Data for automated feed intake detection in European Seabass (Dicentrarchus labrax) with computer vision

The data contains videos, models, training datasets and results of the development to detect feed intake in European seabass.
Averagely 20 grams European seabass reared in 3 different aquariums for 5 days. There are 10 fish in aquarium 1, 15 fish in aquarium 2 and 20 fish in aquarium 3. The aquarium are 240L in total volume but modified to have a green background and a slope at the bottom.
Recording are done with two See3CAM_CU27 until apparent satiation. The models are trained to detect and tracked the pellets. When the pellet disappeared, it triggers a search proximity area and window to search for a confirming “open mouth” for a feeding action. If there is an "open mouth" detected, then feeding event is counted.
The model folder contain the weights of the pellet detection model and the open mouth detection model (.pt or .onnx file) that can be used directly with a new input data.
The video folder contain the raw input video of European seabass during feeding period and the output videos of the computer vision automatic detection of feed intake. The videos are representatives. One video per aquarium chosen by the largest number of feeding event detected for that particular aquarium. For access to all of the videos, contact the authors.
The results are the csv files containing all of the recorded feed intake with headings:
aquarium. The number of aquarium (1, 2 or 3) Day. Correspond to the date (24, 25, 26…) Feeding_Event_Id. The number of the feeding event per video Frame_Number. The number of the frame that the feeding event detected for the particular video Pellet_Disappearance_X. X coordinate of the last detection of the pellet Pellet_Disappearance_Y. Y coordinate of the last detection of the pellet Open_Mouth_X. X coordinate of the open mouth bounding box that confirmed the feeding event. Open_Mouth_Y. X coordinate of the open mouth bounding box that confirmed the feeding event. Open_Mouth_Confidence. The confidence of the particular open mouth detection that confirm the feeding event. Distance_Pellet_Mouth. The pixel distance from the pellet last detection to the open mouth bounding box of the feeding event Video_Filename. The video filename of the corresponding feeding event Pellet_Conf_Threshold. The confidence threshold for pellet detection. Open_Mouth_Conf_Threshold. The confidence threshold for open mouth detection. Proximity_Threshold. The proximity search for open mouth from the coordinate of the pellet last detection (in pixels). Pellet_Disappearance_Frames. Number of frames where a particular tracked pellet undetected for before considered disappeared. Open_Mouth_Detection_Window. A frame number where an open mouth is search for from the key frame where the pellet disappear. Min_Pellet_Detection_Frames. A minimum number of frame to validate a pellet detection as a true pellet not just a debris. Priority_Area_Ratio. Ratio of a priority area on top of the video where pellet detection become valid immediately bypassing Min_Pellet_Detection_Frames Waste_Threshold_Y_Ratio. Ratio of the zone at the bottom of the video where pellet disappearance at that area will count as pellet waste.
The training_datasets folder contains the annotated images and label to train the pellet detection and open mouth detection models.

本数据集包含用于检测欧洲海鲈（European seabass）摄食量的研发相关视频、模型、训练数据集及实验结果。
实验选取平均体重20克的欧洲海鲈，在3个不同水族箱中饲养5天。其中1号水族箱放养10尾鱼，2号水族箱15尾，3号水族箱20尾。所有水族箱总容积为240升，经改造后带有绿色背景与底部斜坡。
采用两台See3CAM_CU27摄像机进行录制，直至鱼群达到表观饱食状态。所训练的模型用于检测并追踪饲料颗粒，当饲料颗粒消失时，系统将触发邻近区域搜索窗口，以寻找可确认摄食行为的“张嘴”动作。若检测到“张嘴”动作，则计为一次摄食事件。
模型文件夹内包含可直接用于新输入数据的饲料颗粒检测模型与张嘴检测模型权重文件（格式为.pt或.onnx）。
视频文件夹包含欧洲海鲈摄食期的原始输入视频，以及计算机视觉自动检测摄食量的输出视频。所选视频均具有代表性，每个水族箱对应一支视频，选取标准为该水族箱中检测到的摄食事件数量最多。如需获取全部视频，请联系论文作者。
实验结果为包含所有记录摄食量数据的CSV文件，其表头字段及说明如下：
aquarium：水族箱编号（1、2或3）；
Day：实验日期（如24、25、26……）；
Feeding_Event_Id：单支视频内的摄食事件编号；
Frame_Number：对应摄食事件被检测到时的视频帧编号；
Pellet_Disappearance_X：饲料颗粒最后一次被检测到的X坐标；
Pellet_Disappearance_Y：饲料颗粒最后一次被检测到的Y坐标；
Open_Mouth_X：确认摄食事件的张嘴边界框（bounding box）的X坐标；
Open_Mouth_Y：确认摄食事件的张嘴边界框的Y坐标（原文此处描述存在笔误，已修正为Y坐标）；
Open_Mouth_Confidence：确认摄食事件的张嘴检测置信度；
Distance_Pellet_Mouth：饲料颗粒最后一次检测位置与摄食事件张嘴边界框之间的像素距离；
Video_Filename：对应摄食事件的视频文件名；
Pellet_Conf_Threshold：饲料颗粒检测的置信度阈值；
Open_Mouth_Conf_Threshold：张嘴检测的置信度阈值；
Proximity_Threshold：以饲料颗粒最后检测坐标为基准的张嘴邻近搜索阈值（单位：像素）；
Pellet_Disappearance_Frames：某一被追踪饲料颗粒被判定为消失前，连续未被检测到的帧数；
Open_Mouth_Detection_Window：以饲料颗粒消失的关键帧为起始，搜索张嘴动作的帧范围；
Min_Pellet_Detection_Frames：验证饲料颗粒检测结果为真实颗粒而非碎屑所需的最小帧数；
Priority_Area_Ratio：视频顶部优先区域的比例，该区域内的饲料颗粒检测可直接跳过Min_Pellet_Detection_Frames验证；
Waste_Threshold_Y_Ratio：视频底部区域的比例阈值，若饲料颗粒在该区域内消失，则计为饲料浪费。
training_datasets文件夹包含用于训练饲料颗粒检测与张嘴检测模型的标注图像及标注标签。