Giunchi, Daniele; (2021) Application of Machine Learning within Visual Content Production. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text FINAL_Daniele_Giunchi_PhD_Thesis.pdf - Accepted Version Download (52MB) | Preview

Abstract

We are living in an era where digital content is being produced at a dazzling pace. The heterogeneity of contents and contexts is so varied that a numerous amount of applications have been created to respond to people and market demands. The visual content production pipeline is the generalisation of the process that allows a content editor to create and evaluate their product, such as a video, an image, a 3D model, etc. Such data is then displayed on one or more devices such as TVs, PC monitors, virtual reality head-mounted displays, tablets, mobiles, or even smartwatches. Content creation can be simple as clicking a button to film a video and then share it into a social network, or complex as managing a dense user interface full of parameters by using keyboard and mouse to generate a realistic 3D model for a VR game. In this second example, such sophistication results in a steep learning curve for beginner-level users. In contrast, expert users regularly need to refine their skills via expensive lessons, time-consuming tutorials, or experience. Thus, user interaction plays an essential role in the diffusion of content creation software, primarily when it is targeted to untrained people. In particular, with the fast spread of virtual reality devices into the consumer market, new opportunities for designing reliable and intuitive interfaces have been created. Such new interactions need to take a step beyond the point and click interaction typical of the 2D desktop environment. The interactions need to be smart, intuitive and reliable, to interpret 3D gestures and therefore, more accurate algorithms are needed to recognise patterns. In recent years, machine learning and in particular deep learning have achieved outstanding results in many branches of computer science, such as computer graphics and human-computer interface, outperforming algorithms that were considered state of the art, however, there are only fleeting efforts to translate this into virtual reality. In this thesis, we seek to apply and take advantage of deep learning models to two different content production pipeline areas embracing the following subjects of interest: advanced methods for user interaction and visual quality assessment. First, we focus on 3D sketching to retrieve models from an extensive database of complex geometries and textures, while the user is immersed in a virtual environment. We explore both 2D and 3D strokes as tools for model retrieval in VR. Therefore, we implement a novel system for improving accuracy in searching for a 3D model. We contribute an efficient method to describe models through 3D sketch via an iterative descriptor generation, focusing both on accuracy and user experience. To evaluate it, we design a user study to compare different interactions for sketch generation. Second, we explore the combination of sketch input and vocal description to correct and fine-tune the search for 3D models in a database containing fine-grained variation. We analyse sketch and speech queries, identifying a way to incorporate both of them into our system's interaction loop. Third, in the context of the visual content production pipeline, we present a detailed study of visual metrics. We propose a novel method for detecting rendering-based artefacts in images. It exploits analogous deep learning algorithms used when extracting features from sketches.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as JSONCSVXML

Downloads by country - last 12 months

Export as XMLCSVJSON

Archive Staff Only

View Item