It is a time-consuming and laborious task to manually count the number of sheep in the process of pastoral livestock production. Counting the number of sheep in grassland animal husbandry has been helpful for overgrazing monitoring and grassland ecology assessment. A unmanned aerial vehicle (UAV) is used to obtain aerial images, and a UAV sheep counting (USC) dataset is made to provide data support for the study of flock dense counting. Based on the USC dataset, experiments with network models, included MCNN, CSRNet, SFANet, and Bayesian Loss are carried out. The experimental results show that due to occlusion, irregular distribution of sheep, great changes in sheep size, shape, and density, MCNN, CSRNet ,and SFANet models apply the assumed Gaussian kernel to point labeling to calculate the truth density map, which are difficult to achieve high quality. However, the Bayesian loss function proposed by Bayesian Loss model supervises the counting expectation of each sheep’s labeling points. The average absolute error (MAE) of the density map obtained by the Bayesian Loss model is 3.56, the mean square error (MSE) is 5.46, and the average relative error (MRE) is 1.86%, which provides a useful reference for the dense counting of grassland sheep.