With the increasing complexity of VR model, limitations appear both in memory loading and remote access of such data. The basic solution is to reduce the amount of model information (3D/Texture) by either compression or detail elision means. This approach is quite similar to video coding in the sense that people aim to make a lossy reduction of information, taking account the visual artifacts which are generated by the loss of information. The main difference with video coding is that VR data intrinsically provides critical information about the scene content while for video, these information have to be extracted during a sophisticated analysis phase. These information are typically the followings:

           VR Info: (Known)             Video Info: (Analysis)                  
                Shape                   Contour extraction                      
                Object                  Segmentation & tracking                 
                Speed                   Motion estimation                       
               Rotation                 Motion estimation (Affine model)        
                 Zoom                   Motion estimation (Affine model)        
             Illumination                                                       
               Texture                  Mathematical transform  (Gabor for      
                                        instance)                               

The optimal use of such data will lead not only to allow visualization of more complex scenes. It will open the door to the transmission of virtual environment with astounding compression ratio, thus enabling to get 3D remote capabilities with access of limited bandwidth.

[1] T.A. Funkhouser, C .H. Sequin, Adaptive Display Algorithm for Interactive Frame Rates During Visualization of Complex Virtual Environments, University of California at Berkeley, Proceedings of SIGGRAPH'93 Computer Graphics 17, pp 247-254.

* Keywords:

Feed-back: Action to take into account the computing time at previous frame for decisions made on current frame.

Cost/Benefit function: Ratio used to select LOD of different objects in a scene depending on a targeted frame rate.

* Type of processed Data:

Unstructured mesh.

* Targeted Problem:

The paper presents a technique which enable to keep a constant frame rate while displaying 3D complex scenes.

* Solution:

The solution proposed is based on the appropriate selection of LOD of 3D objects of the scene. Each object has a hierarchical representation with different LOD.

* Originality:

The selection criteria is based on the optimization of Cost/Benefit function. The Cost is defined empirically for a given machine as the time required to render a object with a given LOD. Benefit is defined as the visual enhancement due to an object rendering with a given LOD. The approach proposed allows to give a good approximation of the NP-complete problem "which LOD for which object".

* Results:

Results are shown in different cases of LOD management for an animation created by a given path into the auditorium model and compared to a so-called feed-back method. The feed-back method selects the LOD of each object depending both and the area covered on the screen and the time required to render previous image. Results clearly show improvement of frame rate regulation.

* Open Issues:

"The automatic Generation of multi-resolution models and experiments to develop measures of image quality and image differences"

[2] P. Lindstrom, D. Koller, Level of detail Management for Real-time Rendering of Phototextured Terrain, Georgia Institue of Technology

* Keywords:

Cell: A power of 2 set of vertices defining a square area.

* Type of processed Data:

Structured grid mesh and Texture.

* Targeted Problem:

The paper presents a technique to reduce mesh and texture data.

* Solution:

The technique is based on a quadtree multiresolution of the grid. This allows to scale by a power of 2 the number of polygons per cell. This scalability is then used to adjust LOD depending on visual criteria. The visual criteria is based on area covered on the screen by each cell.

* Originality:

By taking into account orientation of cells between them, the technique allows to use lower resolution when the surface is close to a plane. Computation of the orientation of cell in respect to the view point enables to adjust both quadtree LOD and texture resolution.

* Results:

Results show both number of polygons rendered and bytes of texture rendered. The scene is a 120 frames sequence of a path on a surface. Thanks to the technique, only 1% of polygons which are in view frustum are rendered and 10% of texture information is used.

* Open Issues:

Temporal and spatial artifacts generated by resolution changes of the grid: Connection of different resolution cells, texture block-effects, ... Generalization to unstructured meshes.

[3] G. Schaufler, Exploiting Frame-to-Frame Coherence in a Virtual Reality System, GUP, Johannes Kepler Universitat Linz, Proceedings of VRAIS'96, pp. 95-102

* Keywords:

Impostor: A texture mapped on a tranparent plane object wich looks like an object.

* Type of processed Data:

Unstructured mesh.

* Targeted Problem:

The paper presents a technique to reach a given frame rate.

* Solution:

The solution is based on a similar approach as [1]. The algorithm simplifies the model in order to reach the targeted frame rate. The choice of LOD for each object is made using the optimization algorithm proposed by [1].

* Originality:

The simplification is based on replacement of some objects by textured planes which are called impostors and can take benefit of the frame-to-frame coherence of rendered scenes. It the object is sufficiently small or the motion sufficiently slow, the impostor looks like the object. Several LOD impostors are used for each corresponding object LOD. A computation of the view angle to the impostor bounding box is made. It enables to determine when to update the texture of the impostor as viewer is moving.

Since, from one frame to the other, the same impostor can be used, this leads to a significant decrease of the computing cost to render the whole scene. It is then possible to choose higher LOD for impostors which gives an increasing quality of scene through time.

* Results:

The results are shown with a 10600 polygons and 34 objects scene. For a viewer translation, results show an increasing quality of images during the sequence as the frame rate is kept constant. Funkhouser algorithm gives constant and lower quality.

* Open Issues:

The stereo view management with impostors different for each eye could be difficult. Problem arise when replacing intersecting objects by impostors. Animated objects and texture remains to study.

[4] P. Lindstrom, D. Koller, W. Ribarsky, L. Hodges, Real-Time, Continuous Level of Detail Rendering of Height Fields, Georgia Tech.

* Keywords:

* Type of processed Data:

Structured grid mesh.

* Targeted Problem:

Reduction and compression of grid mesh information.

* Solution:

The solution is based on a quadtree LOD simplification as shown in [2] followed by vertex based simplification. The visual Benefit function is defined as the projected delta due to the mesh simplification.

* Originality:

The vertex based LOD simplification which colapses 2 triangles into one, thus providing a mean to decrease "continously" the LOD. The paper also describes a method to compress delta information by using a non-linear quantizer.

* Results:

The difference between non-simplified rendered image and simplified version is used to quantify the error. Tests are made with a 3000 frames sequence. They show a number of rendered polygon reduced by 2 orders of magnitude with an error rate of 5% which is presented to be better when compared to classical quadtree multiresolution approach.

* Open Issues:

Gaps between polygons of different LOD as well as temporal artifacts.

[5] Reinhard Klein, Tobias Huttner, Simple camera-dependent approximation of terrain surfaces for fast visualization and animation, Tubingen Uniersity, Germany, IEEE visualization'96, "Late breaking hot topics papers", pp. 20-24

* Keywords:

TIN: Triangulated Irregular Network

* Type of processed Data:

Grid mesh and TIN.

* Targeted Problem:

Reduction of grid mesh information depending on viewer position.

* Solution:

Multiresolution preprocessing of the elevation field and then choice of proper resolution for each cell, given the view point.

* Originality:

Instead of using a quadtree approach as [2], a TIN is used.

* Results:

With an initial terrain of 1.440.000 vertices, a simplified model was built with 213.000 vertices. The update as viewer is moving takes less than 200 ms on a 150Mhz impact SGI.

* Open Issues:

Temporal aliasing due re-triangularisation.

[6] Julie C.Xia, Amitabh Varshney, Dynamic View-Dependent Simplification for Polygonal Models, State university of New York at Stony Brook, IEEE Proceedings of Visualization'96, pp. 327-334, October 1996, San Francisco, CA.

* Keywords:

Merge Tree: Parents-children relationships between vertices. Used to represent variable resolution of mesh.

Image-space-guided simplification: Mesh ellision based on what the viewer can see after rendering.

* Type of processed Data:

Irregular mesh.

* Targeted Problem:

View dependent simplification of a mesh.

* Solution:

A continous LOD representation of the mesh is done off-line and used at run time to display the most important vertices.

* Originality:

The continous LOD mechanism coupled with local image-space-guided simplifications. It enables, for instance, to adapt to local illumination of an object region.

* Results:

With 4 different objects, a decrease by more than one order of magnitude of the number of triangles is shown. Real-time computation is reached on a 250 Mhz Impact SGI.

* Open Issues:

Use of normals and colors in the merge tree.

Visibility Computations in Densely Occluded Polyhedral Environments

Feel free to send an email if you have any comment or information !