MotoReID is a computer vision pipeline for MotoGP bike team detection, tracking, and re-identification from race footage. It uses YOLOv8 for detection, official Ultralytics BoT-SORT + ReID for tracking, and DINOv3 embeddings with a lightweight classifier for team identity. It targets key high-speed sports CV challenges: stable team recognition through heavy occlusions, rapid camera cuts, and motion blur.

Overview

MotoReID runs a multi-stage perception pipeline on race video:

• Detection (YOLOv8): localizes motorcycles frame by frame.
• Tracking (BoT-SORT + ReID): assigns short-term track IDs using the official Ultralytics tracker path.
• Feature Extraction (DINOv3 ViT-S/16): produces semantic embeddings from tracked bike crops.
• Team Classification (Logistic Regression): predicts team identity from DINO features.

This design combines strong detection, official tracking, and lightweight identity classification for practical sports-video analytics.

Dataset & Annotation

This project uses a manually curated MotoGP dataset built from race broadcast footage.

• Source: 1080p/60fps highlight footage.
• Frame extraction: OpenCV script via src/data/extract_frames.py.
• Annotation platform: Manually annotated with Roboflow.
• Detection set: 501 images with 1 class (motorbikes), 1,583 motorcycle box annotations, and 97 null/background images.
• Identity set: 668 curated team crops across 11 teams.
• Challenging views: rear-view, lean-angle, motion-blur, and occluded bike samples are included within the curated team crop folders to improve difficult-view robustness.

Training Pipeline

The training workflow is:

1. Download race highlight videos.
2. Extract representative frames from race videos.
3. Train/fine-tune YOLOv8 for motorcycle detection.
4. Auto-harvest bike crops with detector-assisted extraction.
5. Manually clean and sort crops into team folders.
6. Extract DINOv3 embeddings and train the team classifier.
7. Run full inference with BoT-SORT tracking and team classification.

Setup

Requirements

• Python 3.10+
• PyTorch 2.4+
• CUDA 12.x (recommended)
• Linux (tested on Ubuntu 22.04)

Install

git clone https://github.com/johnamit/motoreid.git
cd motoreid

conda create -n motoreid_env python=3.10 -y
conda activate motoreid_env
pip install -r requirements.txt

DINOv3 dependency

This project uses DINOv3, which requires a local clone of the official repository:

# Clone DINOv3 into the project root
git clone https://github.com/facebookresearch/dinov3.git

# Download Pre-trained Weights (ViT-Small)
mkdir -p models/DINO
wget -O models/DINO/dinov3_vits16_pretrain_lvd1689m.pth \
    https://huggingface.co/facebook/dinov3-vits16-pretrain-lvd1689m/resolve/main/dinov3_vits16_pretrain_lvd1689m.pth

MotoReID model artifacts

Download the trained detector and classifier from Hugging Face and place them at the paths expected by the inference scripts:

mkdir -p models/YOLO/motogp_yolov8m/weights models/classifier

wget -O models/YOLO/motogp_yolov8m/weights/best.pt \
    https://huggingface.co/johnamit/motoreid/resolve/main/motogp_yolov8m_detector.pt

wget -O models/classifier/dinov3_identity_model.pkl \
    https://huggingface.co/johnamit/motoreid/resolve/main/dinov3_team_classifier.pkl

Usage

Download race clips

bash src/scripts/download_race_highlights.sh

Full pipeline (detection + tracking + classification)

Run the complete pipeline. This script initializes the models, processes the video with BoT-SORT tracking, and outputs annotated results.

python src/inference/main.py \
  --video data/race_highlights/test/grandprix/2025_australian_gp_sprint.mp4 \
  --output results/pipeline/2025_australian_gp_sprint_full_pipeline.mp4 \
  --yolo_weights models/YOLO/motogp_yolov8m/weights/best.pt \
  --classifier_path models/classifier/dinov3_identity_model.pkl \
  --dino_weights models/DINO/dinov3_vits16_pretrain_lvd1689m.pth \
  --tracker models/YOLO/trackers/botsort_reid.yaml

YOLO-Only Detection

For simple detection without team classification:

python src/inference/detector.py \
    --video data/race_highlights/test/grandprix/2025_british_gp.mp4 \
    --output results/yolo_detection/2025_british_gp_yolov8m.mp4

Train Team Classifier

If you add new images to data/teams/, retrain the identity head:

python src/training/train_identity_model.py \
    --data_dir data/teams \
    --model_dir models/classifier \
    --dino_weights models/DINO/dinov3_vits16_pretrain_lvd1689m.pth \
    --batch_size 64

Extract Frames from Video

python src/data/extract_frames.py \
    --input data/race_highlights/train \
    --outdir data/race_frames \
    --interval 150

Harvest Bike Crops

Auto-crop detected bike regions for building the identity dataset:

python src/data/harvest_bikes.py \
    --video data/race_highlights/train/2025_qatar_gp.mp4 \
    --outdir data/teams/unlabeled \
    --gp qatar \
    --yolo_weights models/YOLO/motogp_yolov8m/weights/best.pt

Performance

Detector Performance

Current trained detector runs on the validation split:

Model	Best Epoch	Precision	Recall	mAP50	mAP50-95
YOLOv8n	89	0.9398	0.8974	0.9404	0.7659
YOLOv8m	92	0.9618	0.8910	0.9459	0.7735
YOLOv8l	100	0.9309	0.9039	0.9386	0.7496

Team Classifier Performance

Model	Crops	Teams	Accuracy	Macro F1	Weighted F1
DINOv3 + Logistic Regression	668	11	0.90	0.89	0.90

Teams

The system identifies all 11 teams from the 2025 MotoGP grid:

TEAM_COLORS = {
    'aprilia_factory':    (0, 255, 0),      # Green
    'aprilia_trackhouse': (0, 200, 100),    # Teal
    'ducati_lenovo':      (0, 0, 255),      # Red
    'ducati_gresini':     (0, 100, 255),    # Orange-Red
    'ducati_vr46':        (0, 255, 255),    # Yellow
    'honda_hrc':          (0, 0, 0),        # Black
    'honda_lcr':          (200, 200, 200),  # Light Gray
    'ktm_factory':        (0, 165, 255),    # Orange
    'ktm_tech3':          (0, 140, 255),    # Dark Orange
    'yamaha_monster':     (255, 0, 0),      # Blue
    'yamaha_pramac':      (255, 100, 100),  # Light Blue
}

Citation

DinoV3:

@article{simeoni2025dinov3,
  title={Dinov3},
  author={Sim{\'e}oni, Oriane and Vo, Huy V and Seitzer, Maximilian and Baldassarre, Federico and Oquab, Maxime and Jose, Cijo and Khalidov, Vasil and Szafraniec, Marc and Yi, Seungeun and Ramamonjisoa, Micha{\"e}l and others},
  journal={arXiv preprint arXiv:2508.10104},
  year={2025}
}

YOLOV8:

@software{yolov8_ultralytics,
  author = {Glenn Jocher and Ayush Chaurasia and Jing Qiu},
  title = {Ultralytics YOLOv8},
  version = {8.0.0},
  year = {2023},
  url = {https://github.com/ultralytics/ultralytics},
  orcid = {0000-0001-5950-6979, 0000-0002-7603-6750, 0000-0003-3783-7069},
  license = {AGPL-3.0}
}

BoT-SORT

@article{aharon2022bot,
  title={BoT-SORT: Robust associations multi-pedestrian tracking},
  author={Aharon, Nir and Orfaig, Roy and Bobrovsky, Ben-Zion},
  journal={arXiv preprint arXiv:2206.14651},
  year={2022}
}

License

This project is released under the MIT License

Note: MotoGP broadcast footage is copyrighted material. This project is intended for educational and research purposes only. All models were trained on fair-use excerpts for non-commercial analysis.