Automated recognition of postures and drinking behaviour for the detection of compromised health in pigs

The Dataset.zip file comprises of three folders as the following:

1) Animal trial
This folder includes the full image dataset (with .mp4 format) of the food restriction trial. Fifteen pigs were
housed, under commercial conditions, in a single, fully-slatted pen from 9 to 14 weeks of age. Food and water were provided to the pens using 4 drinking
sources and 4 feeding troughs respectively. Pigs were fed ad-libitum
a commercial food appropriate for their stage of growth. Food that
remained in the feeding troughs was removed, weighed and replaced, with a
predefined amount of new food at approximately 09:30 of every morning.
The ad-libitum protocol applied for 4 days. Subsequently, pigs were quantitatively food-restricted with the pen receiving 80% of their daily ad-libitum feed for 4 consecutive days at 12 weeks of age. Food allocation at 12 weeks of age consisted of ad-libitum
daily intake: 0.059–0.070 kg feed/kg initial total pen body weight;
restriction daily intake: 0.047 kg feed/kg initial total pen body
weight. Immediately following the 4 days of food restriction, pigs were
returned to ad-libitum feeding. No adverse effects were recorded
at any point during the food restriction protocol. After completion of
the study, all the pigs were checked by a veterinarian and released back
into the commercial stock.

2) Annotation for postures and drinking

Sample frames were selected from our database of video sequences to construct a data set for training and testing. Our behavioural dataset comprised a total of 7,35,094 instances (an instance denotes an individual pig label with its bounding box coordinates in a given image; this is provided with .mat format) across 1,13,379 images (with .jpg format); each pig within an image was manually annotated into one of five categories. By definition, the bounding box specifies the location of a pig within an image. It contains a vector in the format [x y width height], where x and y correspond to the upper left corner of the bounding box while the width and height denote width and height of the rectangular-shaped box around each pig.

3) Code of the detection method

The code used to detect pig postures and drinking behaviour.

本数据集压缩包包含以下三个文件夹：
1. 动物试验
该文件夹包含限饲试验的完整视频数据集（格式为.mp4）。试验期间，15头生猪在商业化养殖条件下饲养于单栏全漏缝地板栏舍中，饲养周期为9至14周龄。该栏舍分别配备4个饮水点与4个饲槽，用于供应饮水与饲料。生猪自由采食适配其生长阶段的商业饲料，每日约9:30时，工作人员会清除饲槽内剩余饲料，称重后更换预设分量的新鲜饲料。自由采食模式持续4天。
在12周龄时，随后连续4天对生猪实施定量限饲，当日饲槽供应的饲料为该组生猪当日自由采食量的80%。12周龄时的饲料配给标准为：自由采食模式下，每日每千克栏舍初始总体重对应投喂0.059~0.070 kg饲料；限饲模式下，每日每千克栏舍初始总体重对应投喂0.047 kg饲料。4天限饲期结束后，立即恢复生猪的自由采食模式。整个限饲试验期间未记录到任何不良反应。试验结束后，所有生猪均经兽医检查合格后放回商业种猪群。

2. 姿态与饮水行为标注数据集
本数据集从视频序列数据库中抽取样本帧，用于构建训练与测试数据集。本行为数据集共包含735094个标注实例（每个实例对应单头生猪的标签及其在对应图像中的边界框坐标，以.mat格式存储），覆盖113379张.jpg格式图像；图像内的每头生猪均经人工标注为5个类别之一。根据定义，边界框用于标注生猪在图像中的位置，其采用[x y width height]格式的向量表示：x与y分别代表边界框左上角的横、纵坐标，width与height分别代表该矩形框的宽度与高度。

3. 检测方法代码
本代码用于检测生猪姿态与饮水行为。