imageomics/mmla-pose

该数据集包含来自MMLA无人机拍摄的斑马裁剪图像，手动标注了八种离散的姿态方向（前、前左、左、后左、后、后右、右、前右），相对于摄像头的视角。数据集旨在训练用于WildWing无人机自主导航系统的姿态/视角分类器（基于DINOv2骨干网络和MLP头部），其中了解动物哪一侧可见对于个体重识别（匹配侧面条纹图案）和行为感知飞行决策（例如避免从正面接近）至关重要。数据集包含988张手动标注的224x224 RGB裁剪图像，这些图像来自肯尼亚的Mpala研究中心和Ol Pejeta保护区以及俄亥俄州的Wilds保护中心的无人机镜头。每个裁剪图像被分配一个姿态方向标签，描述斑马哪一侧对摄像头可见（等效于斑马头部相对于摄像头的方向）。此外，还有35个模糊或不可用的裁剪图像保存在`_skip/`文件夹中，以记录标注决策。数据集的类别分布不平衡，侧面（左和右）占主导（约67%），反映了无人机观察的自然几何特征。This dataset contains cropped zebra images captured by MMLA drones, which are manually annotated with eight discrete pose directions relative to the camera's viewpoint: front, front-left, left, rear-left, rear, rear-right, right, and front-right. The dataset is intended to train a pose/viewpoint classifier for the WildWing UAV autonomous navigation system, which is built on a DINOv2 backbone and an MLP head. Knowing which side of the animal is visible is critical for individual re-identification (matching lateral stripe patterns) and behavior-aware flight decision-making, such as avoiding approaching from the front. The dataset consists of 988 manually annotated 224×224 RGB cropped images sourced from drone footage collected at Mpala Research Centre in Kenya, Ol Pejeta Conservancy, and The Wilds Conservation Centre in Ohio. Each cropped image is assigned a pose direction label that describes which side of the zebra is visible to the camera, which is equivalent to the direction of the zebra's head relative to the camera. Additionally, 35 blurry or unusable cropped images are stored in the `_skip/` folder to document annotation decisions. The dataset has an imbalanced class distribution, with the side classes (left and right) being dominant (approximately 67%), which reflects the natural geometric characteristics of UAV observation.