beketov027/soccer_player_ball_keypoint_detection_models

YOLOv11 Soccer Detection & Keypoints

Inference pipeline for soccer match images using two YOLOv11 models:

• Player / Ball / Referee detection — yolov11_sahi_1280
• Field keypoint detection — yolov11_keypoints_29 (29 pitch keypoints)

Project Structure

yolo26_ball_player_keypoint_detection/
├── inference.py
├── original.jpg              # default input image
├── output.jpg                # annotated output (generated)
├── yolov11_sahi_1280/
│   └── Model/weights/best.pt
└── yolov11_keypoints_29/
    └── Model/weights/best.pt

Requirements

• Python 3.8+
• Ultralytics
• OpenCV

pip install ultralytics opencv-python

Usage

Place your input image in the project folder (or update INPUT_IMAGE in inference.py), then run:

python inference.py

The script will:

1. Run object detection at 1280px resolution (Player, Ball, Referee)
2. Run keypoint detection (29 field keypoints)
3. Draw both results on the image
4. Save the annotated image to output.jpg

Example output:

Input:  .../original.jpg
Output: .../output.jpg
Detections: {'Referee': 2, 'Player': 14, 'Ball': 1}
Keypoint sets: 1
Keypoints per set: 29

Configuration

Edit the constants at the top of inference.py:

Constant	Default	Description
INPUT_IMAGE	original.jpg	Path to the input image
OUTPUT_IMAGE	output.jpg	Path to save the annotated output
PLAYER_CONF	0.57	Minimum confidence for player detections
DEFAULT_CONF	0.25	Minimum confidence for ball and referee

Models

Detection (yolov11_sahi_1280)

Class	ID
Player	0
Ball	1
Referee	2

Trained with SAHI-style slicing at 1280px resolution for better small-object detection (e.g. the ball).

Keypoints (yolov11_keypoints_29)

Pose model that detects 29 field keypoints per image (pitch lines, corners, penalty areas, etc.).

---

Goldenboy