PCR: A Large-Scale Benchmark for Pig Counting in Real World

Automatic pig counting with pattern recognition and computer vision techniques, despite its significance in intelligent agriculture, remains to be a relatively unexplored area and calls for further study. In this paper, we propose a large-scale image-based Pig Counting in Real world (PCR) dataset, covering a variety of real-world scenarios and environmental factors. The dataset consists of two subsets, i.e., PartA captured on real-world pig pens and PartB collected from the Internet, with center point annotations of pig torsos in 4844 images. Moreover, we develop an automatic pig counting algorithm based on weakly-supervised instance segmentation, which can output a single segmentation blob per instance via the proposed Segmentation-Split-Regression (SSR) loss, utilizing point-level annotations only. Experiments show that the proposed algorithm achieves state-of-the-art counting accuracy and exhibits superior robustness against challenging environmental factors. The dataset and source codes are available at https://github.com/jierujia0506/PCR .