Nathanael Bloch

I was responsible for designing and constructing an AI face tracker specifically tailored for infrared images based on deep neural networks.

This face tracker became a vital component of the company product I was working for. The face tracker had the capability to transform 2D videos into real-time 3D face representations, providing essential information such as bounding box, landmarks, head pose and position, as well as 3DMM coefficients. Notably, this new model achieved an accuracy improvement of up to 200% compared to the previous version.

I proposed, designed, and developed a 3D image augmentation system that revolutionized the generation of synthetic face data. This system addressed the challenges of data scarcity and non-uniformity by synthesizing a wide range of head poses, expressions, and illuminations from a single real picture. Additionally, it had the capability to add 3D masks and glasses to the synthesized images realistically.

To achieve this, I created and implemented various mathematical models for optimization and rendering accelerated by GPU. The system demonstrated exceptional performance, being approximately 100 times faster than existing solutions available in the market.

I enhanced the performance of an existing face recognition system through several key improvements. I fine-tuned the loss function, the system architecture, and the inference algorithm and increased the dataset size. These enhancements generated a threefold improvement in the false positive rate, making the system more reliable and effective in face recognition tasks.

A Python repository that showcases how to use synthetic data and apply it to several use cases. This includes dataset analysis, small network training, head pose, eye gaze, face landmarks, body key points, camera configuration, and segmentation use cases.

I ideated and implemented the first version of Datagen SDK, which provides a layer of abstraction for the data and lets the user seamlessly load an existing synthetic dataset into Python.

All the parsing and loading functions were implemented behind the scenes, and each data point was structured into a simple and easy-to-access Python object.

The SDK was backward compatible with any data version.

This project aimed to take an existing synthetic face image with all its modalities—normals, depth, segmentation, and camera position—and re-render it into a realistically-looking infrared image. This infrared image could be trained to train neural networks and improve their performance in the target domain.

The modeling was based on computer graphics and 3D principles, making the image look close to the target domain and configurable to the customer's needs.

This project was highly relevant to customers working in the infrared domain since infrared data is very hard to obtain.

This project aimed to take synthetic images of the same scene from different camera positions and turn them into a 3D point cloud based on camera intrinsic and extrinsic parameters.

It provided a significant added value to customers by letting them train their neural networks on supervised 3D reconstruction, which was impossible without the 3D ground truth.

I built a fully-fledged C++ simulator based on OMNet++. It accurately models a complete network stack, from the network interface controller to the switches and software protocols.

The project required a high proficiency in C++ programming and debugging. A thorough understanding of complex protocols such as TCP and RoCE and the hardware at the register level was also needed.

This project's goal was to artificially increase the resolution of ultrasound medical devices using super-resolution and image-to-image techniques.

A neural network based on the encoder-decoder architecture remapped the raw data from the ultrasound device into a high-resolution image. It enabled a fourfold increase in the original number of pixels while keeping the original image quality so physicians could use smaller, cheaper ultrasound devices.