Development of an AI-based generation process of digital 3D garments for any posture and its dy[...]

Development of an AI-based generation process of digital 3D garments for any posture and its dynamic evolution

Réf ABG-127379

Sujet de Thèse
04/12/2024

Financement public/privé
Ecole Nationale Supérieure des Art et Industries Textiles
Lieu de travail: Roubaix - Les Hauts de France - France

Champs scientifiques:

• Sciences de l’ingénieur

Mots clés: 3D digital garments, image analysis and generation, AI, deep learning, text/image relationship exploitation

Description du sujet

This PhD study aims at exploiting relations between a 3D digital garment representation, its geometric structure, technical parameters and annotated semantics in order to develop an automatic generation process of digital garments for any posture and its dynamic evolution (with and without fitting on human bodies).

Deep neural models heavily depend on datasets, which is also the case here. While there exist several datasets that are publicly available for static and dynamic garment modeling, most of them are for garments that are either in their canonical forms (i.e. reference or template shapes prior to any deformation) or draped/worn on human bodies. The project will propose new datasets in this context, comprising both 4D (3D+time) digital and 2D real-world video datasets. In particular, the digital dataset requires the integration of professional expertise on fabric physical properties and construction of garment patterns. For this purpose, we plan to use garment 3D CAD software (e.g. Style3D, Clo3D, Modaris 3D Fit), which are equipped with several datasets of representative digital fabrics and garments. To optimize efficiency, we will initially leverage existing 3D digital datasets, combined with strategic methods for obtaining new data. Such strategies include: guidance of the initial part labeling by a devoted mesh segmentation network model, and scripted simulation sessions with several predefined external forces and material properties. In order to narrow down the gap between real-world videos and the rendered videos from digital data, we intend to generate new garment models with enhanced realism by recursively running a cycle of garment model generation – 3D garment demonstration and sensory evaluation – knowledge-based garment parameters adjustment until satisfaction of the user. This addresses a limitation in the current digital datasets, which lack elements like collar, cuff, buttons, pocket, etc. Note that all these are important elements for the semantic labeling of clothes in a real-world context.

Main Activities:

1) Digital dataset construction towards the First Dataset

A first dataset composed of a few garments and a few motions, typically between 5 garment types combined with 4-5 material properties and 4-5 environments, will be simulated and processed to provide an initial dataset for the project.

Generating the 3D digital garments corresponding to their real prototypes by using the selected 3D simulation software and collecting them in the digital garment dataset. The technical parameters of the 3D digital garments (usually incomplete or inaccurate for real garments) will by repeatedly optimized by following the cycle of 3D garment demonstration – sensory comparison with pictures and videos of the real garments – garment parameters adjustment based on the garment knowledge model. A sensory evaluation process will be defined for quantitatively identifying the difference of the digital and real garment on each garment geometric key position. A garment knowledge model will be set up for identifying relations of annotated semantic parts of each garment, their geometric positions and related technical parameters (fabrics, patterns).

2) Digital dataset acquisition towards the Second Dataset

The first dataset will be extended with new garments and new motions. After the deployment by partners, it will be made available to the related scientific communities. The first real and digital garment datasets will be extended to other garment types (shirt, leggings, …) and other causal postures. The garment knowledge model will also be updated by introducing relations of geometric key positions on the garment at causal postures and annotated semantic causal postures.

Prise de fonction :

06/01/2025

Nature du financement

Financement public/privé

Précisions sur le financement

Projet GarSEM (ANR 2024-2028)

Présentation établissement et labo d'accueil

The ENSAIT Textile Engineer School is the largest textile higher education and research institution in France and Europe. It trains about 70% of engineers and managers in French textile/fashion companies.

The GEMTEX Laboratory of ENSAIT is an interdisciplinary national research unit specialized in textile engineering and materials as well as fashion digitalization. It is currently working on more than 20 research collaborative projects funded by the European Commission, French government and industrial companies. The main research themes of GEMTEX concerned by this project are: (i) computerized fashion product design and (ii) AI-based textile modeling and optimization.

Site web :

http://www.gemtex.fr

Intitulé du doctorat

Doctorat d'Automatique et de Traitement du Signal

Pays d'obtention du doctorat

France

Etablissement délivrant le doctorat

Université de Lille

Ecole doctorale

Sciences pour l'ingénieur (SPI)

Profil du candidat

Requested training / diploma: Master of Research on AI/Data Science/Image Analysis

Requested academic competences: Artificial intelligence, data mining, basic knowledge on garments

Requested operational know-how: basic techniques on computer programming and 3D image/video analysis and synthesis

Date limite de candidature: 01/07/2025