Crack detection in marble images using semantic segmentation

Improved performance of models by 40%, by improving the data labeling and organization process

Design and implementation of Reinforcement Learning environment for the autonomous field coverage with one or multiple drones (Parrot Anafi Drone)

• Texture image classification using ML

• Improved classification rate by 30%, by applying transfer learning and feature aggregation techniques

Research and development of an algorithm for the classification of marble tiles according to their according to their aesthetic criteria. The research was consisted of 3 different approaches:

1. Extraction of Local Binary Patterns and training of conventional machine learning models (related article)

2. Transfer learning using Tensorflow on pre-trained deep Convolutional Neural Networks (related article)

3. Metric learning using PyTorch and PyTorch Metric Learning and pre-trained deep Convolutional Neural Networks (related article)

Independent development of the software, as well as proposition of the different approaches. The software for the first approach heavily utilized parallelization techniques and data management.

Role focused on developing an Augmented Reality app using Unity3D and Vuforia SDK and applying Reinforcement Learning to train autonomous agents to shape their behaviors in fighting scenarios inside an arena.

• Quickly adapted to the new environment and familiarized myself with the new technologies of AR development

• Application of Reinforcement Learning to shape the behavior of autonomous agents in fighting scenarios

Worked in multidisciplinary teams for the development of software for the social robot NAO, to interact autonomously (or with minor interventions from the therapist) with children in special education (autism and learning disabilities). Research and development of A.I. algorithms for action recognition using computer vision. Applied machine and deep learning for engagement detection and engagement level estimation from data collected

• Initiative allowed us to collect data during the scenarios and conduct studies for engagement detection and engagement level estimation

• Responsibilities extended to the managing and maintaining of the software of the whole project

• Adapted to the new environment and learned the new technologies for the development of the required scenarios for the robot early on

• Application of ML for engagement detection (93% F1-score) and engagement level estimation (2.92% MSE)

• Took the lead of the software development process of the whole project

• Quickly adapted and learned new technologies

• Development of an automatic dartboard game scoring system using single-board computers

• Improved accuracy and speed of the calibration system by 50%, by improving the developed algorithms

• Improved accuracy of the detection system to 96%, by applying pre- and post-processing algorithms

The system was consisted of:

1. The scorer: consisted of 4 cameras, their dynamic detection and calibration of the dartboard and detection of the dart and the area it hit inside the board.

2. The player and play validation: consisted of a single camera that detected the oche and validated the player and their position during the throw.

• Worked within a team of 5 people (including me)

• Optimized code to a high degree, applying parallelization techniques and array operations, reaching a satisfactory performance on a micro-computer.