YOLOv8s MMA Fighter Detector

Fine-tuned YOLOv8s for detecting fighters in MMA and boxing footage. Part of the fight-judge project — an AI pipeline for automated combat sports scoring.

Model Description

This model detects the bounding box(es) of fighter(s) in a video frame. It is the first stage in the fight-judge pipeline, providing fighter crops that feed into the downstream pose estimation model.

Architecture: YOLOv8s (small)
Task: Object detection (single class: fighter)
Input: RGB image, 640×640 px
Output: Bounding boxes [x_center, y_center, width, height] (normalized), confidence scores
Base model: yolov8s.pt (COCO pretrained)
Finetuned on: MMA Fighter Detection Dataset — 5,106 images, 20 UFC fights

Training

Parameter	Value
Epochs	100
Batch size	64 (multi-GPU)
Image size	640
Optimizer	Auto (SGD)
LR initial	0.01
LR final	0.01 (cosine off)
Warmup epochs	3
Weight decay	0.0005
AMP	true
Hardware	2× GPU (Kaggle)
Training time	~1.2 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 Detection Dataset.

Metric	Value
mAP50-95 (box)	0.9836
mAP50 (box)	0.9950
Precision	0.9983
Recall	0.9990
Val box loss	0.1864
Val cls loss	0.1317

Training converged smoothly. mAP50-95 reached 0.72 by epoch 4 and exceeded 0.98 by epoch 24, stabilizing around 0.983–0.984 in the final epochs.

Dataset

MMA Fighter Detection Dataset

5,106 images extracted from 20 UFC fights (stand-up phases only)
640×640 px, YOLO format, single class: fighter
~2 fighter instances per image (10,186 total annotations)
Published on Mendeley Data (CC BY-NC-SA 4.0)
Fights span multiple weight classes — Poirier, Adesanya, Pereira, Topuria, Gaethje, and others

Split:

Split	Images
Train	3,635
Valid	980
Test	491

Usage

from ultralytics import YOLO

model = YOLO("hasanfaesal/yolov8s-mma-fighter-detector")

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

# Batch inference on a video
results = model("fight.mp4", stream=True)
for r in results:
    boxes = r.boxes.xyxy   # [x1, y1, x2, y2]
    confs = r.boxes.conf

Limitations

Trained exclusively on UFC stand-up footage. Performance may degrade on:
- Ground game sequences (grappling, guard, mount)
- Non-UFC broadcasts with different camera setups or overlays
- Boxing-specific camera angles (ringside, overhead)
Single-class model — does not distinguish between the two fighters (no ReID)
Not designed for crowd scenes or multi-person environments outside of 1v1 combat

Pipeline Context

[This model] → YOLOv11x-pose MMA Pose Estimator → Action Recognition (planned)

Fighter bounding boxes from this model are used as inputs to the top-down pose estimation stage. See the fight-judge repository for the full pipeline.

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

Downloads last month: 4