Synthesizing and Editing Photo-Realistic Visual Objects

Lead Research Organisation: University College London
Department Name: Computer Science

Abstract

Current computer graphics techniques allow us to render almost any object at near photo-realistic quality. However, the standard approach necessitates that the user painstakingly specifies all aspects of the geometric and material properties of the object. This is time-consuming and needs skilled human operators. It is hard to edit the resulting models at anything other than the low level of geometry and materials at which they are specified. Moreover, we cannot edit real photographs without reverse engineering the underlying model and this is very difficult.In this proposal we investigate a radically different pipeline for computer graphics that will allow non-experts to rapidly create and edit photo-realistic two dimensional images of objects. The crux of our approach is to provide the computer with a deeper understanding of the class of objects under consideration. This knowledge (which takes the form of a statistical model) is then leveraged to help the user achieve their goals more easily. The impact of this project is potentially enormous. Such a technology could become a standard tool installed on every home and business computer. Some of the many potential applications are:- Conceptual design. Manufacturing industries often need to sketch new product ideas and refine existing designs. Our system could help a fashion designer produce and manipulate photo-realistic images of new garments.- Clipart objects. Stock images are required for on-line and real-world publishing and these are often sought via search engines (e.g. Google Images). However, the returned results are often not ideal and may be subject to copyright. Our approach will allow the user to design bespoke images to exactly their specifications.- Photo and movie editing. Digital editing of images and movies is commonplace, but requires considerable skill. Our techniques could be used to modify facial expressions in portrait photography or apply digital cosmetics in movie post-production.- Content for virtual worlds. The trend towards larger 'sandbox' environments in video games has created an explosive demand for graphical content. Our system could allow automated or semi-automated creation of photorealistic building facades for a large virtual environment.

Planned Impact

The proposed work has a wide scope and if successful the results of this project will become part of a standard tool installed on every home and business computer. We expect that the project will benefit the following groups in particular. 1. The UK economy - Commercial exploitation of this project has potential to create wealth and we are actively exploring business opportunities. 2. Graphics and games communities - The project will decrease both the time and skill required to create images of objects. This will potentially save these industries enormous sums of money, and allow them to achieve results faster. Moreover, the methods will enable new creative possibilities. 3. Film and photography communities - the ability to change the characteristics of objects in existing images will be of interest to commercial photographers and film makers. One example of this would be the ability to edit facial expressions in existing footage. Postproduction is becoming an increasing part of the film making process but is currently limited by what is practical. This work will expand those limits. 4. Design Communities - the project will provide methods that are suitable for conceptual design: it will allow photos of existing products to be easily manipulated with the goal of creating a photo-realistic image of a modified design. Examples of possible communities who might benefit are architects and the fashion industry. 5. General Public - There are a plethora of practical and artistic uses for this technology: people need images of objects to illustrate all kinds of conventional and on-line documents. Currently, they are faced with two options: (i) they can either laboriously create an image themselves or (ii) they can use Google Images or similar to find one. However, the result may not be exactly what they require and may be subject to copyright. With our technology, it will be possible to easily create bespoke images of objects for publishing.

Publications

10 25 50
publication icon
Campbell N (2014) Learning a manifold of fonts in ACM Transactions on Graphics