Insights on back marking for the automated identification of animals

arXiv:2603.25535 · cs.CV, cs.LG · Submitted 2026-03-26 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Insights on back marking for the automated identification of animals".

Jane: Here's the gist of the paper. The gist This study provides insights on effective back mark design for automated identification of pigs by analyzing a machine learning model trained to distinguish pigs via their unique back marks, highlighting that designs must account for motion blur, diverse view angles, occlusions, and common data augmentation strategies.

Tom: <ref:2603.25535#pg2> Experimental Setup and Back Marks The data was collected during an observational study on social behaviour in pigs to help behaviour <ref:2603.25535#pg3>.

Jane: The setup comprised two identical pens observed by identical cameras mounted in an elevated side-view angle <ref:2603.25535#pg2>.

Tom: To support individual identification, the pigs received regular back marks using symbols instead of numbers for practicality and reproducibility <ref:2603.25535#pg2>.

Paper summary: Tom: So, looking at the title and the authors for "Insights on back marking for the automated identification of animals," it seems they are giving us a blueprint for practical application in livestock monitoring >

Jane: It’s about moving beyond just slapping random symbols on pigs and thinking about how those symbols need to be robust against real-world noise when an AI is trying to identify them >

Lu: The implication is that future work needs to focus on designing mark sets that are inherently resistant to the specific types of errors—motion blur, angle changes—that we see in animal behavior data >

Meng: For us shipping models, this means when we design our training data or our input pipeline, we have to explicitly account for those kinds of visual corruptions so the model doesn't fail at scale because of a specific viewing condition >

Lalam: I'm pretty sure the core message is that back mark design has important choices that directly impact an AI model’s ability to successfully differentiate a set of marks >

Tom: Yeah, it’s about matching the physical design of the identifier with what we know about how animals move and how data gets messed up in production >

Jane: So for someone just listening, this paper shows that if you want an AI to reliably identify individual pigs from back marks, you need to make sure those marks are designed to survive motion blur, different angles, and being partially hidden when the animal is moving around >

Conclusion: Tom: So we’ve been looking at how this study actually works, and now we need to talk about what they actually found in their conclusion for this paper, "Insights on back marking for the automated identification of animals."

Jane: Yeah, so basically, what does it mean when they wrap up all that training and testing? What's the big picture here?

Tom: They show that designing those physical back marks isn't just about picking any symbols; it’s a very specific set of choices. They use ten different marks to test how well an AI, specifically a ResNet-fifty neural network, can actually tell them apart.

Lu: The core message is that the design has to anticipate what happens in the real world. They stress that the marks have to stay clear even when there's motion blur or weird angles from which you’re viewing the pig.

Meng: From a deployment standpoint, that means any system we build needs to account for those visual distortions because if it doesn't, the model just fails in production when things get a little messy.

Jane: So, what does this actually change for someone who just listens to the show? It’s telling us that back mark design is a critical part of the AI problem, not just something you can ignore.

Lalam: I'm pretty sure their main conclusion is that if you want reliable individual identification with machine learning, you have to align the physical mark with how data gets corrupted by animal movement and camera angles.

Tom: Exactly. It’s a design constraint for the whole system. It moves it from just being a computer vision problem to being a hardware and behavioral problem combined.

Jane: So, if we take that idea—designing marks to survive those specific visual failures—what kind of real-world applications do you think this opens up next?

Lalam: It paves the way for better, individual-level monitoring in future studies because the data quality starts at the design stage rather than just hoping the AI can fix blurry images later.

D. Brunner, M. Bordes, E. Mayrhuber, S. M. Winkler, V. Dorfer, M. Oczak

Bioinformatics Research Group, PLFDoc, University of Applied Sciences Upper Austria · Computer Vision Lab, TU Wien, Vienna, Austria · Centre for Animal Nutrition and Welfare, The University of Veterinary Medicine Vienna · Precision Livestock Farming Hub, The University of Veterinary Medicine Vienna

cs.CV, cs.LG

Submitted: 2026-03-26

Updated: 2026-10-21

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 77/100

The gist: The gist This study provides insights on effective back mark design for automated identification of pigs by analyzing a machine learning model trained to distinguish pigs via their unique back marks,

Key concepts

Back Marks
These are specific patterns or symbols placed on an animal's back used for individual identification. The study tested various simple shapes like dots, lines, and letters to see which ones work best for automated recognition by computer vision models.
ResNet-50 Neural Network
This is a type of deep learning model used in the study to classify pigs based on their back marks. It is a sophisticated algorithm designed to recognize complex patterns in images, helping the system distinguish between different pigs with high accuracy.
Data Augmentation
These are techniques used during model training to artificially increase the size and diversity of a dataset by applying transformations like flipping, rotation, brightness changes, and cropping. The study found these methods are crucial for making the model robust against real-world variations.
Motion Blur/Occlusion
These refer to real-world issues where an animal's movement blurs the image or when other animals block the view of a mark. The research showed that back mark designs must be specifically chosen to remain recognizable even when these physical obstructions occur.

Terminology

Summary

The gist

This study provides insights on effective back mark design for automated identification of pigs by analyzing a machine learning model trained to distinguish pigs via their unique back marks, highlighting that designs must account for motion blur, diverse view angles, occlusions, and common data augmentation strategies. <ref:2603.25535#pg2>

Experimental Setup and Back Marks

The data was collected during an observational study on social behaviour in pigs to help behaviour <ref:2603.25535#pg3>. The setup comprised two identical pens observed by identical cameras mounted in an elevated side-view angle <ref:2603.25535#pg2>. To support individual identification, the pigs received regular back marks using symbols instead of numbers for practicality and reproducibility <ref:2603.25535#pg2>. The back marks were inspired by patterns used in mice <ref:2603.25535#pg2>. The specific back marks used included dot dot, dot line horizontal, i, line line horizontal, o, reverse t, s, v, vertical line, and x <ref:2603.25535#pg2>.

Data Preparation and Model Training

A human labeller placed bounding boxes with class IDs on all frames of 11 video clips <ref:2603.25535#pg3>. Nine of the clips were selected for the training set, and two for validation and test sets, respectively <ref:2603.25535#pg3>. The raw training set amounted to a total of 3750 frames, while the validation and test sets comprised 250 and 750 frames <ref:2603.25535#pg2>. The final training set consisted of 26,260 crops, with the validation and test sets reduced to 500 crops respectively <ref:2603.25535#pg2>. A PyTorch implementation of a neural network of type ResNet-50 was selected as the image classification model <ref:2603.25535#pg2>. Extensive data augmentation was used in training, including horizontal and vertical flipping, random rotation, brightness, contrast, saturation and hue modifications <ref:2603.25535#pg2>.

Model Performance and Error Analysis

The trained model reached a classification accuracy of 91% on the validation set and 69% on the test set <ref:2603.25535#pg2>. There were conspicuous differences across classes, such as vertical line achieving 88% versus reverse t at 38% on the test set <ref:2603.25535#pg2>. The challenge posed by reverse t was likely due to its tendency to look similar to other back marks in specific situations tied to animal behaviour <ref:2603.25535#pg2>. Three kinds of behaviour were identified as especially relevant, which can lead to motion blur, diverse view angles and occlusions, respectively <ref:2603.25535#pg2>. Specifically, motion blur can cause reverse t to look like vertical line <ref:2603.25535#pg2>. A certain view angle can cause reverse t to look more like v or x, which is reflected in both the model confidence and the pixel importance <ref:2603.25535#pg2>. Even minor occlusions can make it impossible for the model to distinguish between line line horizontal and o <ref:2603.25535#pg2>.

Design Recommendations

Back mark design must control for predictable animal behaviour such as fast movement, diverse poses and proximity to other animals <ref:2603.25535#pg2>. Furthermore, the back mark design should consider common data augmentation strategies applied during model training <ref:2603.25535#pg2>. This includes photometrical augmentations like colour, brightness and contrast adjustments as well as geometrical augmentations like cropping and flipping <ref:2603.25535#pg2>. Colour jitter and grayscale modifications might interfere with coloured back marks <ref:2603.25535#pg2>. Crop augmentation can make it impossible to distinguish reverse t from vertical line, and line line horizontal from o <ref:2603.25535#pg2>. Back marks that are mirror images of each other must be avoided in the presence of flip augmentation because an instance of class s looks like class 2 after a horizontal flip <ref:2603.25535#pg2>.

Conclusion

The study demonstrated that back mark design has important choices impacting a ML model’s ability to successfully differentiate a set of back marks <ref:2603.25535#pg2>. The design should control for predictable animal behaviour and consider common data augmentation strategies to ensure data compatibility with a wide range of ML methods <ref:2603.25535#pg2>.

Acknowledgements

This research was funded in whole or in part by the Austrian Science Fund (FWF) [https://doi.org/10.55776/DFH34] <ref:2603.25535#pg2>. The data used in this study originates from the Let me out project, funded by the Austrian Science Fund (FWF) [https://doi.org/10.55776/I6488] <ref:2603.25535#pg2>.

References

[1] Caroline Clouard, Auriane Foreau, Sébastien Goumon, Céline Tallet, Elodie Merlot, and Rémi Resmond. Evidence of stable preferential affiliative relationships in the domestic pig. Animal Behaviour, 213:95–105, 2024 <ref:2603.25535#pg2>.

[2] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016 <ref:2603.25535#pg2>.

[3] Dong Liu, Maciej Oczak, Kristina Maschat, Johannes Baumgartner, Bernadette Pletzer, Dongjian He, and Tomas Norton. A computer vision-based method for spatial-temporal action recognition of tail-biting behaviour in group-housed pigs. Biosystems Engineering, 195:27–41, 2020 <ref:2603.25535#pg2>.

[4] Shay Ohayon, Ofer Avni, Adam L Taylor, Pietro Perona, and SE Roian Egnor. Automated multi-day tracking of marked mice for the analysis of social behaviour. Journal of neuroscience methods, 219(1):10–19, 2013 <ref:2603.25535#pg2>.

[5] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In European conference on computer vision, pages 818–833. Springer, 2014 <ref:2603.25535#pg2>.

--- Page 1 ---

Insights on back marking for the automated identification of animals∗

D. Brunner1,2,†

, M. Bordes3

, E. Mayrhuber1,2

, S. M. Winkler1

, V. Dorfer1 and M. Oczak3,4 <ref:2603.25535#pg2>. 1Bioinformatics Research Group <ref:2603.25535#pg2>. 2Computer Vision Lab, TU Wien, Vienna, Austria <ref:2603.25535#pg2>. 3Centre for Animal Nutrition and Welfare, The University of Veterinary Medicine Vienna, Vienna, Austria <ref:2603.25535#pg2>. 4Precision Livestock Farming Hub, The University of Veterinary Medicine Vienna, Vienna, Austria <ref:2603.25535#pg2>. †david.brunner@fh-hagenberg.at <ref:2603.25535#pg2>.

Abstract

To date, there is little research on how to design back marks to best support individual-level monitoring of uniform looking species like pigs <ref:2603.25535#pg2>. With the recent surge of machine learning-based monitoring solutions, there is a particular need for guidelines on the design of marks that can be effectively recognised by such algorithms <ref:2603.25535#pg2>. This study provides valuable insights on effective back mark design, based on the analysis of a machine learning model, trained to distinguish pigs via their back marks <ref:2603.25535#pg2>. Specifically, a neural network of type ResNet-50 was trained to classify ten pigs with unique back marks <ref:2603.25535#pg2>. The analysis of the model’s predictions highlights the significance of certain design choices, even in controlled settings <ref:2603.25535#pg2>.

Improvements for AI systems

  1. Develop a robust back mark design system that explicitly incorporates avoidance strategies for behavior-based error cases, specifically addressing motion blur, diverse view angles, and occlusions caused by animal behavior to ensure classification stability under real-world conditions. This improved AI system can perform individual-level monitoring of uniform species like pigs with higher reliability in dynamic environments.

  2. Integrate data augmentation strategies directly into the back mark design pipeline to proactively test and mitigate interference from common photometrical augmentations such as colour jitter and grayscale conversion before model training begins. This allows the system to be more resilient when deployed in varied lighting conditions or when using diverse ML architectures that employ standard augmentation techniques.

  3. Implement a design constraint layer in the back mark generation process that forbids mirror image patterns for any given class, directly addressing the issue where back marks that are mirror images of each other must be avoided in the presence of flip augmentation. This prevents catastrophic misclassification during training when horizontal flipping is applied to the input data.

Abstract

To date, there is little research on how to design back marks to best support individual-level monitoring of uniform looking species like pigs. With the recent surge of machine learning-based monitoring solutions, there is a particular need for guidelines on the design of marks that can be effectively recognised by such algorithms. This study provides valuable insights on effective back mark design, based on the analysis of a machine learning model, trained to distinguish pigs via their back marks. Specifically, a neural network of type ResNet-50 was trained to classify ten pigs with unique back marks. The analysis of the model's predictions highlights the significance of certain design choices. Most importantly, the set of back marks must be designed such that each mark remains unambiguous under conditions of motion blur, diverse view angles and occlusions, caused by animal behaviour. The generated insights can support individual-level monitoring in future studies and real-world applications by optimizing back mark design.

Related papers