Efficient automated texture recognition and mapping for 3D city models


B808b42e 7994 4b7c 90e4 a944de94e65d.png 500 500

Starting Date: earliest start: (2017-11-01) latest end: (2018-07-31)

Organization: Photogrammetry and Remote Sensing (Prof. Schindler)

Involved Host(s): Berger Monique

Abstract: The purpose of dense 3D reconstruction is to obtain models that can be visualized on computers. Research on the topic focuses especially on improving the quality of the geometry in order to obtain models that are close to reality or making it faster in real-time.

Description: The purpose of dense 3D reconstruction is to obtain models that can be visualized on computers. Research on the topic focuses especially on improving the quality of the geometry in order to obtain models that are close to reality or making it faster in real-time. Recently, the introduction of semantics in the process led to even better reconstruction of scenes with missing data. Indeed, semantic allows the introduction of class-specific shape priors (e.g., the ground is flat and horizontal, while a building has vertical facades and stands on the ground…) in order to encode urban structure (c.f. Figure 1.b). Three-dimensional virtual models have a significant value for a wide range of applications such as urban planning, virtual reality, navigation or preservation of archeological monuments, where realism is one the key component. It is thus desirable to capture not only the geometry, but also the visual appearance (e.g. photo-realistic, texture...) or higher-level information (e.g. material composition …). This can be done, for example, by mapping a texture to a geometric or semantic 3D model with acquired images in order to generate a photo-realistic look. In this project, we are also interested in identifying the materials that constitute the 3D objects. For instance, a building could be made of brick, concrete or wood and we would be interested in displaying such information on the model. These materials can be first identified from texture and then directly visualized on the 3D model using texture mapping.

Goal: - State-of-the-Art for texture mapping methods of 3D models, texture and material recognition from RGB images (c.f. Figure 1.c) and even hyperspectral images. - Implementation of one of the state-of-art texture mapping methods - Implementation of a technique to identify materials from texture - Visualization of the final results: - Photo-realistic from RGB images (in the spirit of Figure 1.e) - "Material-realistic" (in the spirit of Figure 1.a) using a texture bank (e.g. brick or concrete in Figure 1.d)

Contact Details: Audrey Richard (audrey.richard@geod.baug.ethz.ch) Ian Cherabier (ian.cherabier@inf.ethz.ch)

Material/Texture Recognition Texture Mapping Segmentation 3D Models Machine Learning Computer Vision


Labels: Master Thesis CLS Student Project (MPG ETH CLS)
Topics: Information, Computing and Communication Sciences Engineering and Technology

© 2017, Copyright Max-Planck-Gesellschaft - Imprint