3D shape analysis of cattle to improve data collection

To improve genetic gain in cattle breeding, Geno SA and TYR wants to improve the data collection. A lot of manual work is currently needed to record some traits, and we want to want to automatize this data collection by utilizing 3D imaging technology. To extract traits for breeding from 3D images, we need to develop algorithms, predicting meaningful quantitative measures about the animal's size shape or quality, based on 3D images. In this project, we want to utilize machine learning techniques to predict traditional traits, like the bull's weight or the teat positioning of a cow. We have built data sets consisting of both 3D images and the traditional measures we want to improve. On these data sets, we are using supervised learning to train the deep neural networks to predict the traditional traits based on the 3D images. Given a relevant 3D image, the trained model should be able to predict the traditional trait, and the traditional recording could then be skipped, which could both improve and streamline data collection on cattle. Until now, we have been developing deep learning based classifiers to distinguish between useful and non-useful images. Since some blurry or irrelevant test-images are taken during the acquisition phase, such images need to be automatically filtered out before further post-processing. Further, we have been working on prediction of supermummerary teats based on the images. We also used the data augmentation approaches to increase the number of images that help in improving the performance of deep learning models. We also used RGB images, depth images, and fusion RGB and depth images, to improve the performance of deep learning model for cattle teat length analysis. Considering front teat placement trait, we explored different deep learning models, with a specific focus on transfer learning. By employing this methodology, we were able to leverage the existing knowledge contained within pretrained models. For cattle weight estimation, we developed and analyzed a range of data modalities, such as RGB images, depth images, merged RGB and depth images, segmentation images, and fused segmentation with depth images. The importance of color information in producing precise outcomes is revealed by our research, which demonstrates the effectiveness of RGB, RGB and depth, segmentation, and merged segmentation with depth modalities. However, depending solely on depth data was found to be inadequate, as it could not capture the complex patterns required for precise weight estimates.

With this project we aim to build up competence and tools that allow us to improve and streamline our routine data collection. The outcome of this project is therefore reduced amount of manual work, which could lead to increased amount of data and reduced labour costs. In addition, by digitizing the value chain, the quality of data is going up. This would impact the quality of the selection process, and further, improve the products which allow us to increase sales and/or prices in the international market. To further elaborate, the primary objective of this project is to foster proficiency and create instruments that enable the improvement and streamlining of our regular data collection procedures in the livestock sector. Through the utilization of a range of deep learning models and diverse image modalities, encompassing RGB and depth data, in conjunction with the application of transfer learning and data augmentation techniques, our objective is to significantly enhance the effectiveness and precision of numerous tasks. These tasks encompass the classification of udder and non-udder cattle images, supernumerary teat classification, analysis of front teat placement traits, teat length traits, and estimation of cattle weight. We have achieved good results in the aforementioned areas. The results and completion will result in a substantial decrease in the need for human work, therefore allowing us to handle higher quantities of data while also reducing labor expenses. Now it is a matter of deploying these developed techniques and models in the future. This will not only make our operations more efficient but also improve our competitiveness and profitability in the worldwide market.

To improve genetic gain within cattle breeding, relevant data need to be collected on a large number of animals. However, for some traits, data collection requires a lot of manual work, and automation of these processes has the potential to both save costs and improve the genetic gain. Examples of such traits are weight measurements on beef cattle (where the cattle needs to be located on the weight) and udder conformation on dairy cows, which is currently scored by visual inspection by technicians. Geno, TYR and other collaborators are now in another project testing out whether use of 3D cameras in combination with computer vision could replace traditional recording systems. Main objective: In this project, we want to develop automatic computer vision algorithms that utilize the surfaces to predict geometric properties relevant for the cattle breeding. Reference data sets are under constrcion, where both 3D images and traditional recordings are taken from the same animals. Based on computer vision and machine learning techniques (like e.g. supervised deep learning), we aim to train regression models that could be used to predict relevant phenotypes, based on 3D images. • Secondary objective1: Develop models to predict traditional udder conformation traits, based on 3D images. Traditional conformation score data and predicted traits from 3D images should be used into genetic analyses for comparison and estimation of e.g. heritability. • Secondary objective 2: Develop regression models that are able to predict weight, carcass-quality and body condition score from 3D images of cattle, taken from above. • Secondary objective 3: Develop new relevant traits for breeding. To do this, we aim to build a reference atlas of the udder surface to achieve point correspondence between individuals. Further, by using dimension techniques (like Partial least squares regression), we want to combine the 3D surfaces with health data, to find health risk indicators in complex 3D data.