Experience

The system follows a hybrid computer-vision approach combining person detection, multi-object tracking, and pose estimation for behavioral analysis in crowded railway environments.

• Delivered an automated evaluation pipeline for eye products, from per-image calibration and keypoint extraction through metric computation and BI reporting. Applied factorial Design of Experiments to assess product, acquisition, and operator effects; product formulation dominated the measured variance while operator influence remained below the noise threshold, supporting a reproducible ranking framework.
• Designed a data-centric approach for foundation evaluation from only a few hundred target images. Combined synthetic-data generation, transfer learning from a related imaging domain, and a lightweight MobileNetV3-SSD region detector suitable for GPU-free robotic deployment; also defined the downstream segmentation and active-learning strategy for defect localization and interpretability.
• Built an operator-assistance tool for laboratory evaluation of sun-care products, enabling faster and more consistent assessments and receiving qualitative validation from users.
• Worked closely with laboratory operators and multidisciplinary research teams to translate experimental protocols into maintainable software, supporting adoption and alignment with real workflows.

Impact

Standardized measurements, reduced operator dependence, accelerated product comparisons, and established reusable computer-vision components for laboratory automation and resource-constrained robotics.

What I did

• Audited 11,000 scanned and photographed documents, mapping blur, glare, shadows, cropping, and complex backgrounds while capturing business and regulatory requirements. Manually annotated a balanced 3,000-image MRZ/background subset selected to represent these failure modes.
• Fine-tuned six YOLOv7 variants for MRZ detection using MS COCO transfer learning, testing SGD and Adam, learning-rate schedules, batch sizes, and frozen layers. Adapted the training code to run on Djezzy’s CPU-only Red Hat environment because confidential data could not leave company infrastructure.
• Compared autoencoder, histogram equalization, CLAHE, Retinex, and K-means enhancement strategies. Rejected the autoencoder after observing altered character content and selected a fidelity-preserving Retinex, erosion, and Otsu-thresholding pipeline.
• Added Hough-based deskewing and evaluated orientation correction with a custom CNN, VGG16 transfer learning, and PaddleOCR. Applying PaddleOCR at individual-line level generalized substantially better than classifying the complete MRZ.
• Evaluated U-Net, DBNet, and EAST for text-line detection. Retained EAST with bounding-box union refinement after it outperformed the K-means refinement strategy.
• Developed four CRNN-CTC recognizers. The experiments progressed from 50,000 synthetic images to transfer learning on 7,000 augmented real and 3,000 synthetic samples, then to a 320,000-image mixed training run; results were benchmarked against EasyOCR and PaddleOCR.
• Built a modular PyQt application for batch and on-demand processing. Complete 30-character MRZ lines were parsed into structured fields, while truncated outputs were retained in JSON for review and downstream database handling.

Evaluation highlights

• The selected YOLOv7 configuration reached 0.988 precision, 0.95 recall, 0.9812 mAP@0.5, and 0.8088 mAP@0.5:0.95.
• Deskewing correctly aligned 323 of 330 evaluated images (97.8%). PaddleOCR line-level orientation classification reached 87% accuracy and F1 score.
• EAST with union refinement detected all three MRZ lines correctly on 99.09% of 330 images, compared with 90.3% using K-means refinement.
• The best custom CRNN reached 3.56% CER on the mixed dataset and 9.85% on unseen samples. PaddleOCR reached 4.61% and 5.77% respectively, demonstrating stronger generalization.
• The complete five-module pipeline extracted identity-document information successfully in 70% of a 100-image end-to-end evaluation sample.

Impact

• Improved data quality and traceability upstream of core systems while reducing manual-entry errors and accelerating KYC controls.
• Standardized the extraction, parsing, and validation of MRZ fields to support ARPCE compliance.

Academic result

High Distinction — 17/20.

Scope of the study

• Established the distinction between 2D and 3D pose estimation and compared top-down and bottom-up pipelines for detecting, associating, and tracking semantic keypoints.
• Reviewed both human and object pose-estimation approaches, including OpenPose, BlazePose, PoseCNN, 6D-VNet, and LatentFusion.
• Analyzed how PoseCNN combines semantic labeling with 3D translation and rotation estimation for end-to-end 6-DoF pose prediction, and how 6D-VNet couples Mask R-CNN features with rotation and translation branches for monocular RGB inputs.

LatentFusion analysis

• Conducted a detailed architectural review of LatentFusion for pose estimation of previously unseen objects from a small set of RGB-D reference views.
• Traced the pipeline from multi-view observations to a canonical latent feature volume, followed by 3D and 2D U-Net decoding for depth, segmentation, and image-based rendering.
• Studied its two-stage inference strategy: forward inference for a coarse pose, then gradient-based refinement of rotation, translation, and viewport parameters with the Adam optimizer.
• Examined published ADD and ADD-S benchmark results on LINEMOD and MOPED, where LatentFusion remained competitive without object-specific pretraining and outperformed PoseRBPF on most reported MOPED objects.

Recommendation

Identified LatentFusion as a strong candidate for the target augmented-reality context because it can estimate the pose of new objects without per-object retraining and requires only a limited number of reference views, supporting a flexible deployment workflow.