Adsa Logo White Adsa Title White

Evaluating the predictive ability of point cloud deep learning to identify individual animals using surface-based body shape of dairy calves.

R. E. P. Ferreira

Abstract:

135
Evaluating the predictive ability of point cloud deep learning to identify individual animals using surface-based body shape of dairy calves.
R. E. P. Ferreira*1, T. Bresolin1, L. G. Pereira2,1, J. R. R. Dorea1. 1University of Wisconsin-Madison Madison, WI, 2Embrapa Dairy Cattle Ju�z de Fora, MG, Brazil.

Advances in machine learning techniques have allowed the development of computer vision systems (CVS) that can potentially monitor growth development of livestock animals. In this context, depth images from dorsal view are the most used inputs to estimate animal BW and biometrics. Frequently, such CVS rely on identification (ID) systems as a way to match animal ID and predicted phenotype. However, the use of surface-based body shape acquired from dorsal images to predict BW could also be adopted for animal recognition. Such alternative would optimize CVS to recognize animal ID and monitor growth development at the same time. Nonetheless, growing animals are continuously changing body shape, which could limit its use as an invariant feature for pattern recognition. Thus, the objectives of this study were: (1) to investigate the use of 3D dorsal images from calves as a potential tool for animal identification; and (2) to investigate if changes in body shape due to growth affect the prediction accuracy for animal identification. Images from 5 Holstein calves were acquired over 3 weeks using a depth sensor. From each image, the background was removed and only 3D data points of the animal's back were used. For objective 1, the algorithm was trained (n = 111), validated (n = 23) and tested (n = 22) using images within wk 1. For objective 2, the algorithm was trained (n = 221) and validated (n = 53) using images from wk 1 and 2, and tested using images from wk 3 (n = 172). Accuracy within wk 1 was 90.9% on the testing set. In objective 2, the overall accuracy decreased to 54.7% on the testing set. The prediction accuracy for each respective animal was: 38.5% (n = 32); 97.1% (n = 55); 88.4% (n = 62); 3.1% (n = 38); and 32.4% (n = 34). Animals with reduced number of images in the training data set presented lower accuracies on the testing set. Thus, increasing the amount of training samples can potentially improve predictive performance when later weeks are evaluated. The results show that the use of animal's 3D body surface is a promising tool for animal recognition.

Keywords: growth, calf, deep learning.