Gabriel Nascarella Hishida

gabrielnhn @ duck . com

LinkedIn/gabriel-hishida

GitHub/gabrielnhn

I'm a masters student in Data Science and Artificial Intelligence at Delft University of Technology 🇳🇱.
My interests lie in computer vision, robotics and deep learning.

Before this, I graduated with honours in BSc Computer Science at UFPR 🇧🇷,
and was doing research on ML/CNNs/GANs in the Laboratory of Vision, Robotics and Imaging (VRI).

“Blurred Lines: Does Blurring preserve privacy?”

Does Gaussian Blur protect your privacy? In this HackMD blog post, my colleagues and I try to identify for a normal (sharp) image and another blurred image whether they depict the same person by applying triplet loss on two distinct CNNs.

”Using Conditional Generative Adversarial Networks for Data Augmentation in Pneumothorax Classification”

Experimented on GAN architectures in order to generate lung images from poor segmentation inputs for pneumothorax classification. Several experiments on how to integrate positional encoding in Generative Adversarial Networks to generate consistent but dissimilar results. In the end we produce lung images, but unfortunately the output images didn't grasp the concept of pneumothorax. (PDF, GitLab)

“Look where you’re going: Classifying drivers' attention through 3D gaze estimation".

2023 BSc thesis. Trained simple ML models used in a pipeline after a Gaze Estimation model to classify the output vector as 'distracted' or not. (PDF | GitHub Repo)

“Camera-based line follower”

I was part of a robotics team, Yapira UFPR .
We received the Innovation Award for our image-based line follower in RCX 2023. I've implemented many handcrafted methods to robustly identify the track under inconsistent and adverse lighting condition / framerate.

“VRI-GazeNet: 3D Gaze Estimation for Real-time Applications" LARS 2023

Trained a MobileNetv2 on the Gaze360 dataset. Achieved decent results using a relatively small architecture. PDF | GitHub Repo

“Photomosaic from dominant patch color”

Replace image patches with other 20x20 .ppm tiles based on dominant colour. GitHub