The Problem Every Dairy Farmer Knows
Lameness is widespread on dairy farms, costing the industry millions of dollars every year. Lameness causes pain to dairy cows, but it is difficult to devise effective approaches to prevention and treatment without reliable data on when cows first become lame.
Several companies have tried to develop automatic lameness detection systems, but few seem to work effectively on real farms. An important bottleneck to developing automatic detection systems is the availability of reliable data that can be used to ‘train’ the computer vision models. To make better models, we need a large collection of accurately labeled videos for these models to learn from.
The development of this data has been limited because current cattle lameness assessment methods have major flaws:
- Unreliable and inconsistent results: Even trained and experienced lameness assessors often disagree when scoring the same cow. The same assessor will often give different scores when watching the same cow (or even the same video) when scoring on different days
- Time-consuming and labour intensive: Scoring cows requires assessors to visit farms and manually assess each cow
- Identifying early cases: By the time lameness is obvious, the cow has already endured a period of pain, recovery takes longer, and cases become longer or impossible to treat.
This inconsistency in manual assessment creates a negative-feedback loop. Automatic detection systems depend on having good training data. If the manual assessment methods used to create the training data are inconsistent and unreliable, the automated systems will be flawed too.
To help the dairy industry detect lameness reliably and early, we need to rethink how we assess lameness. First, we need a manual method that allows us to detect lameness in a consistent manner, and then we can use this improved method to create reliable training data needed for effective automation.
A New Approach: Lameness Hierarchy
Researchers at UBC Animal Welfare Program have developed a completely different way to assess lameness. Instead of trying to score each cow individually against some standard scale, we simply present assessors videos of two cows walking side by side and ask: “Which cow is more lame?” The key idea behind the project is that comparison makes everything more obvious.
It is much more intuitive for people to compare two things side by side than to give an absolute score for one object. UBC research has now shown that pairwise lameness assessments are indeed more reliable than traditional absolute gait scoring of individual cows. This work also showed that pairwise assessments can be accurately performed by “crowd workers” recruited from online platforms where people with no special training can complete simple tasks.
How It Works
We tested this new method using videos of 30 cows walking. Here is what we did:
- Traditional scoring: 5 expert lameness assessors scored all 30 cows using a traditional 5-level gait scoring method, with each assessor evaluating each cow 3 times (15 total scoring rounds)
- Created video pairs: We paired up every cow with every other cow, creating 435 different pairwise comparisons
- Pairwise lameness assessment: For each video pair, assessors chose which cow looked more lame and by how much (Figure 1)
- Multiple assessments: 4 expert assessors and on average 18 untrained crowd workers evaluated each video pair
- Created a lameness hierarchy: We used the Elo-rating method (similar to that used to rank players in chess and on line video games) to rank all cows from the most healthy to the most lame
The Results
The new method outperformed traditional scoring:
The lameness hierarchy constructed from pairwise lameness assessments achieved high agreement among different experienced assessors (intraclass correlation coefficient (ICC) = 0.81; the closer the ICC is to 1 the higher the agreement). In comparison, we found that different assessors gave different scores to each cow when using the traditional, subjective gait scoring method (ICC = 0.44 ± 0.02). Even when the same expert used traditional scoring to score the same cow on different days, score did not always agree (ICC = 0.62 ± 0.09).
Traditional scoring method identified only 3 out of 30 cows as clinically lame (gait score ≥ 3 on 5-level scale), with average gait scores ranging from 2.0 to 3.9. The new ranking method detected more subtle differences in lameness among different cows. The new hierarchy method still correlated well with traditional scores (Spearman rs = 0.77, P < 0.01; Figure 2), but provided finer detail.
The lameness hierarchy created by untrained online workers reached high agreement with that constructed by experienced assessors (ICC = 0.85; Figure 2). And secondary analysis showed that only a few crowd workers were needed to create reliable rankings.
What This Means for Dairy Farms
Potential benefits on our new approach include lower costs, more consistent data. faster processing and earlier detection of lameness. Specifically, individual assessments by online workers is much cheaper than hiring trained assessors, and this method results in reliable rankings of lameness. Crowd workers can also evaluate large numbers of cows quickly. Lameness hierarchy method provides finer detail, which should promote earlier identification of lame cows.
In new work we hope to develop this method at a larger scale, ultimately leading to the creation of a large, reliable dataset ideal for training automatic detection systems. Automated systems built on this foundation have the potential to catch lameness with greater granularity and will facilitate regular monitoring to make routine lameness assessment feasible.
While our method still needs to be scaled up and tested on larger groups of cows from different farms, it represents a promising step toward solving one of the dairy industry’s most persistent challenges.
Next Steps
Our team is now working on streamlining the video collection process to automatically gather, sort, and identify cows based on video footage from multiple farms. This will help us rapidly expand the video database for the next phase of development.
For further information please contact Dan Weary (danweary@mail.ubc.ca), Marina (Nina) von Keyserlingk (nina@mail.ubc.ca), or Kehan (Sky) Sheng (skysheng@mail.ubc.ca). The results described in this article are based on the study Kehan Sheng, Borbala Foris, Marina A.G. von Keyserlingk, Tiffany-Anne Timbers, Varinia Cabrera, Daniel M. Weary, (2025). Redefining lameness assessment: Constructing lameness hierarchy using crowd-sourced data. Computers and Electronics in Agriculture, 234, 110206. https://doi.org/10.1016/j.compag.2025.110206. This study was done in collaboration with UBC Department of Statistics, Faculty of Science.
This research was supported by a Natural Science and Engineering Research Council (NSERC) Discovery grant (RGPIN-2023-04350). General funding for the UBC Animal Welfare Program is provided by the NSERC Industrial Research Chair in Dairy Cattle Welfare together with our industrial partners: the Dairy Farmers of Canada (Ottawa, ON, Canada), Saputo Inc. (Montreal, QC, Canada), British Columbia Dairy Association (Burnaby, BC Canada), Alberta Milk (Edmonton, AB, Canada), Intervet Canada Corporation (Kirkland, QC, Canada), Boehringer Ingelheim Animal Health (Burlington, ON, Canada), BC Cattle Industry Development Fund (Kamloops, BC, Canada), The Semex Alliance (Guelph, ON, Canada), Lactanet (Sainte-Anne-de-Bellevue, QC, Canada), Dairy Farmers of Manitoba (Winnipeg, MB, Canada), and the Saskatchewan Milk Marketing Board (Regina, SK, Canada). MvK, DMW and TAT also received funding from the UBC Land and Food system Internal Research Grant.
Research Reports
Research Reports are published throughout the year by UBC’s Dairy Education and Research Centre (DERC), a centre affiliated with the Faculty of Land and Food Systems, to share applied aspects of research from published articles in refereed scientific journals. The Dairy Education and Research Centre is used by several research groups on campus including Animal Reproduction and Animal Welfare and Behaviour. Other groups interested in conducting research at the Centre are encouraged to contact the UBC DERC Manager of Research, Health and Animal Welfare – Dr. Julia Lomb (Julia.Lomb@ubc.ca)
