UM-SWIM-360/swimmer-head-detection-face-blurring

Swimmer Head Detection (Face Blurring) Models

Object detection models that locate the head of a swimmer in video frames. They were trained as part of the SWIM-360 / DIVE project, a collaboration between the University of Malta and the Aquatic Sports Association of Malta, and are intended to support face anonymisation (blurring) of swimmers in swimming footage.

This repository contains several detectors from two model families, so they can be compared on accuracy and speed:

Family	Architectures	GitHub
Ultralytics YOLO	YOLO11m, YOLO26m, YOLO26L	ultralytics/ultralytics
Roboflow RF-DETR	RF-DETR Small, RF-DETR Large	roboflow/rf-detr

Repository structure

.
├── Ultralytics/
│   ├── YOLO11m/
│   │   └── swimmer_head_detection_yolo11m_exp1.pt
│   ├── YOLO26m/
│   │   ├── swimmer_head_detection_yolo26m_exp2.pt
│   │   └── swimmer_head_detection_yolo26m_with_augmentations_exp4.pt
│   └── YOLO26L/
│       └── swimmer_head_detection_yolo26L_exp3.pt
└── Roboflow/
    ├── RF-DETR Small/
    │   ├── swimmer_head_detection_rfdetr_small_best_ema_exp5.pth
    │   └── swimmer_head_detection_rfdetr_small_checkpoint_best_total_exp5.pth
    └── RF-DETR Large/
        ├── swimmer_head_detection_rfdetr_large_best_ema_exp6.pth
        └── swimmer_head_detection_rfdetr_large_checkpoint_best_total_exp6.pth

Available models

Experiment	Model	File	Notes
exp1	YOLO11m	Ultralytics/YOLO11m/swimmer_head_detection_yolo11m_exp1.pt	First experiment
exp2	YOLO26m	Ultralytics/YOLO26m/swimmer_head_detection_yolo26m_exp2.pt	Baseline
exp3	YOLO26L	Ultralytics/YOLO26L/swimmer_head_detection_yolo26L_exp3.pt	Larger YOLO26 variant
exp4	YOLO26m	Ultralytics/YOLO26m/swimmer_head_detection_yolo26m_with_augmentations_exp4.pt	YOLO26m (best model out of "exp1" - "exp3"). Trained with geometric augmentations
exp5	RF-DETR Small	Roboflow/RF-DETR Small/..._best_ema_exp5.pth and ..._checkpoint_best_total_exp5.pth	Best EMA weights / best overall checkpoint
exp6	RF-DETR Large	Roboflow/RF-DETR Large/..._best_ema_exp6.pth and ..._checkpoint_best_total_exp6.pth	Best EMA weights / best overall checkpoint

For the RF-DETR models, two checkpoints are provided per experiment: the best EMA weights and the best total checkpoint saved during training.

Task

• Task: object detection
• Class: face (a single class: the swimmer's head)
• Input: RGB video frames of swimmers, recorded on 120fps and 4k resolution; 3840x2160.
• Output: bounding boxes around swimmers' heads with confidence scores

Usage

Ultralytics YOLO

pip install ultralytics huggingface_hub

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

weights = hf_hub_download(
    repo_id="UM-SWIM-360/swimmer-head-detection-face-blurring",
    filename="Ultralytics/YOLO26m/swimmer_head_detection_yolo26m_with_augmentations_exp4.pt",
)

model = YOLO(weights)
results = model.predict("swimmer.jpg", conf=0.25)

for box in results[0].boxes:
    print(box.xyxy, box.conf)

RF-DETR

pip install rfdetr huggingface_hub

from huggingface_hub import hf_hub_download
from rfdetr import RFDETRSmall  # or RFDETRLarge for the Large models
from PIL import Image

weights = hf_hub_download(
    repo_id="UM-SWIM-360/swimmer-head-detection-face-blurring",
    filename="Roboflow/RF-DETR Small/swimmer_head_detection_rfdetr_small_best_ema_exp5.pth",
)

model = RFDETRSmall(pretrain_weights=weights)
detections = model.predict(Image.open("swimmer.jpg"), threshold=0.5)
print(detections.xyxy, detections.confidence)

Example: blur detected heads

import cv2
from ultralytics import YOLO

model = YOLO("swimmer_head_detection_yolo26m_with_augmentations_exp4.pt")
frame = cv2.imread("swimmer.jpg")

for box in model.predict(frame, conf=0.25)[0].boxes.xyxy.int().tolist():
    x1, y1, x2, y2 = box
    roi = frame[y1:y2, x1:x2]
    frame[y1:y2, x1:x2] = cv2.GaussianBlur(roi, (0, 0), sigmaX=15)

cv2.imwrite("swimmer_blurred.jpg", frame)

Training data

Models were trained on video frames with the ground truth located on a swimmer's head. The dataset included frames from the dynamic camera setup and a static camera setup datasets ; frames that do not have a swimmer's head visible are completely discarded.

Evaluation on Test Set

Best model (Based on mAP@50): YOLO26m

Worst model (Based on mAP@50): RF-DETR Small

Model	Experiment	Precision	Recall	mAP@50	mAP@50-95	Inference speed
YOLO11m	exp1	0.895	0.882	0.904	0.597	49.7 ms (20.1 FPS)
YOLO26m	exp2	0.895	0.880	0.912	0.607	49.0 ms (20.4 FPS)
YOLO26L	exp3	0.905	0.881	0.905	0.604	54.1 ms (18.5 FPS)
YOLO26m + augmentations	exp4	0.887	0.886	0.909	0.562	51.1 ms (19.6 FPS)
RF-DETR Small	exp5	0.902	0.871	0.871	0.519	50.7 ms (19.7 FPS)
RF-DETR Large	exp6	0.916	0.882	0.890	0.545	53.0 ms (18.9 FPS)

Citation

@misc{swim360facedetectionmodels2026,
  author       = {Camilleri, Vanessa and Seychell, Dylan and Montebello, Matthew and Azzopardi, Liam and Bartolo, Matthias and Camilleri, Reno Yuri and Pace, Daniel and Fialovszky, Mark and Gineste, Thomas and Scicluna, Justine and Aquilina, Gianluca},
  title        = {{SWIM-360 Swimmer Head Detection Face Blurring Models}},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/UM-SWIM-360/swimmer-head-detection-face-blurring}},
  note         = {Accessed: 2026-09-10}
}