El Centro de Inteligencia Artificial y Neurociencia de la Universidad Torcuato Di Tella (CIAN Di Tella) tiene el agrado de invitar a la comunidad a un seminario especializado a cargo de dos referentes en el campo de la visión por computadora: Victoria Fernández Abrevaya y Achuta Kadambi.

El seminario contará con dos bloques de exposiciones:

"Neural approaches for inverse graphics", por Victoria Fernández Abrevaya.

Abstract
Traditional computer graphics focus on generating images from 3D assets, where the scene is defined by each object’s geometric representation, material properties, and the overall illumination. The opposite task, known as inverse graphics, involves recovering these properties from a single, natural image. The field of inverse graphics has recently made impressive progress, thanks to novel neural techniques ranging from implicit representations and neural rendering, to the use of foundational models for text and text-to-image (e.g Stable Diffusion). In this talk, I will present two of our recent works on the topic of inverse graphics. First, I will discuss an approach for generating a high-quality avatar of a person's head from a single monocular video. This method combines classic rendering techniques with novel neural approaches to produce a detailed geometry of the person’s face, along with an animation model and material properties. The resulting avatar can be manipulated to generate novel poses and inserted seamlessly into new scenes, demonstrating the potential of hybrid approaches in achieving realistic digital humans.Second, I will present a work that leverages large-language models to reconstruct the objects within a scene. By analyzing a single image, our method generates pseudo-code that describes the discrete properties, locations, and orientations of each object primitive. Although the results focus on relatively simple experimental setups, they illustrate the remarkable potential of foundational models to tackle traditionally complex problems in scene understanding and reconstruction.

Victoria Fernández Abrevaya es investigadora postdoctoral en el Max Planck Institute for Intelligent Systems, Alemania, en el grupo del Dr. Michael J. Black. Sus temas de investigación se centran en la visión por computadora 3D, y en el modelado y análisis de rostros tridimensionales. Antes de eso, realizó su doctorado en Inria, Grenoble, donde trabajó en modelos generativos 3D para el rostro. Victoria es egresada de la Universidad de Buenos Aires, donde estudió Ciencias de la Computación.

Website: https://is.mpg.de/~vabrevaya

"Toward ´Sensing for All´ in Computer Vision and Graphics Pipelines", por Achuta Kadambi.

Abstract
Vision pipelines do not work for everyone. Today, billions of light-based medical sensors are used by hospitals to measure quantities like blood flow, temperature, oxygenation and more. Unfortunately, the accuracy of light-based devices varies across demographics. Just as a soap dispenser may not always work for those with dark skin, a light-based medical device has fundamental challenges with signal-to-noise (SNR) ratio, and measurement accuracy. To address this problem, we need to rethink the sensing process, a. We draw from a paradigm of “equitable computational imaging”, where we co-design the optical sensing and computation layers to resist bias. Removing biases in both hardware and software, attacks the root causes of bias at the physical layer (e.g. light-based reflections). We will discuss new types of equitable imaging systems that measure heart rate and blood volume (contact-free and wirelessly); synthetic data pipelines that model melanin content; and theoretical results on dataset composition. 

Achuta Kadambi Achuta Kadambi (PhD, MIT ‘18) es Profesor Asociado en la Universidad de California en Los Ángeles (UCLA) en Ingeniería Eléctrica y Ciencias de la Computación. Lidera un grupo de investigación en la intersección de visión por computadora, física, procesamiento digital de imágenes, inteligencia artificial y dispositivos médicos. Ha recibido numerosos premios, incluidos el NSF (CAREER), DARPA (YFA), ARO (YIP), IEEE (premio HKN para menores de 35 años) y Forbes (30 menores de 30). Ha cofundado dos empresas en California para comercializar tecnologías de investigación. Una de ellas, Akasha Imaging, fue adquirida por Alphabet por su tecnología de automatización robótica. La otra es Vayu Robotics, que trabaja en la navegación de robots impulsados por la percepción. Kadambi ha coescrito recientemente un libro de texto de acceso abierto sobre Procesamiento Computacional de imágenes.                                                           

Website: https://www.ee.ucla.edu/achuta-kadambi/

*El seminario se dictará en idiomas inglés, sin traducción simultánea.