Williamsanderson/PoultryVision-Dataset

PoultryVision统一数据集是一个大规模、多模态的家禽养殖数据集，整合了六个公共来源的数据，用于鸡（肉鸡、母鸡、公鸡）和鸡蛋的检测、分类、多摄像头跟踪和行为分析。该数据集旨在训练Williamsanderson/PoultryVision模型，这是一个YOLOv11m模型，性能优于Cardoen等人（MVBroTrack论文，2025年）报告的微调YOLOv11x模型，mAP@50-95提高了8.5%。数据集包含21,586张用于目标检测的图像（训练集15,987张，验证集3,706张，测试集1,893张）和2,539张用于图像分类的图像（训练集1,832张，验证集444张，测试集263张）。此外，还包括来自4个同步摄像头的24个MP4视频、相机校准文件、重投影掩码、跟踪地面实况以及预计算的每帧YOLO检测结果。数据集采用统一的YOLO检测格式和ImageFolder分类格式，并分为训练集、验证集和测试集。

The PoultryVision Unified Dataset is a large-scale, multi-modal poultry-farm dataset that unifies six public sources for detection, classification, multi-camera tracking, and behavior analysis of chickens (broilers, hens, cocks) and eggs. This dataset was built to train Williamsanderson/PoultryVision, a YOLOv11m model that beats the fine-tuned YOLOv11x reported by Cardoen et al. (MVBroTrack paper, 2025) by +8.5 mAP@50-95. The dataset includes 21,586 images for object detection (15,987 train, 3,706 val, 1,893 test) and 2,539 images for image classification (1,832 train, 444 val, 263 test). Additionally, it contains 24 MP4 videos from 4 synchronized cameras, camera calibration files, reprojection masks, tracking ground truth, and pre-computed YOLO detections per frame. The dataset is standardized to a unified YOLO detection format and/or ImageFolder classification format, deduplicated, and split into train/val/test.