Gael Sourimant - Homepage

• Abstract
  

  Virtual worlds feature increasing geometric and shading complexities, resulting in a constant need for effective solutions to avoid rendering objects invisible for the viewer. This observation is particularly true in the context of real-time rendering of highly occluded environments such as urban areas, landscapes or indoor scenes. This problem has been intensively researched in the past decades, resulting in numerous optimizations building upon the well-known Z-buffer technique. Among them, extensions of graphics hardware such as early Z-culling efficiently avoid shading most of invisible fragments. However, this technique is not applicable when the fragment shader discards fragments or modifies their depth value, or if alpha testing is enabled.

  We introduce Triple Depth culling for fast and controllable perpixel visibility at the fragment shading stage using multiple depth buffers. Based on alternate rendering of object batches, our method effectively avoids the shading of hidden fragments in a single pass, hence reducing the overall rendering costs. Our approach provides an effective control on how culling is performed prior to shading, regardless of potential discard or alpha testing operations. Triple Depth culling is also complementary with the existing culling stages of graphics hardware, making our method easily integrable as an additional stage of the graphics pipeline.

• Location Technicolor Research & Innovation
• Downloads [ Paper ] [ Video (Available soon) ] ACM Siggraph'11
• Links [ View Paper ] [ Video Stream ] [ Book Publisher ] [ Buy Me! ]

• Abstract
  

  We introduce procedural geometry mapping and ray-dependent grammar development for fast and scalable render-time generation of procedural geometric details on graphics hardware. By leveraging the properties of the widely used split grammars, we replace geometry generation by lazy per-pixel grammar development. This approach drastically reduces the memory costs while implicitly concentrating the computations on objects spanning large areas in image space. Starting with a building footprint, the bounding volume of each facade is projected towards the viewer. For each pixel we lazily develop the grammar describing the facade and intersect the potentially visible split rules and terminal shapes. Further geometric details are added using normal and relief mapping in terminal space. Our approach also supports the computation of per-pixel self shadowing on facades for high visual quality. We demonstrate interactive performance even when generating and tuning large cityscapes comprising thousands of facades. The method is generalized to arbitrary mesh-based shapes to provide full artistic control over the generation of the procedural elements, making it also usable outside the context of urban modeling.

• Location Technicolor Research & Innovation
• Downloads [ Paper ] [ Video (Results) ] GI'11
• Links [ Video Stream ]

• Abstract
  

  Ambient occlusion improves the realism of synthetic images by accounting for the overall light attenuation due to local occlusions. This technique recently became a popular tool to produce realistic images, either for offline rendering or in interactive applications such as games. For offline rendering, ambient occlusion (AO) is traditionally computed by casting many rays from each visible point and analyzing the amount of rays hitting the geometry in a close vicinity around the considered point. To achieve interactive frame rates, AO is approximated in screen-space, using the depth buffer to represent the scene geometry, and built upon analytical or numerical tricks to replace the ray-traced occluder search.

  We present a novel method for computing screen-space ambient occlusion (SSAO) at interactive rates, while preserving the original formulation of the standard ray-traced ambient occlusion. Based on ray-marching on graphics hardware, our method achieves interactive frame rates with physically consistent results, even with few samples. Our SSAO computation is settled within a standard deferred shading framework, making it easily integrable into post-production / previsualization workflows, or in computer games.

• Location Technicolor Research & Innovation
• Downloads [ Paper ] [ Video (Results) ] [ Video (Fast-forward) ] I3D '11
• Links [ View Paper ] [ Video Stream (Results) ]

• Abstract

  We explores in this work several depth maps estimation methods, in different video contexts. For standard (monocular) videos of fixed scene and moving camera, we present a technique to extract both the 3D structure of the scene and the camera poses over time. These information are exploited to generate dense depth maps for each image of the video, through optical flow estimation algorithms. We also present a depth maps extraction method for multi-view sequences aiming at generating MVD content for 3DTV. These works are compared to existing approaches used at writing time in the 3DV group of MPEG for normalization purposes. Finally, we demonstrate how such depth maps can be exploited to performrelief auto-stereoscopic rendering, in a dynamic and interactive way, without sacrifying the real-time computation constraint.

• Location INRIA Rennes Bretagne Atlantique
• Downloads [ Technical Report ] [ View Paper ]
• Links [ View Report ]
• Thanks This work could not have been achieved without these guys' work

Reconstruction of urban areas based on GPS, GIS and Video registration

• Abstract

  This thesis presents a new scheme for 3D buildings reconstruction, using GPS, GIS and Video datasets. The goal is to refine simple and geo-referenced 3D models of buildings, extracted from a GIS database (Geographic Information System). This refinement is performed thanks to a registration between these models and the video. The GPS provides rough information about the camera location.

  First, the registration between the video and the 3D models using robust virtual visual servoing is presented. The aim is to find, for each image of the video, the geo-referenced pose of the camera (position and orientation), such as the rendered 3D models projects exactly on the building images in the video.

  Next, textures of visible buildings are extracted from the video images. A new algorithm for façade texture fusion based on statistical analysis of the texels color is presented. It allows to remove from the final textures all occluding objects in front of the viewed building façades.

  Finally, a preliminary study on façades geometric details extraction is presented. Knowing the pose of the camera for each image of the video, a disparity computation using either graph-cuts or optical flow is performed in texture space. The micro-structures of the viewed façades can then be recovered using these disparity maps.

• Advisors Luce Morin and Kadi Bouatouch
• Location INRIA Rennes Bretagne Atlantique
• Downloads

  [ PDF File ] (in French, 25.4MB)

  [ PDF File ] (in French, Low resolution version, 9.4MB)

• Links [ View Manuscript ]