Skip to main content
Ko Lab

Research

Francis builds computer vision systems that replace subjective, manual welfare assessment with automated measurement, working with veterinary collaborators at the University of Saskatchewan and the Prairie Swine Centre.

Publications

Computer vision system for assessing pig welfare indicators on carcasses

F. Ferri, Y. Wang, R. Ko, J. Yepez, M. Lagoda, Y. M. Seddon, S.-B. Ko. Veterinary and Animal Science, 2026.

System diagram with a video decoder, pig detector, tracker and head pose classifier.
  • Assessment at slaughter covers far more animals than on-farm inspection, which is subjective, slow and a biosecurity risk
  • A modular pipeline of YOLOv4 detection, U-Net segmentation, and colorimetric and geometric analysis scores skin lesions, tail lesions, tail length and hernias
  • 93.0% accuracy on hernias, 86.3% and 90.4% on dorsal and lateral skin lesions, 86.8% on tail lesions
  • Tail length is estimated by segmentation and curve fitting, to within 4.45 cm RMSE, with the whole system running at 30.31 frames per second
Diagram of a pig with a tail bounding box and the steps used to segment and measure tail length. Dorsal and lateral photographs of a pig carcass with the anatomical regions outlined and labelled.

Automated detection and imaging of dorsal and lateral carcass views

F. Ferri, J. Yepez, M. Ahadi, Y. Wang, R. Ko, Y. M. Seddon, S.-B. Ko. Computers and Electronics in Agriculture, 222, 109058, 2024.

  • Captures both dorsal and lateral views of each carcass as it passes the camera, so no indicator is missed to occlusion

Instance-aware coordinate attention for efficient pig pose estimation

F. Ferri, M. Mostafavi, B. Predicala, S.-B. Ko. Preprint, 2026.

Overhead camera view of pigs in a barn pen, used as input for pose estimation.
  • Occlusion, overlapping animals and varying body shapes make multi-instance pig pose estimation hard
  • A lightweight instance-aware head replaces the usual heavy backbone dependence, using Coordinate Attention and Conditional Channel Weighting
  • Accuracy comparable to CiD with 68.09% fewer FLOPs
Architecture diagram of the instance-aware coordinate attention head used for pig pose estimation.

Lightweight network for anterior mediastinum segmentation

S. Soleimani-Fard, W. G. Jeong, F. Ferri, H. Sasani, Y. Choi, S. Deiva, et al. Neural Computing and Applications, 37(27), 22875-22889, 2025.

  • A U-shaped convolutional network with an attention mechanism, kept small enough for clinical deployment