Unsupervised Canine Emotion Recognition Using Momentum Contrast
Name
sensors-24-07324.pdf
Size
2.68 MB
Format
Adobe PDF
Checksum (MD5)
01c1e1acb1aa604a1b7aa3f833e87ea5
Author(s)
Bhave, Aarya
Hafner, Alina
Bhave, Anushka
Gloor, Peter A.
Date Issued
November 16, 2024
Journal
Sensors
Publisher
Multidisciplinary Digital Publishing Institute
Citation
Bhave, A.; Hafner, A.; Bhave, A.; Gloor, P.A. Unsupervised Canine Emotion Recognition Using Momentum Contrast. Sensors 2024, 24, 7324.
Version
Final published version
Abstract
We describe a system for identifying dog emotions based on dogs’ facial expressions and body posture. Towards that goal, we built a dataset with 2184 images of ten popular dog breeds, grouped into seven similarly sized primal mammalian emotion categories defined by neuroscientist and psychobiologist Jaak Panksepp as ‘Exploring’, ‘Sadness’, ‘Playing’, ‘Rage’, ‘Fear’, ‘Affectionate’ and ‘Lust’. We modified the contrastive learning framework MoCo (Momentum Contrast for Unsupervised Visual Representation Learning) to train it on our original dataset and achieved an accuracy of 43.2% and a baseline of 14%. We also trained this model on a second publicly available dataset that resulted in an accuracy of 48.46% but had a baseline of 25%. We compared our unsupervised approach with a supervised model based on a ResNet50 architecture. This model, when tested on our dataset with the seven Panksepp labels, resulted in an accuracy of 74.32%
MIT Department
System Design and Management Program.
Terms of Use
Creative Commons Attribution
Persistent DSpace Link
DOI of Published Version
https://doi.org/10.3390/s24227324
