Rapid Visualization of Large Point-Based Surfaces Tamy Boubekeur Florent Duguet Christophe Schlick Presented by Xavier Granier.

Rapid Visualization of Large Point-Based Surfaces

Tamy Boubekeur

Florent Duguet

Christophe Schlick

Presented by Xavier Granier

Large Data Acquisition

• Sub-millimeter acquisition and saving of statues and archaeological artifact

• Billions of sample for accurate representation• Need of specific methods for visualizing such

objects

David laser scansDigital Michelangelo Project

Complex archeological artifactDuguet et al. VAST 2004

Large Objects - Prior Art

• Visualization of gigantic meshes– Out-of-core rendering [Lindstrom 2003], [Tetra-Puzzles -

Cignoni 2004]

– Multiscale approach [Far Voxels - Gobetti 2005]

• Using points– As rendering primitives [QSplat - Rusinkewicz 2000]

[Dachsbacher 2003]

– As surface representation [VAST - Duguet 2004]

Acquisition & Visualization

2.5D Scans

RegistrationSurface reconstruction

Merging

Appearance preservingSimplification

Real-time rendering

Large meshLarge PBS

RealWorld

3D Image

Intermediate large mesh generation and storing

In-core rendering

PBS : Point-based surface

Problem

• Surface reconstruction

• Simplification of resulting large meshes

• Preprocessing for out-of-core rendering

…are time-consuming tasks Hours to week of computation on single workstations for scanned statues.

Bottleneck for visualization

2.5D Scans

RegistrationSurface reconstruction

Merging

Appearance preservingSimplification

Real-time rendering

Large meshLarge PBS

RealWorld

3D Image

Intermediate large mesh generation and storing

Time consuming tasks

Appearance Preserving Simplification

• Reducing the complexity of 3D objects• Maintaining as much as possible their appearance• Usual solution

– Large mesh >>> coarse mesh + high resolution textures (normal, color)

– Requires mesh generation and simplification

Normal mappingCoarse mesh

Time consuming tasks

Solution: removing the bottleneck

2.5D Scans

RegistrationSurface reconstruction

Merging

Appearance preservingSimplification

Real-time rendering

Large meshLarge PBS

RealWorld

3D Image

Intermediate large mesh generation and storing

Removing the intermediate mesh

2.5D Scans

Registration

Appearance preservingGeneration

Real-time rendering

Large PBS

RealWorld

3D Image

Using only the registered point-cloud

Direct processing for visualization

Our fast conversion pipeline

• No surface reconstruction at full resolution

• No global surface parameterization

• Direct PBS to appearance preserving representation conversion

Our approach

Surfel : Surface Element, sampled point with associated sampled normal, color, etc…

1 - Out-Of-Core Resampling

• The first reading pass– Filtering the registered PBS

through a regular grid– Keeping at most one

represent per cell

• Similar to out-of-core simplification for meshes [Lindstrom 2000]

• Typical output : few tens thousands points In-core point cloud

2 - Local coarse mesh generation

• Building an octree over the simplified point cloud

• Local generation of pieces of surfaces : Surfel Strips [Boubekeur 2005] - Lower dimensional triangulation

• Overlapping between neighboring pieces of mesh– Provide hole free visualization– Each piece processed

independantly

Collection of surfel strips

2 - Surfel Stripping [Boubekeur 2005]

– Partitioning criterion

Height-field predicate

– Local 2D Delaunay triangulation

– Fast cache-coherent stripping [Reuter 2005]

Local partition Projection 2D triangulation Fast stripping Surfel Strip

2 - Local coarse mesh generation

• Overlapping decimation– Reducing redundant/non useful

triangles

• Output : mesh clustered in an octree– Mesh = collection of Surfel

Strips– Each surfel strip independently

generated– Bounding quad in the average

plane

Direct visualization of surfel strips

3 - Out-Of-Core normal mapping

• Second reading pass• Filtering all the point through the octree• Projecting point’s normals on textures of intersected

leaves• Output : textured surfel strips

– Coarse mesh + sparse normal map

Holes in normal map = no normal projected

4 - Normal map diffusion

• Filling hole in normal maps : diffusion with push-pull• Per surfel strip diffusion

1. Quad-tree construction

2. Hierarchical hole filling

3. Smoothing

Coarse surfel strip’s topology Normal mapping

4 - Normal map diffusion

1 .Quad-Tree construction (PUSH) 2. Hierarchical hole filling (PULL)

3. Iterative smoothing (gradient constrained)

Output :

Coarse mesh + high resolution normal map

Surfel Strips Diffused per-surfel strip normal map

View-coherent packing of texures

Results

Real-time rendering of archaeological artifacts on a single workstation Dancer (30 M samples)

Drums (20 M samples)Omphalos (10 M samples)

Results

Real-time rendering of archaeological artifacts on a single workstation

St Matthew (186 M samples)

Performance

Timing, frame-rate and memory space for a single workstation Intel PIV 3.4 GHz, 1.5 Go ram

Details preserving

• Small details represented only in normal map, stored on GPU texture memory

Surfel strips only Normal surfel strips

Comparison to QSplat

• Better hardware support (coarse mesh + normal maps)

• Realtime rendering in high resolution at high framerate (2 to 3 orders of magnitude faster)

• Mipmapping = Automatic Hardware Filtering

Our approach QSplat

Avantages

• No surface reconstruction of full PBS

• No complex processing on full PBS

• No global parameterization for normal mapping, only local planar projection

• Very fast processing

• Final in-core model mostly stored as texture on GPU memory

• Automatic hardware filtering

Limitations

• Still a simplification approach

• Out-of-core resampling can miss small topological details– Using adaptive out-of-core resampling

methods [Schaeffer 2003]– After tests : no significant difference with our

data sets – Very large object can be resampled in a simple grid [Lindstrom 2000]

Conclusion

• An easy-to-implement pipeline for visualizing large scanned objects

• Suitable for very large and dense point clouds• Can preserve any sampled surface property :

– Normal – Color– Etc…

Scanned objects such as statues and other archaeological artifact can be stored as simple unorganized point clouds

On-going work

• Mutli-scale GPU-friendly structures– See Surfel Stripping [Boubekeur Graphite 2005]

• Larger scenes processing and visualization on single workstations – 10 Billions ?– 100 Billions ? (on-the-fly surface synthesis)

• Advanced comparison– Sequential Point Tree + Splatting on Today’s GPU

[Dachsbacher 2003] [Botsch 2003]… still less efficient than ours (no true hardware support for point-based rendering)

Thank you for your attention !

http://www.labri.fr/~boubek