Banca de DEFESA: MANOLO CANALES CUBA

Uma banca de DEFESA de DOUTORADO foi cadastrada pelo programa.
STUDENT : MANOLO CANALES CUBA
Date: 08/12/2025
TIME: 09:00
LOCAL: https://meet.google.com/sup-yxrg-rrk
TITLE:

Flow Matching for Text-Driven 3D Human Motion Generation: Towards High-Quality, Diverse, and Natural Synthesis

PAGES: 85
BIG AREA: Ciências Exatas e da Terra
AREA: Ciência da Computação
SUBÁREA: Metodologia e Técnicas da Computação
SPECIALTY: Processamento Gráfico (Graphics)
SUMMARY:

The generation of human body motions guided by descriptive text has a wide range of applications, including robotics, augmented and virtual reality, and entertainment. However, achieving human motions that exhibit both diversity and fidelity to the input text while maintaining high naturalness and avoiding artifacts remains a challenging task. This challenge stems from the need to go beyond merely retrieving motions similar to the text or performing simple operations on recorded human movements; instead, it requires demonstrating creativity by synthesizing novel movements. Generative models based on artificial intelligence thus offer a promising approach to address this challenge; these methods, including Diffusion Models, VAEs, GANs, and their combinations, have been explored for this purpose. Despite advances, challenges have remained, including artifacts, involuntary movements, and difficulties maintaining high fidelity to the text. Some approaches suggest post-processing generated motions with filters to reduce jitter, but this often compromises quality or coherence. More recently, Flow Matching (FM) models have emerged as a promising generative approach across many domains, offering simpler mathematics and faster generation. FM trains a neural network to transform, via a vector field, a simple distribution corresponding to noisy data into the complex distribution of the training data. Motivated by their characteristics, we have adopted FM in this work. However, the high jitter caused by limited computational resources imposes challenges on our human motion generation task. To address this, we proposed a novel training and sampling strategy. This strategy allows the model to first estimate noise-free data before defining the vector field for generation, rather than simulating the vector field directly. Results demonstrate high-quality motion generation, improved diversity and naturalness (low jitter), and more efficient resource utilization. On the HumanML3D and KIT datasets, our method achieves top jitter reduction on KIT and near-state-of-the-art reduction on HumanML3D, while maintaining an optimal fidelity-naturalness balance and avoiding post-processing.

COMMITTEE MEMBERS:
Presidente - Interno ao Programa - 1672977 - JOAO PAULO GOIS
Membro Titular - Examinador(a) Interno ao Programa - 1773182 - SAUL DE CASTRO LEITE
Membro Titular - Examinador(a) Externo à Instituição - FERNANDA MIYUKI YAMADA
Membro Titular - Examinador(a) Externo à Instituição - SORAIA RAUPP MUSSE - PUC-RS
Membro Titular - Examinador(a) Externo à Instituição - PAULO ARISTARCO PAGLIOSA - UFMS
Membro Suplente - Examinador(a) Interno ao Programa - 1169328 - MATEUS COELHO SILVA
Membro Suplente - Examinador(a) Externo à Instituição - AFONSO PAIVA NETO - USP
Membro Suplente - Examinador(a) Externo à Instituição - LEONARDO KOLLER SACHT - UFSC