MMA Fighter Pose Estimation Dataset: Keypoint-Annotated UFC Stand-Up Combat Images for Computer Vision

Dataset Description:
This dataset is the second version of the MMA Fighter Detection Dataset, extending Version 1 (bounding box annotations) with 17-point COCO-style skeletal keypoint annotations for each detected fighter. It addresses the need for fine-grained body pose data in combat sports computer vision, enabling research into strike biomechanics, fighter posture analysis, and AI-assisted judging systems for mixed martial arts.

Research Hypothesis and Purpose:
Building on the hypothesis that automated fighter detection can serve as a foundation for AI-assisted judging, this version advances that goal by capturing full-body skeletal pose per fighter. Keypoint data enables quantification of body mechanics — guard positioning, weight distribution, striking extension, and defensive posture — directly relevant to judging criteria such as effective striking and aggression.

What the Data Shows:
The dataset contains 5,109 images from the same 20 professional UFC fights as Version 1, each annotated with 17 COCO-format keypoints per fighter instance in YOLO pose format. It captures stand-up striking exchanges, with keypoints following the standard COCO-17 skeleton (nose, eyes, ears, shoulders, elbows, wrists, hips, knees, ankles) and a per-point visibility flag (0 = not labeled, 1 = occluded, 2 = visible).

Data Collection Methodology:
Images were extracted from the same official UFC broadcast footage used in Version 1, selected to ensure diversity in fighter body types and styles (orthodox vs. southpaw), weight classes (bantamweight through light heavyweight), camera angles, and venue settings. All images are 640×640 pixels, split into train/valid/test sets (3,636 / 981 / 492). Keypoint annotations were generated through a semi-automated pipeline described in the Steps to Reproduce section.

Notable Characteristics and Limitations:
As with Version 1, the dataset is scoped to stand-up striking scenarios; ground grappling and cage clinch situations are excluded. Additional limitations: keypoints falling outside the Version 1 ground-truth bounding box were discarded to avoid cross-fighter contamination; heavily occluded keypoints carry lower visibility scores and should be used with caution; annotations are model-generated (YOLOv11x-pose) and may contain minor errors in complex poses.

How to Interpret and Use the Data:
Each .txt label file follows YOLO pose format — one instance per line:
<class> <cx> <cy> <w> <h> <kp0_x> <kp0_y> <kp0_v> ... <kp16_x> <kp16_y> <kp16_v>
All coordinates are normalised to [0, 1]. The included data.yaml enables immediate integration with Ultralytics YOLOv11 pose pipelines. For bounding boxes only, refer to Version 1 (DOI: 10.17632/c456bnk8bm.1).

数据集说明：本数据集为MMA格斗选手检测数据集的第二版，在第一版（仅包含边界框（bounding box）标注）的基础上，为每个检测到的选手新增了17点COCO（Common Objects in Context）格式骨骼关键点标注。该数据集填补了格斗运动计算机视觉领域对细粒度人体姿态数据的需求，可支撑混合格斗中的击打生物力学研究、选手姿态分析以及AI辅助裁判系统等方向的研究。

研究假设与研究目标：本数据集基于「自动化选手检测可作为AI辅助裁判系统的基础」这一学术假设，通过为每位选手采集全身骨骼姿态数据推进了该目标的实现。关键点数据可实现对身体力学的量化分析——包括护架姿势、体重分布、击打伸展幅度与防御姿态，这些指标均与有效击打、进攻性等裁判评判标准直接相关。

数据集内容说明：本数据集包含与第一版相同的20场UFC职业赛事的5109张图像，每张图像中每个选手实例均采用YOLO姿态格式标注17个COCO格式关键点。数据集涵盖站立击打对抗场景，关键点遵循标准COCO-17骨骼结构（包含鼻子、双眼、双耳、双肩、双肘、双腕、双髋、双膝、双踝），并为每个关键点配备可见性标记（0=未标注，1=被遮挡，2=可见）。

数据采集方法：图像提取自第一版所用的官方UFC赛事转播素材，筛选时确保覆盖不同体型与打法的选手（正架 vs 左架）、不同体重级别（雏量级至轻重量级）、不同拍摄角度与赛事场地。所有图像分辨率均为640×640像素，并划分为训练集、验证集与测试集（分别为3636张、981张、492张）。关键点标注通过「复现步骤」章节中描述的半自动化流程生成。

显著特征与局限性：与第一版一致，本数据集仅涵盖站立击打场景，不包含地面缠斗与笼边缠抱场景。其他局限性包括：超出第一版真实边界框范围的关键点会被剔除，以避免跨选手标注混淆；被严重遮挡的关键点可见性分数较低，使用时需谨慎；标注由模型（YOLOv11x-pose）生成，在复杂姿态下可能存在少量误差。

数据解读与使用方法：每个.txt标签文件遵循YOLO姿态格式，每行对应一个选手实例：
<类别> <中心x坐标> <中心y坐标> <宽度> <高度> <kp0_x> <kp0_y> <kp0_v> ... <kp16_x> <kp16_y> <kp16_v>
所有坐标均归一化至[0, 1]区间。附带的data.yaml文件可直接适配Ultralytics YOLOv11姿态处理流程。若仅需边界框标注，请参考第一版数据集（DOI: 10.17632/c456bnk8bm.1）。