Titan - MTB Rider Detection & Attribute Classification (Prototype)

Computer-vision pipeline for downhill MTB rider detection and multi-label attribute classification. Paused prototype.

Problem

Downhill MTB events produce large volumes of race-day photography, but tagging riders by bike, helmet, apparel, sponsor text and competitor number is still a manual job for broadcasters, scoring crews and photo-sales workflows. The goal of this prototype was to automate that attribution so a photo set could be made searchable by rider attributes end-to-end.

Approach

Two-stage pipeline:

1. Detection - Ultralytics YOLOv11m detects 9 object classes on the rider: bicycle, bicycle_text, clothes_text, competitor_number, cyclist, cyclist_clothes, cyclist_with_bike, helmet, helmet_text (boxes + polygon labels).
2. Classification - An EfficientNet-B2 backbone with a multi-label head (BCEWithLogitsLoss, auto-computed positive-class weights, tunable inference threshold) classifies 116 attributes per image: bike/helmet/apparel colors, bike brands (e.g. SantaCruz, Trek, Mondraker, RockShox), apparel brands (FOX, LEATT, TroyLeeDesigns, ...) and helmet brands.

Both stages support ONNX export out of the box (src/detection/train.py --export onnx, equivalent path on the classifier) and ship modular single-image and batch inference utilities. A Colab-oriented notebook drives the full training/eval flow end to end.

Stack

Component	Tech
Language	Python 3.10+
Detection	Ultralytics YOLOv11 (yolo11m)
Classification	EfficientNet-B2 (PyTorch / torchvision)
Loss	BCEWithLogitsLoss (auto pos_weight)
Optimization	AdamW + cosine schedule, warmup
Export	ONNX (also TorchScript / TFLite / CoreML via Ultralytics)
Notebooks	Jupyter / Google Colab
Data processing	OpenCV, Pillow, pandas, NumPy
Metrics	scikit-learn (F1 macro, Hamming loss, mAP)

Repo layout

configs/
  detection_config.yaml         # YOLOv11 hyperparameters + class list
  classification_config.yaml    # EfficientNet-B2 + attribute groups
src/
  detection/                    # YOLOv11 train / validate / inference
  classification/               # multi-label dataset, model, train, inference
  utils/                        # metrics + visualization
notebooks/
  titan_training.ipynb          # Colab-oriented driver notebook
requirements.txt

Local setup

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Training and inference entry points (Python module form):

# Detection
python -m src.detection.train     --config configs/detection_config.yaml
python -m src.detection.validate  --config configs/detection_config.yaml
python -m src.detection.inference --model runs/detection/.../best.pt --source path/to/images

# Classification
python -m src.classification.train     --config configs/classification_config.yaml
python -m src.classification.inference --model runs/classification/.../best.pt --source path/to/images

# ONNX export (after training)
python -m src.detection.train --config configs/detection_config.yaml --export onnx

Colab usage

1. Upload the two datasets to Google Drive at:
   • MyDrive/titan_project/models_datasets/titan_detection.yolov11/
   • MyDrive/titan_project/models_datasets/titan_labels.multiclass/
2. Open notebooks/titan_training.ipynb in Google Colab.
3. Run the cells in order - the notebook handles mounts, configs and both training loops.

Datasets (training-time, internal)

Stage	Images (with augmentation)	Labels
Detection	~222	9 classes, YOLO polygon format
Classification	~240	116 trainable attributes, one-hot CSV (117 cols incl. filename)

The datasets themselves are not included in this repository.

Status

• Two-stage CV prototype (detection + multi-label attribute classification) with ONNX export - currently paused.
• Training targets defined: mAP@0.5 for detection, F1 macro and Hamming loss for classification.
• Kept public as a record of the architecture.

Why this is in the portfolio

A standalone computer-vision prototype outside my usual web/backend stack - broadens AI breadth beyond LLM work, in a non-trivial vertical (downhill MTB) that doesn't have off-the-shelf rider-attribution models. See the portfolio entry.

License

MIT