MOO (Multi-view Oriented Observations)
收藏资源简介:
MOO是由巴黎萨克雷大学与索邦大学联合构建的大规模合成牛只重识别数据集,包含1,000个虚拟牛个体从128个均匀采样视角(方位角360°、俯仰角-25°至90°)渲染的128,000张标注图像。数据集通过Blender生成512×512像素的RGB图像及深度图,并附有精确的几何元数据(角度、相机参数等),旨在消除背景干扰和视角偏差。其创新性在于首次系统量化了俯仰角对动物重识别的影响,揭示了30°关键阈值规律,并通过零样本和微调实验验证了合成数据对真实场景(如牲畜管理)的迁移价值。
MOO is a large-scale synthetic cattle re-identification dataset jointly developed by Paris-Saclay University and Sorbonne University. It contains 128,000 annotated images rendered from 128 uniformly sampled viewpoints (azimuth range: 360°, pitch angle range: -25° to 90°) for 1,000 virtual cattle individuals. The dataset generates 512×512 pixel RGB images and depth maps using Blender, and is accompanied by precise geometric metadata including angles, camera parameters and others. Its core goal is to eliminate background interference and viewpoint bias. The innovation of this dataset is that it systematically quantifies the impact of pitch angle on animal re-identification for the first time, reveals the 30° critical threshold law, and verifies the transferability of synthetic data to real-world scenarios such as livestock management through zero-shot and fine-tuning experiments.
数据集概述
基本信息
- 数据集状态:准备中,即将发布。
- 数据集地址:https://github.com/TurtleSmoke/MOO
联系信息
- 紧急事宜联系人:William Grolleau
- 联系人邮箱:william.grolleau@cea.fr
备注
- 数据集详情页面内容显示数据集尚未发布,具体信息待后续更新。
- 1MOO: A Multi-view Oriented Observations Dataset for Viewpoint Analysis in Cattle Re-Identification巴黎萨克雷大学·CEA List实验室; 索邦大学·CNRS ISIR实验室 · 2026年



