BlueRev waste detection pipeline

Improved a PyTorch segmentation workflow for underwater waste removal, with synthetic data generation and a modular training pipeline.

Multicultural team internship project (PER BlueRev).

Python
PyTorch
Segmentation
TensorBoard
Data augmentation

Machine learning pipeline diagram for BlueRev waste detection

Problem and dataset

Started from a small, noisy image dataset (~95 samples) with duplicates and inconsistent preprocessing. The goal was to improve segmentation quality for detecting waste (déchets) in underwater scenes.

ML pipeline redesign

Refactored the codebase into config, dataset, model, train, test, callback, and logger modules. Added hyperparameter config, EarlyStopping, best-epoch checkpointing, TensorBoard logging, LR tuning, and CPU/GPU-agnostic training. Built a synthetic data generator combining animal backgrounds, debris masks, and random compositing to expand the training set.

BlueRev training and inference pipeline — From raw images through synthetic augmentation to deployment.

Outcomes

Trained and compared multiple model configurations, logged metrics for reproducibility, and integrated the best checkpoints into the prediction application. The modular structure makes it easier to iterate on models and hyperparameters.