OutofCore Compression for Gigantic Polygon Meshes

1
Out-of-Core Compression for Gigantic Polygon
Meshes

• Martin Isenburg, Stefan Gumhold
• Siggraph 2003 Accepted

Present by Pin Ren Apr 18, 2003
2
The Problem I

• Huge Size of 3D Mesh Data From
• 3D Scanning Technology
• Digital Michelangelo Project at Stanford
• Large Scale CAD Application
• Power Plant Model at UNC
• Compression / Decompression
• Without loss of fidelity
• Efficiency

3
The Problem II

• Limited Memory on Common PC
• Typically 128MB--1GB
• St.Matthew statue gt6GB (raw,186M vertices)
• Out-of-Core Vs. In-Core
• Things to be considered
• Small Memory Foot-Print (cant in the size of the
  Mesh!)
• Accuracy (Connectivity and Geometry)
• How to solve the side effects of techniques
  introduced (partitioning introduce
  discontinuities)

4
Previous Work

• Traditional Representation of Meshes
• Array of vertex positions
• Array of indices of vertex array to specify
  polygons
• Mesh Compression
• Deering95, Taubin and Rossignac98, Touma and
  Gotsman98, Brodsky and Watson00, Lee et.
  al02 Many nice works, still cannot fill the
  gap size of model and limited memory size of
  common PC.
• Ho et.al 02 address the problem by partitioning
  the manageable pieces. Then apply existing
  approaches. Artificial Discontinuities!

5
Previous Work cont.

• Out-of-Core algorithms
• Pointer De-referencing vertex soup.
• I/O efficient external algo eg. external merge
  sort
• Two main paradigms
• Batched data is streamed in one or more passes,
  small amount of data at memory at one time
• On-line data is processed through a series of
  queries. Common data structure B/B tree

6
Where to go?

• Limited memory vs. Huge Model Size
• Eliminate popular schemes
• Multi-pass approaches that has to store the
  entire mesh in the middle.
• Two-pass schemes that decompresses connectivity
  and Geometry
• The (only) promising way
• One-Pass single and memory-limited pass

7
Where to go? Cont.

• Common problem lower quality results caused by
  different handling schemes. Like partitioning,
  vertex clustering.
• We dont want to sacrifice accuracy when using
  out-or-core algorithms.
• How to achieve?
• Out-of-core meshes should provide the compressor
  transparent access to the info (c g)
• Octree-based external memory? Cignoni et.al 03
  (lack of explicit connectivity info)

8
Where to go? Cont.

• Partition into pieces vs. treat as whole
• Partition introduces discontinuity
• Compression Rate
• Decompression speed
• Compared with Ho et.al01
• 25 better compression rate
• 100 times fast (decompression speed)

9
Whatre their contributions? I

• An external memory data structure that provides
  transparent access to connectivity and geometry
  of gigantic meshes
• A method to construct this out-of-core mesh from
  an indexed mesh representation

10
Whatre their contributions? II

• A compression scheme that uses the out-of-core
  mesh to compress gigantic models in one piece
• A streamable, highly compressed mesh format that
  allows decompression at 2 million triangles per
  second
• The concept of a processing sequence for
  high-speed, limited memory foot-print mesh
  computations with access to connectivity

11
Out-of-Core MeshData Structure 1

• Major structure Half-Edge
• Information to retain
• Ability to enumeration of all half-edges and mark
  as visited
• Access to next, inverse half-edge and to origin
  vertex
• Access to the position of a vertex and whether it
  is non-manifold
• Knowledge of border edges

12
Out-of-Core MeshData Structure 2
13
Out-of-Core MeshData Structure 3

• Two mode
• Implicit for pure triangular meshes
• Explicit for polygonal meshes (has next)
• Capacity analysis
• 12 byte for each vertex or triangle, 8(I)/12(E)
  bytes for each half-edges
• St. Matthew 186M Vertices/372M Tri,
• At least 2.4GB4.8GB9.6GB/14.4GB

14
Out-of-Core MeshClustering

• Partition the mesh into a set of clusters
• Cache over clusters

15
Out-of-Core MeshCaching

• LRU replacing strategy
• Vertex, half-edge data, binary-flag of a cluster
  are kept separately in files
• Read/Written pattern
• Vertex data only read
• Half-edge data R/W when mesh built, only R by
  compressor
• Binary-flag of cluster R/W by compressor

16
Building Out-of-core MeshOverview

• Six stages
• 3 vertex passes
• 1 face pass
• 1 matching stage of incident half-edges
• Stage of linking and shortening of borders,
  search for non-manifold vertices
• All restricted to the in-core memory

17
Building Out-of-core MeshVertex Passes

• First
• determine of vertices,
• Find bounding box
• Second
• Computer balanced, spatial clustering of v
• K-nearest neighbors k6
• Similar to Ho et.al 01
• Third
• Sort the v into clusters
• Determine their index-pair
• Store mapping file (map index of v to
  index-pairs)

18
Building Out-of-core MeshFace Passes

• Create half-edges
• One half-edge for each of faces edges
• Store it into cluster it belongs to
• Primary file store it into cluster
• Secondary file store copies of H-E (crossing
  H-E) from other clusters temporary file for later
  matching stages.
• Decided order of half-edges in cluster
• Explicit mode sort by its origin vertex
• Implicit mode half-edges of one triangle has to
  be in successive order.

19
Building Out-of-core MeshMatching of Incident
Half-edges

• Get info from previous stored files
• Mixed sorting stratege
• A singled bucket-sort over all edges
• Quick-sort over edges of each buckets.
• O(nlogDmax)
• After sort, what info we get
• Looking at their number and orientation,we can
  distinguish 4 different types of edges
• Pair the half-edges to guarantee manifold
  connectivity

20
Building Out-of-core MeshBorder Loops and
Non-Manifold Vertices

• Link all border loops
• Simply cycle for each border half-edge
• Shorten border loops
• Detects and Mark non-Manifold vertices using two
  binary flags
• First specifies whether it was visited before
• Second specifies whether it is non-manifold

21
Compression--overview

• Single Pass to handle both connectivity and
  geometry info
• Do decompressor a favor
• At any time, it any need to have access to the
  boundaries of this region.
• Maintain some extra information at compression
  time

22
Decompression--Process
23
Connectivity Coding-degree coder

• Active Boundary
• Maintained as loop of boundary edges that are
  doubly-linked
• Coder grows a region by always including the face
  adjacent to the gate of the A-B
• Holes
• Special handling, otherwise leads to bad access
  pattern, store extra info
• Non-Manifold vertices
• Whenever a vertex is processed, store whether it
  is manifold
• Only first-time non-manifold vertex is
  compressed.

24
Geometry Coding-predictive coder

• Quantization precision
• For Scanned data, keep the quantization error
  just below scanning error
• Quantized vertices compressed with Parallelogram
  Rule Touma98
• First three of each mesh cannot
• Prediction value is simple
• Other properties similarly to vertex pos

25
Conclusions

• First technique that is able to compress St.
  Matthew in one piece on desktop PC
• New out-of-core mesh data structure
• Efficient building process
• Compress scheme
• The concept of processing sequence

26

• Thank you very much.
• Any Questions?