Across-species pose estimation in poultry based on images using deep learning

Published on
June 2, 2022

Jan Erik Doornweerd started working as a PhD for Animal Breeding and Genomics in 2020. He recently published his first paper entitled “Across-Species Pose Estimation in Poultry Based on Images Using Deep Learning” in Frontiers of Animal Science. An important milestone, which is why we asked Jan Erik to share a bit more about his research.

Extracting walk and stance information via video

“Poultry breeders are interested in the way an animal walks and stands because it gives them information that can be used to improve the health and welfare of the animal,” Jan Erik explains. “In practice, the walkability of an animal is assessed at one point in time by a human expert. Consequently, the expert assigns a certain walking score to the animal which can then be used for selection.”

“However,” he continues, “this process requires a considerable amount of time and effort, is prone to subjectivity, and only provides information about the animal’s walk and stance at one point in time. Alternatively, video cameras can be used to collect information on walk and stance more frequently and on a larger number of animals.”

Pose-estimation network

Extracting walk and stance information from video can be done with a pose-estimation network. But how does that work exactly? Jan Erik explains: “A pose-estimation network can be trained to estimate the location of key body points. The key body points related to an animal’s walk and stance are, for example, the knees, hocks and feet. However, a pose-estimation network needs to be trained on examples of videos where a human marked the location of the knees, hocks and feet on a video frame. This process is called annotation and it is quite laborious; multiple examples are required, and each video frame requires multiple accurate marks of the location of the key body points.”

He continues: “The data from which a pose-estimation network can be trained are not always available, or at least not the large numbers needed to train an accurate pose-estimation network. However, data from another species might be useful, either by itself or combined with data from the target species, given that the combined data could reduce the amount of annotation that is required.”

Combining data

Jan Erik: “In this study we investigated the performance of animal pose-estimation networks trained on data of broilers, turkeys, or a combination of broiler and turkey data. The networks were tested within species and across species. The within-species test provided us with a baseline of the network, which was then tested across species. The network trained on the combined data was tested on both broilers and turkeys.”

Promising results for pose-estimation networks

Results of the study were promising, although Jan Erik says that further research is necessary before pose-estimation networks can be applied in practice: “Across species the networks trained on a single species did not provide a directly applicable model. The network trained on the combined data reduced the annotations per species and approached within species performance. However, even though a network trained on one species and tested on another did not result in a directly applicable model, it was still able to estimate key body point locations that seemed relatively informed. Thus, a network trained on one species could be a jumping point from which a network for a different but somewhat similar species can be trained.”

Jan Erik’s study was conducted in collaboration with Hendrix Genetics and Breed4Food.