Multi-Center Agent Loss for Visual Identification of Chinese Simmental in the Wild

Area of Science:

• Computer Vision
• Animal Science
• Machine Learning

Background:

• Real-time cattle monitoring is crucial for precision livestock farming.
• Chinese Simmental cattle possess unique coat patterns (white stripes/spots) suitable for biometric identification.
• Existing datasets lack multi-view images for robust cattle identification.

Purpose of the Study:

• To develop a novel method for visual identification of individual cattle from any viewpoint.
• To create a new dataset for training and evaluating multi-view cattle identification models.
• To improve the accuracy and efficiency of cattle identification in livestock farming.

Main Methods:

• Proposed a multi-center agent loss function to jointly train deep convolutional neural networks (DCNNs).
• Reformulated SoftMax with multiple centers to reduce intra-class variance.
• Utilized an agent triplet loss to enforce inter-class separation.
• Created the CNSID100 dataset with 11,635 multi-view images of 100 Chinese Simmental cattle.

Main Results:

• The proposed multi-center agent loss outperformed state-of-the-art methods on the CNSID100 and OpenCows2020 datasets.
• Reduced intra-class variance and improved inter-class separability.
• Demonstrated promising performance in an engineering application for continuous cattle identification.

Conclusions:

• The developed method offers a robust solution for individual cattle identification using visual characteristics.
• The new CNSID100 dataset facilitates research in multi-view animal identification.
• The proposed pipeline shows significant potential for practical application in precision livestock farming.