Skip to Main content Skip to Navigation

Photometric registration of indoor real scenes using an RGB-D camera with application to mixed reality

Salma Jiddi 1, 2
2 RAINBOW - Sensor-based and interactive robotics
Abstract : The overarching goal of Mixed Reality (MR) is to provide the users with the illusion that virtual and real objects coexist indistinguishably in the same space. An effective illusion requires an accurate registration between both worlds. This registration must be geometrically and photometrically coherent. In this thesis, we propose novel photometric registration methods to estimate the illumination and reflectance of real scenes. Specifically, we propose new approaches which address three main challenges: (1) use of a single RGB-D camera. (2) estimation of both diffuse and specular reflectance properties. (3) estimation of the 3D position and color of multiple dynamic light sources. Within our first contribution, we consider indoor real scenes where both geometry and illumination are static. As the sensor browses the scene, specular reflections can be observed throughout a sequence of RGB-D images. These visual cues are very informative about the illumination and reflectance of scene surfaces. Hence, we model these cues to recover both diffuse and specular reflectance properties as well as the 3D position of multiple light sources. Our algorithm allows convincing MR results such as realistic virtual shadows and correct real specularity removal. Shadows are omnipresent and result from the occlusion of light by existing geometry. They therefore represent interesting cues to reconstruct the photometric properties of the scene. Presence of texture in this context is a critical scenario. In fact, separating texture from illumination effects is often handled via approaches which require user interaction or do not satisfy mixed reality processing time requirements. We address these limitations and propose a method which estimates the 3D position and intensity of light sources. The proposed approach handles dynamic light sources and runs at an interactive frame rate. The existence of a light source is more likely if it is supported by more than one cue. We therefore address the problem of estimating illumination and reflectance properties by jointly analysing specular reflections and cast shadows. The proposed approach takes advantage of information brought by both cues to handle a large variety of scenes. Our approach is capable of handling any textured surface and considers both static and dynamic light sources. Its effectiveness is demonstrated through a range of applications including real-time mixed reality and retexturing. Since the detection of cast shadows and specular reflections are at the heart of this thesis, we further propose a deep-learning framework to jointly detect both cues in indoor real scenes.
Complete list of metadatas

Cited literature [205 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Thursday, June 27, 2019 - 2:28:12 PM
Last modification on : Wednesday, September 9, 2020 - 4:06:17 AM


Version validated by the jury (STAR)


  • HAL Id : tel-02167109, version 1


Salma Jiddi. Photometric registration of indoor real scenes using an RGB-D camera with application to mixed reality. Computer Vision and Pattern Recognition [cs.CV]. Université Rennes 1, 2019. English. ⟨NNT : 2019REN1S015⟩. ⟨tel-02167109⟩



Record views


Files downloads