― About
I'm a Telecommunications and Electronics Engineer with a Master’s Degree in Artificial Intelligence, Pattern Recognition and Digital Imaging.
Deeply interested in Image Processing and Computer Vision, which is what I have been doing for the past 7 years, in academia and commercial businesses, in Cuba, Spain and Germany.
Continually embracing opportunities for intellectual growth and welcoming challenges with enthusiasm.
― Skills
Communication and team work
Teaching Research Computer Vision Machine Learning Linux Python
Software Development
― Work experience
Computer Vision Engineer
Tyris AI / Jan 2021 - Aug 2021
Quality inspection for pavement and traffic signs in highway, namely crack segmentation, pothole detection and measuring retro reflectance level of vertical and horizontal signals.
Real time detection of patients’ and doctors’ faces and texts in screens in surgical operations to provide anonymization.
― Education
Universitat Politècnica de València (UPV)
Master's Degree in Artificial Intelligence, Pattern Recognition and Digital Imaging / 2018 - 2019 / Passed with 9.4 / 10.0
Universidad Tecnológica de La Habana, CUJAE
Engineer's degree in Telecommunications & Electronics / 2012 - 2017 / Passed with 4.72 / 5.0
Lead Computer Vision Engineer
AI Superior / Jan 2023 - Present
Senior Computer Vision Engineer
AI Superior / Sep 2021 - Jan 2023
Geo-spatial analytics for different types of objects such as debris, plants, buildings, vehicles, roof planes and others using satellite and drone images.
Image to image translation with paired and unpaired medical hyper-spectral data.
Multiple projects involving precise detection and fine grained segmentation of objects in different scenarios.
Technical interviewer and junior and mid engineers mentoring.
Machine Learning | Computer Vision Engineer
Solver Intelligent Analytics / May 2020 - Jan 2021
Anomaly detection and localization in images of electric towers with unannotated datasets.
EDA on tabular data for multiple PoCs.
Development of a forecaster for retail in supermarkets.
Machine Learning Engineer
Everon Quantitative Analysis / Sep 2019 - May 2020
EDA on financial data.
Algorithmic trading programming.
Research And Development Associate
Universitat Politècnica de València (UPV) / Oct 2018 - Aug 2019
Investigation, research and project development at GTIA research group.
Mostly involved in MultiAgent Systems and simple computer graphics with Unity.
Graduate Teaching Assistant
Universidad de las Ciencias Informáticas / Nov 2017 - Aug 2018
Introduction to Computer Networks.
Introduction to Basic Electronics.
Computer Vision Researcher And Development
University of Havana / Jan 2017 - Jul 2018
Associate Computer Vision investigator at Grupo de Sistemas Complejos y Física Estadística of the Physics Faculty. Worked on precise tracking of millimetric sized insects with static and dynamic camera setups.
Computer Vision and Machine Learning are not just fields of study or work for me, they are hobbies that fuel my curiosity.
― Solar Panel Installation
A solar energy company needed to accelerate the process of analyzing residential roofs for solar panel installations. The traditional manual methods employed were time consuming, making it necessary to find an automated and precise solution that could provide reliable and efficient roof analysis. The developed solution can automatically identify the boundaries of roofs and segment them into distinct areas, providing accurate measurements for each roof plane, which is essential for determining the optimal placement of solar panels. Below are some roof segments predicted by the AI model and measurements of each side.
Some of the professional projects I have lead
― Geospatial analysis with drone and satellite images
Trained and deployed custom AI models for gaining geospatial insights in different use cases such as agricultural, city municipality, construction sites and others. Here some results for getting areas of buildings in an entire city, detecting vehicles parked in forbidden areas, estimating materials left over in construction sites and monitoring amount of rocks in dikes.
A research department of the pharmaceutical company for which the virtual stain transfer project mentioned earlier was developed, was interested in performing image translation on hyper-spectral images. The acquired hyper-spectral cubes were some times as big as 330 GB, since they had up to 50 channels of information at different wavelengths and could have over 100k x 100k pixel matrices. Below are some examples of different real immunofluorescence channels and fake generated ones from their corresponding autofluorescence. The large image is a real whole scale image and a fake one.
― Hyper-Spectral virtual stain transfer
― Cleanliness in the work place
― smart bin recycling
A company which facilitates insights in multiple areas of Soft Facilities Management (SFM), including cleaning, room occupancy, space utilization, temperature and others required a machine vision system to detect when different parts in the office where clean or not, and needed maintenance. A custom AI model was developed to meet this request.
The goal of this project was to develop a smart vision system for detecting the different types of materials thrown in a recycling bin. Additionally, the brands that could be detected should be reported, for which an image embedding retrieval system using vector databases was setup to accommodate for scalability for new brands. Below are some results on materials and brands detected.
― VIrtual stain transfer
A pharmaceutical company needed to generate virtual (fake) stained histopathology images from real slices of tissue that were stained with different reactants. This would allow them to use AI models, for example cancer detection models, that were trained with tissues stained on one specific type of reactant to work with tissues stained with different types of reactant, by making predictions on the virtually stained counterpart, rather than retraining the existing AI model or developing a new one. Multiple SOTA models suitable for this unpaired image to image translation were trained and the fake slices generated by them were analyzed. The results obtained were promoted to be published and a paper reflecting the work was published at Medical Imaging with Deep Learning. The code was released and is available here. Below are some examples of real tissues and their virtually stained pairs generated by one of the AI models.
― Proptosis Assesment
A web app was developed to accurately assess eyeball positioning and identify abnormalities related to proptosis, a condition characterized by the forward displacement of the eye. By obtaining very precise segmentations of the eye and the eyebrow key metrics relevant to the ophthalmological domain could be computed. Some of these are:
Margin Reflex Distance 1 (MRD1) and Margin Reflex Distance 2 (MRD2), which measure the distance from the centre of the pupil to the upper and lower eyelid margins, respectively.
Eyebrow Height, which determines the distance between the centre of the pupil and the eyebrow.
Proptosis distance, which gauges the distance between the centre of the pupil and the bone of the orbit on the lateral side.
Pupil displacement distances when looking in various directions.
Below are some results of the main metrics monitored on patients who suffer from proptosis.
Overlaying multiple channels we can see different structures and view them as classical RGB images.
― eye fat and muscle volume estimation
This request came from the same Ophthalmology center for which the proptosis diagnosis project was developed. An AI model was trained to estimate the volume of fat and muscle in human eyes using CT and MRI orbit scans. By analyzing these scans, the model successfully segments fat and muscle tissue in each slice, allowing for accurate volume estimation and facilitating before and after volume comparisons following interventions. Below are some results of fat and muscle areas segmented by the AI.
― Chasing Insects: A Survey of Tracking Algorithms
Published on 2017 and available at Revista Cubana de Física
The trajectories described by insects of different species in confined regions have been studied in depth. The main limitation found is the effective area covered while the insect is been tracked. Our work describes different image processing algorithms designed to capture the trajectory of an insect in a sequence of frames, and it focuses on the ones that don't depend on keeping fixed the position of the camera. Finally we propose a method for tracking insects in unconfined regions using a combination of some of the algorithms presented and a mobile camera system.
Scientific Publications
The precise and continuous tracking of millimetric-sized walkers, such as ants, is quite important in behavioral studies. However, due to technical limitations, most studies concentrate on trajectories within areas no more than 100 times bigger than the size of the walker or longer trajectories at the expense of either accuracy or continuity. Our work describes a scientific instrument designed to push the boundaries of precise and continuous motion analysis up to 1000 body lengths or more. It consists of a mobile robotic platform that uses digital image processing techniques to track the targets in real time by calculating their spatial position.
― An autonomous robot for continuous tracking of millimetric-sized walkers
Published on 2019 and available at American Institute of Physics
― A comparative evaluation of image-to-image translation methods for stain transfer in histopathology
Published on 2023 and available at Medical Imaging with Deep Learning
In our work, we compare twelve stain transfer approaches, three of which are based on traditional and nine on GAN-based image processing methods. The analysis relies on complementary quantitative measures for the quality of image translation, the assessment of the suitability for deep learning-based tissue grading, and the visual evaluation by pathologists. Our study highlights the strengths and weaknesses of the stain transfer approaches, thereby allowing a rational choice of the underlying I2I algorithms.