hasanfaesal/fight-judge-yolov11x-pose

YOLO11x-pose MMA Fighter Pose Estimator

Fine-tuned YOLO11x-pose for 17-keypoint pose estimation of fighters in MMA and boxing footage. Part of the fight-judge project — an AI pipeline for automated combat sports scoring.

Model Description

This model jointly detects fighters and estimates their 17 COCO keypoints in a single forward pass. It is the second stage in the fight-judge pipeline, providing skeleton sequences that feed into downstream action recognition.

• Architecture: YOLO11x-pose (extra-large)
• Task: Pose estimation (single class: fighter, 17 COCO keypoints)
• Input: RGB image, 640×640 px
• Output: Bounding boxes + 17 keypoints per fighter [x, y, visibility]
• Base model: yolo11x-pose.pt (COCO pretrained), continued from epoch 90 checkpoint
• Finetuned on: MMA Fighter Pose Estimation Dataset — 5,106 images, 10,155 fighter instances with auto-generated keypoint annotations

Keypoints (COCO 17-point schema)

Index	Keypoint	Index	Keypoint
0	nose	9	left_wrist
1	left_eye	10	right_wrist
2	right_eye	11	left_hip
3	left_ear	12	right_hip
4	right_ear	13	left_knee
5	left_shoulder	14	right_knee
6	right_shoulder	15	left_ankle
7	left_elbow	16	right_ankle
8	right_elbow

Training

Training was conducted in two phases on Kaggle (Tesla P100-PCIE-16GB):

Phase 1 (epochs 1–90): Initial finetuning from COCO pretrained yolo11x-pose.pt

Phase 2 (epochs 91–150): Resumed from the epoch 90 checkpoint with AdamW optimizer

Parameter	Value
Total epochs	150 (resumed at ep. 91)
Batch size	8
Image size	640
Optimizer	AdamW
LR initial	0.001
LR final factor	0.01
Warmup epochs	3
Weight decay	0.0005
Patience	50
Save period	every 10 epochs
AMP	true
Hardware	1× Tesla P100-PCIE-16GB
Training time	~7.4 hours

Augmentations: HSV jitter, horizontal flip (p=0.5), scale, mosaic, random erasing (p=0.4), RandAugment.

Evaluation Results

Evaluated on the held-out test split of the MMA Fighter Pose Estimation Dataset at epoch 150.

Detection (box)

Metric	Value
mAP50-95 (box)	0.9859
mAP50 (box)	0.9950
Precision	0.9990
Recall	0.9995
Val box loss	0.1687
Val cls loss	0.1092

Pose (keypoints)

Metric	Value
mAP50-95 (pose)	0.9198
mAP50 (pose)	0.9932
Precision (pose)	0.9924
Recall (pose)	0.9907
Val pose loss	0.5105
Val kobj loss	0.0016

Training started from a strong checkpoint (epoch 90, pose mAP50-95 ≈ 0.875) and improved steadily to 0.920 by epoch 150, with losses still decreasing at the end of training.

Dataset

MMA Fighter Pose Estimation Dataset

• 5,106 images extracted from 20 UFC fights (stand-up phases only)
• 640×640 px, YOLO-Pose format, single class: fighter
• 10,186 fighter instances; 10,155 (99.7%) successfully labeled with 17 COCO keypoints
• Keypoint annotations auto-generated using pretrained yolo11x-pose with IoU ≥ 0.6 matching against ground-truth fighter bboxes
• Images sourced from the MMA Fighter Detection Dataset (CC BY-NC-SA 4.0)

Split:

Split	Images	Fighter instances
Train	3,635	~7,250
Valid	980	~1,960
Test	491	~980

Usage

from ultralytics import YOLO

model = YOLO("hasanfaesal/yolov11x-pose-mma-fighter")

results = model("fight_frame.jpg")
results[0].show()

# Access keypoints directly
results = model("fight_frame.jpg")
for r in results:
    keypoints = r.keypoints.xy    # [N, 17, 2] — x, y pixel coords
    visibility = r.keypoints.conf # [N, 17]    — confidence per keypoint
    boxes = r.boxes.xyxy          # [N, 4]     — bounding boxes

# Video inference
results = model("fight.mp4", stream=True)
for r in results:
    kps = r.keypoints.data  # [N, 17, 3] — x, y, conf

Limitations

• Trained exclusively on UFC stand-up footage. Performance may degrade on:
  • Ground game sequences (grappling, guard, mount) where limbs are heavily occluded or interleaved
  • Non-UFC broadcasts with unusual camera angles or significant lens distortion
  • Extreme close-ups where only part of the fighter's body is visible
• Keypoint annotations were auto-generated (pseudo-labels) — rare edge cases may contain labeling noise
• Visibility flags may be unreliable for highly occluded keypoints in clinch situations

Pipeline Context

YOLOv8s MMA Fighter Detector → [This model] → Action Recognition (planned)

See the fight-judge repository for the full pipeline including data preparation scripts and methodology.

Citation

If you use this model or the dataset, please cite:

@misc{faisal2025mmafighter,
  author    = {Hasan Faisal},
  title     = {MMA Fighter Detection Dataset},
  year      = {2025},
  publisher = {Mendeley Data},
  version   = {V1},
  doi       = {10.17632/c456bnk8bm.1}
}

License

Code and model weights: MIT License Dataset: CC BY-NC-SA 4.0