Naslov
A deep learning model for estimation of human body measurements from images

Autori
Bartol, Kristijan

Vrsta, podvrsta i kategorija rada
Ocjenski radovi, doktorska disertacija

Fakultet
Fakultet elektrotehnike i računarstva

Mjesto
Zagreb

Datum
24.03

Godina
2022

Stranica
133

Mentor
Tomislav Pribanić

Ključne riječi
anthropometry, deep learning, dissertation, human pose estimation, human shape estimation, statistical body models, computer vision

Sažetak
The analysis of human body measurements and shapes is important for understanding the differences and similarities in and between populations and has many applications in medicine, surveying, the fashion industry, fitness, and entertainment. Recent advances in human body measurement and shape estimation have been significantly driven by statistical body models, created based on large 3D body scanning datasets. As a first contribution, we show that the statistical models can be used to estimate human body measurements by using only a person’s self-estimated height and weight. In order to be able to estimate body measurements for any person for which we do not have their self- estimated height and weight, we have to use a dif- ferent strategy. In this dissertation, we focus on estimating the parameters from the statistical body model based on images. The second contribution is the pose estimation model based on multiple views of the same person. The model is novel in a way that it can take any calibrated set of cameras and estimate feasible 3D human pose, which was not possible in prior work. The third and final contribution is the shape and clothes estimation model from a single image. The model successfully recovers 3D human shape which also allows us to easily extract body measurements, which we describe in detail. The three described models are evaluated in detail and compared to the state-of-the-art. Finally, we compare the proposed approaches and discuss how they can be used together to obtain human body measurements from images and also by using a person’s height and weight on top, if available. The dissertation consists of seven chapters and three Appendices and is structured as fol- lows. In the first chapter of the doctoral dissertation (1. Introduction), an overview of anthro- pometry in the traditional sense, digital anthropometry and 3D scanning, as well as statistical models of the human body, on which the aforementioned 3D network of the human body is based, are given. Finally, a brief general overview of the body measurement procedure in the context of the described terms is given. The second chapter (2. Image-Based Anthropometry) introduces the concept of anthropom- etry from images. Existing methods for estimating human body measurements from images are reviewed and placed in the high-level context of anthropometry. Also, deep learning models are described, which are used to extract significant images from which one can then learn the assessment of the 3D network of the human body, that is, the 3D pose, shape and measurements of the body. In the third chapter (3. Body Measurement Estimation Baseline), a simple linear regres- sion model was proposed that, based on a person’s height and weight, estimates the remaining measurements, e.g., body lengths and girths. Moreover, the model only uses information that a person estimates for himself, in particular, body height and weight, instead of perfectly accurate data. Realistic estimates of height and weight were simulated by adding noise from a normal distribution. The proposed model achieves accuracy comparable to the best existing methods for estimating human body measurements, and is even more accurate than some deep learning models. The fourth chapter (4. Learning Body Pose Estimation from Images) describes the proposed method for 3D pose estimation from multiple views. Specifically, the input model uses images obtained from different synchronized cameras. The conditions for the proposed model to work are that there is only one person in the scene and that the person is visible from at least two views at all times. The proposed model achieves the best result when compared with competing models on images that are not obtained by the same cameras as in the training set. The fifth chapter (5. Learning Body Shape Estimation from Images) proposes a method for estimating the 3D shape of the human body from a single image. Special features of the proposed method compared to existing methods for 3D pose and shape estimation are the ability to estimate the shape of people from images in loose clothing and the ability to estimate the parameters of the clothing itself. These advantages are a significant step towards speeding up and simplifying the estimation of human features from images that do not have significant limitations. The sixth chapter (6. Discussion) discusses the overall dissertation. In particular, it com- ments on the performances of the proposed models w.r.t. body measurement estimation, namely, the baseline described in Chapter 3 and the shape estimation model described in Section 5. The chapter points out general limitations and the assumptions used during the evaluation of the proposed models. Finally, future work is discussed. The last chapter brings concluding considerations, briefly highlights the achievement of scientific contributions and presents guidelines for further scientific research on the topic of assessing the 3D pose, shape and measurements of the human body from images. The three Appendices provide additional details which were logically out of the scope of the thesis or serve as an additional source of information which are not required for the under- standing of the proposed contributions. The first Appendix provides mobile implementations (hardware and software) which came as a result of the proposed methods being implemented on or on top of the smartphone devices. The second Appendix provides more details about the particular types of 3D scanners often used for 3D human body scanning. The third Appendix provides practical recommendations on particular human body measurement applications, such as medicine, fitness, entertainment, etc

Izvorni jezik
Engleski

Znanstvena područja
Elektrotehnika, Računarstvo

POVEZANOST RADA