Project context:
- Enhance FinCarPrint, Asteroidea’s AI-based vehicle and license plate recognition system, focusing on detection accuracy and global adaptability.
- Work with a production AI product that must handle diverse real-world conditions (day/night, full vehicle vs. cropped plates, black & white and color images) across multiple countries' plate formats.
Objectives & responsibilities:
- Build and label a diverse dataset of vehicle and license plate images captured under various conditions (day/night, occlusions, different viewpoints, full vehicles and cropped plates).
- Train, test and fine-tune detection and recognition models (including experiments across country-specific license plate formats) to improve robustness and accuracy.
- Design and implement an automated continuous learning pipeline to periodically retrain models with newly labeled data and deploy updates.
Technical tasks & tools:
- Use Python, PyTorch and OpenCV as primary tools for dataset processing, model development, augmentation, and evaluation.
- Implement reinforcement learning components where applicable to improve model decision policies (for example: active learning selection, data augmentation policy learning, or model adaptation strategies).
- Create data preprocessing, augmentation, and validation scripts to handle color/black-and-white formats and varied image resolutions.
Evaluation, metrics & deliverables:
- Deliver a labeled dataset covering the specified variations (day/night, full/cropped, multi-country plates) with documentation of labeling schema and quality checks.
- Provide trained models, reproducible training scripts, evaluation reports (precision/recall, mAP, robustness tests across countries and lighting conditions) and a CI pipeline for automated retraining.
- Demonstrate improvements in detection accuracy and robustness through comparative benchmarks against baseline models.
Application details:
- To apply, send your CV and a brief motivation referencing the project ID AST-2026-004 and the role AI Developer to careers@asteroidea.co.