Controllable synthetic image datasets generation for advancements in human  pose and shape estimation

Viniavskyi, Ostap

Домівка
→
Students Research & Project Works | Роботи студентів
→
Факультет прикладних наук
→
Освітня програма наук про дані
→
2024
→
Перегляд матеріалів

dc.contributor.author	Viniavskyi, Ostap
dc.date.accessioned	2024-08-23T11:35:25Z
dc.date.available	2024-08-23T11:35:25Z
dc.date.issued	2024
dc.identifier.citation	Viniavskyi Ostap. Controllable synthetic image datasets generation for advancements in human pose and shape estimation. Ukrainian Catholic University, Faculty of Applied Sciences, Department of Computer Sciences. Lviv 2024, 45 p.	uk
dc.identifier.uri	https://er.ucu.edu.ua/handle/1/4679
dc.language.iso	en	uk
dc.subject	datasets generation	uk
dc.subject	Controllable synthetic image datasets generation	uk
dc.subject	advancements	uk
dc.subject	human pose	uk
dc.subject	shape estimation	uk
dc.title	Controllable synthetic image datasets generation for advancements in human pose and shape estimation	uk
dc.type	Preprint	uk
dc.status	Публікується вперше	uk
dc.description.abstracten	Human-centric applications are ubiquitous in the modern world. Myriads of edu- cational, entertainment, e-commerce, and other applications require understanding the measurements of the human body in the image. The deep-learning methods for solving human-centric problems usually rely on supervised learning approaches and need tons of labeled data to excel. Label acquisition for 3D Computer Vision tasks and specifically for the human mesh estimation task became even more difficult and error-prone compared to 2D tasks due to the inherent complexity of working with an additional dimension. That is where the synthetic data steps in, allowing researchers to obtain much more cost- efficient and pixel-perfect annotations. In this work, we utilize the existing Latent Diffusion Model for conditional image generation and create a method for synthesizing a large dataset of humans with 3D mesh labels obtained without the involvement of a human annotator. Further, we show the effectiveness of using such a synthetic dataset and its superiority compared to other synthetic data obtained from the game engines. The implementation of the proposed approach can be accessed on the GitHub	uk