed, age, sex etc.
• LiDAR sensor accuracy
  • The system relies on the mobile device's LiDAR sensor to estimate real-world distances. Consumer grade LiDAR sensors can produce noisy or incomplete point clouds depending on lighting conditions, distance from the subject and the colour and reflectivity of the animal's coat. During testing, the data primarily consisted of Angus cattle, whose dark coats can reduce LiDAR reflectivity and make depth perception less reliable in some cases.
• Evaluation scope
  • The system has only been tested on small number of real-world scans and has not been evaluated under a large and diverse dataset (e.g. different cattle breeds). Future work will involve testing across a wider range of breeds, environments and scanning conditions to better evaluate the accuracy and robustness of the system.