Deep Probabilistic Models for Attribute Learning

Lead Research Organisation: Imperial College London
Department Name: Dept of Computing

Abstract

The current deep learning methodologies lack of two capabilities (a) a rigorous probabilistic methodology that describes the architecture and (b) a rigorous way of incorporating several attributes. This PhD aims to tackle the two above challenges.

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/N509486/1 01/10/2016 31/03/2022
1792614 Studentship EP/N509486/1 01/10/2016 31/03/2020 Stylianos Moschoglou
 
Description My work so far can be summarised in four distinct parts. In the first part, I will describe the work I have accomplished in the so-called area of probabilistic component analysis, in the second part, the work I have carried out in the so-called robust component analysis area, in the third part the work I have done with regards to dataset collection and introduction of new benchmarks in the literature and finally in the fourth part I will describe what I have accomplished in the field of deep learning with applications focused on 3D data generation, translation, and representation.

Part I:
One of the most important problems that researchers in the Computer Vision field need to tackle is with regards to analysing data in three-dimensions (3D). To this end, I introduced Multi-Attribute Probabilistic Linear Discriminant Analysis (MAPLDA), the first method in the literature that is able to carry out classification tasks in 3D data annotated with regards to multiple attributes (such as age, expression, etc.). As I showed in various experiments, MAPLDA achieves state-of-the-art performance against other methods.
Moreover, I introduced a robust method to rigorously initialise probabilistic variants of Linear Discriminant Analysis (such as Probabilistic LDA and MAPLDA). As I demonstrated in a number of experiments, this initialisation methodology outperforms other heuristic methods that are used in the literature in order to initialise the parameters of the aforementioned algorithms.

Part II: I introduced Multi-Attribute Robust Component Analysis (MA-RCA), the first robust method in the literature that is able to accurately reconstruct 2D textures with largely missing parts or noisy corruptions. As I demonstrated in experiments such as image denoising or image completion, MA-RCA performs much better compared to other methodologies.

Part III: We manually collected, annotated, curated and made publicly available the first age database with accurate to the year age annotations. The so-called AgeDB is the first noise-free introduced benchmark to evaluate algorithms in tasks such as age-invariant face recognition and age-group classification. Moreover, apart from pre-processing and curating the age database, this work also provided me with the chance to enhance my skills in deep learning methods, as I had to evaluate AgeDB utilising state-of-the-art deep learning methods tailored for the previously mentioned tasks.

Part IV: Even though deep learning methods have been successfully applied in 2D image analysis tasks (e.g., image generation, image classification, etc.), they cannot be directly applied in non-Euclidean domains such as 3D data (graphs/manifolds). To overcome this, we introduced 3DFaceGAN, a deep learning methodology tailored for 3D facial data manipulation (e.g. 3D data generation, translation, and representation). More specifically, by utilising 3DFaceGAN, one can generate new 3D facial data varying in expressions, age, ethnic-groups etc. Also, 3DFaceGAN can be used on 3D facial data to transfer the expressions they have (for example, if we have a 3D object of a person in a certain expression, e.g., neutral, we can use 3DFaceGAN to transfer this 3D face in a state of happiness or surprise). This work has already been submitted to the International Journal of Computer Vision (IJCV). We have received the first round of reviews which were only minor corrections, we submitted the rebuttal and we are now waiting for the final decision.
Exploitation Route * AgeDB, introduced back in 2017, still remains the only manually collected age database and is widely used amongst the researchers in the field. Even though it is a very good benchmark for age-related tasks, the size of the database is relatively small (circa 17,000 images). AgeDB can be further expanded, by including images originating from more diverse backgrounds (ethnicities, ages, etc.). Moreover, AgeDB should be evaluated in the more recent algorithms and thus the related protocols introduced in the original paper should be updated. This would undoubtedly further strengthen AgeDB as a benchmark.

* Even though MAPLDA achieves state-of-the performance in 3D facial analysis tasks, all of the introduced methods in the literature so far need the 3D data to be pre-processed beforehand (i.e. data need to be registered first utilising a template). It would be very interesting to explore how such methods or variants of them can be modified in order to be used in raw 3D data, without any pre-processing.
Sectors Aerospace, Defence and Marine,Electronics,Financial Services, and Management Consultancy,Healthcare,Security and Diplomacy

URL http://www.moschoglou.com
 
Description In co-operation with the Royal Academy of Engineering (RAEng), for ThisIsEngineering day (https://www.thisisengineering.org.uk/), I undertook a project where the main task was to find out how engineers are being represented online in picture format. To do this, I used an AI algorithm (in particular, a custom Generative Adversarial Network that I developed) to build a picture from images of engineers found on online search engines to see what an engineer looks like. The main finding was that the image of an engineer, made by the AI algorithm, was of a white man wearing a hard hat. The finding of this project was used in a campaign by the RAEng whose main objective was to promote engineering studies in the UK. More info can be found here https://www.imperial.ac.uk/news/193848/what-does-average-engineer-look-like/
First Year Of Impact 2019
Sector Education
Impact Types Cultural,Societal

 
Title AgeDB: the first manually collected, in-the-wild age database 
Description Over the last few years, increased interest has arisen with respect to age-related tasks in the Computer Vision community. As a result, several" in-the-wild" databases annotated with respect to the age attribute became available in the literature. Nevertheless, one major drawback of these databases is that they are semi-automatically collected and annotated and thus they contain noisy labels. Therefore, the algorithms that are evaluated in such databases are prone to noisy estimates. In order to overcome such drawbacks, we present in this paper the first, to the best of knowledge, manually collected" in-the-wild" age database, dubbed AgeDB, containing images annotated with accurate to the year, noise-free labels. As demonstrated by a series of experiments utilizing state-of-the-art algorithms, this unique property renders AgeDB suitable when performing experiments on age-invariant face verification, age estimation and face age progression" in-the-wild". The original paper can be accessed from the following link: http://openaccess.thecvf.com/content_cvpr_2017_workshops/w33/papers/Moschoglou_AgeDB_The_First_CVPR_2017_paper.pdf The license as well as the dataset can be accessed from the following link: https://ibug.doc.ic.ac.uk/resources/agedb/ DOI: http://dx.doi.org/10.1109/CVPRW.2017.250 
Type Of Material Database/Collection of data 
Year Produced 2017 
Provided To Others? Yes  
Impact AgeDB is the first noise-free introduced database/benchmark to evaluate algorithms in tasks such as age-invariant face recognition and age-group classification. It is widely used by researchers in the field and is amongst the standard protocols utilised to benchmark state-of-the-art algorithms in the aforementioned tasks. 
URL https://ibug.doc.ic.ac.uk/resources/agedb/