Scholarlink

<p>Interpreting cat emotions poses a significant challenge due to the subtlety and absence of facial expressions, making it difficult to indicate their feelings. Our study addresses this challenge by experimenting with supervised and semi-supervised models for the classification of cat facial emotions on the Cat Facial Landmark in the Wild (CatFLW) dataset. Due to the lack of ground truth labels, we labeled a portion of the dataset based on shapes of facial features inferred from a previous study on cat behaviors. Subsequently, we examined supervised learning approaches with four model specifications derived from a combination of cat facial images and cat facial landmark vector inputs. The result for the best performing supervised model is the vanilla multi-layer perceptron (MLP) using cat facial landmark vector input with validation accuracy 95.73 ± 1.28%. For semi-supervised learning, we compared two techniques widely used in semi-supervised learning, FixMatch and Mean Teacher. Following the result of the supervised learning, we used the vanilla MLP as our model structure for both semi-supervised methods. Mean Teacher was found to be the better semi-supervised method for our dataset with the unlabeled accuracy of 95.65 ± 0.68% and the validation accuracy only slightly below the supervised model at 90.29 ± 2.00%, demonstrating sufficient generalization performance. Our study demonstrates the effectiveness of the proposed models and introduces a framework for interpreting cat emotions using both image and facial landmark data. We also contributed labeled cat facial expressions to the CatFLW dataset (<a href="https://raw.githubusercontent.com/geexe/cat-facial-expression-recognition/main/CatFLW%20Dataset/psuedo-labels.csv">https://raw.githubusercontent.com/geexe/cat-facial-expression-recognition/main/CatFLW%20Dataset/psuedo-labels.csv</a>). However, since the labels were not provided by feline behavior experts, the dataset may contain subjectivity and limited generalizability, which we aim to address in future work.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Advancing cat facial emotion interpretation: evaluating supervised and semi-supervised models using facial landmark and image data

  • 期刊论文

  • 发表日期：

    2026年1月29日

查看全文
  • Thanyakorn Hovongratana,
  • Budsadee Sareerasart,
  • Therarat Srisaswatakul,
  • Wanlipa Chamchum,
  • Thitirat Siriborvornratanakul

摘要

Interpreting cat emotions poses a significant challenge due to the subtlety and absence of facial expressions, making it difficult to indicate their feelings. Our study addresses this challenge by experimenting with supervised and semi-supervised models for the classification of cat facial emotions on the Cat Facial Landmark in the Wild (CatFLW) dataset. Due to the lack of ground truth labels, we labeled a portion of the dataset based on shapes of facial features inferred from a previous study on cat behaviors. Subsequently, we examined supervised learning approaches with four model specifications derived from a combination of cat facial images and cat facial landmark vector inputs. The result for the best performing supervised model is the vanilla multi-layer perceptron (MLP) using cat facial landmark vector input with validation accuracy 95.73 ± 1.28%. For semi-supervised learning, we compared two techniques widely used in semi-supervised learning, FixMatch and Mean Teacher. Following the result of the supervised learning, we used the vanilla MLP as our model structure for both semi-supervised methods. Mean Teacher was found to be the better semi-supervised method for our dataset with the unlabeled accuracy of 95.65 ± 0.68% and the validation accuracy only slightly below the supervised model at 90.29 ± 2.00%, demonstrating sufficient generalization performance. Our study demonstrates the effectiveness of the proposed models and introduces a framework for interpreting cat emotions using both image and facial landmark data. We also contributed labeled cat facial expressions to the CatFLW dataset (https://raw.githubusercontent.com/geexe/cat-facial-expression-recognition/main/CatFLW%20Dataset/psuedo-labels.csv). However, since the labels were not provided by feline behavior experts, the dataset may contain subjectivity and limited generalizability, which we aim to address in future work.