Photometric Reconstruction from Images: New Scenarios and Approaches for Uncontrolled Input Data.
Technische Universität, Darmstadt
[Ph.D. Thesis], (2014)
Available under Only rights of use according to UrhG.
Download (184MB) | Preview
|Item Type:||Ph.D. Thesis|
|Title:||Photometric Reconstruction from Images: New Scenarios and Approaches for Uncontrolled Input Data|
The changes in surface shading caused by varying illumination constitute an important cue to discern fine details and recognize the shape of textureless objects. Humans perform this task subconsciously, but it is challenging for a computer because several variables are unknown and intermix in the light distribution that actually reaches the eye or camera. In this work, we study algorithms and techniques to automatically recover the surface orientation and reflectance properties from multiple images of a scene.
Photometric reconstruction techniques have been investigated for decades but are still restricted to industrial applications and research laboratories. Making these techniques work on more general, uncontrolled input without specialized capture setups has to be the next step but is not yet solved. We explore the current limits of photometric shape recovery in terms of input data and propose ways to overcome some of its restrictions.
Many approaches, especially for non-Lambertian surfaces, rely on the illumination and the radiometric response function of the camera to be known. The accuracy such algorithms are able to achieve depends a lot on the quality of an a priori calibration of these parameters. We propose two techniques to estimate the position of a point light source, experimentally compare their performance with the commonly employed method, and draw conclusions which one to use in practice. We also discuss how well an absolute radiometric calibration can be performed on uncontrolled consumer images and show the application of a simple radiometric model to re-create night-time impressions from color images.
A focus of this thesis is on Internet images which are an increasingly important source of data for computer vision and graphics applications. Concerning reconstructions in this setting we present novel approaches that are able to recover surface orientation from Internet webcam images. We explore two different strategies to overcome the challenges posed by this kind of input data. One technique exploits orientation consistency and matches appearance profiles on the target with a partial reconstruction of the scene. This avoids an explicit light calibration and works for any reflectance that is observed on the partial reference geometry. The other technique employs an outdoor lighting model and reflectance properties represented as parametric basis materials. It yields a richer scene representation consisting of shape and reflectance. This is very useful for the simulation of new impressions or editing operations, e.g. relighting. The proposed approach is the first that achieves such a reconstruction on webcam data. Both presentations are accompanied by evaluations on synthetic and real-world data showing qualitative and quantitative results.
We also present a reconstruction approach for more controlled data in terms of the target scene. It relies on a reference object to relax a constraint common to many photometric stereo approaches: the fixed camera assumption. The proposed technique allows the camera and light source to vary freely in each image. It again avoids a light calibration step and can be applied to non-Lambertian surfaces.
In summary, this thesis contributes to the calibration and to the reconstruction aspects of photometric techniques. We overcome challenges in both controlled and uncontrolled settings, with a focus on the latter. All proposed approaches are shown to operate also on non-Lambertian objects.
|Place of Publication:||Darmstadt|
|Classification DDC:||000 Allgemeines, Informatik, Informationswissenschaft > 004 Informatik|
|Divisions:||20 Department of Computer Science > Interactive Graphics Systems|
|Date Deposited:||14 Nov 2014 11:43|
|Last Modified:||14 Nov 2014 11:43|
|Referees:||Goesele, Prof. Michael and Klein, Prof. Reinhard|
|Refereed:||16 June 2014|