Bottom-up cattle pose estimation via concise multi-branch network

doi:10.1016/j.compag.2023.107945

2023-08-01

PRE

Sicong Liu

Shuqin Li

Chunjiang Zhao *

DOI：10.1016/j.compag.2023.107945

原文链接

原文求助

摘要

En 中文

There is a significant correlation between the poses of cattle and their health status. Estimating cattle's pose automatically is essential for discovering and detecting diseased animals (e.g. ketosis , foot-and-mouth disease) in a herd. Existing methods concentrate on the top-down paradigm, needing two models to finish the task; we follow the bottom-up paradigm, only one model, to complete the whole process. Based on HRNet, we construct a concise multi-branch network (CMBN) for cattle pose estimation. The entire structure, bottleneck, and basic block are strengthened to reduce parameters and FLOPs, enhance the representation ability of cattle instances, diminish the impact of the external surroundings, and boost the average precision of pose estimation. To evaluate the performance of CMBN, we use the NWAFU-Cattle dataset with more annotated cattle instances and extra supplementary data, containing 2432 images and 3101 instances. Experimental results reveal that the AP of cattle pose estimation arrived at 93.2, extracting efficacious features and locating the joint keypoints of multi-cattle in an image at the same time under complicated environments accurately. The performance of CMBN is superior to that of other state-of-the-art models, such as DEKR, HigherHRNet, and HRViT in the estimation of cattle pose. It has been shown that the parameters and FLOPs are 14.01 M and 9.83 G, respectively, which are far fewer than those of HigherHRNet, DEKR and HRViT. This approach provides a novel resolution for cattle pose estimation.

Keyword:

Cattle pose estimation

Bottom-up

Multi-branch network

Bottleneck