Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Adaptive Texture Maps Martin Kraus and Thomas Ertl VIS Group, Universität Stuttgart

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Introduction WO R EMARKS :  Our goal is to implement “adaptive” texture maps with off-the-shelf graphics hardware.  This research is about an application of programmable graphics hardware, not a suggestion for new graphics hardware. T

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Introduction WO P ROBLEMS OF H ARDWARE -A SSISTED T EXTURE M APPING :  Texture data must be specified on uniform grids. o We restrict ourselves to the mapping from texture coordinates to texture data.  Limited texture memory: o Usually enough for 2d texture data, o Hardly enough for 3d texture data, o Usually not enough for 4d texture data. T

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Introduction WO M AIN P ARTS OF T HIS T ALK :  Adaptive Texture Maps in 2 Dimensions o Adaptivity and Requirements o Data Representation, Sampling, and Generation  Applications in 3 and 4 Dimensions o Volume Rendering o Light Field Rendering T

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Adaptive Texture Maps in 2 Dimensions WO K INDS OF A DAPTIVITY AND C OMPRESSION :  Adaptive domain of texture data (lossless compression).  Locally adaptive resolution of texture data (lossy compression). T

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Adaptive Texture Maps in 2 Dimensions WO R EQUIREMENTS :  Fast random access to texture data. Programmable texturing allows us to decode texture data on-the-fly.  Two-level data representation. Because the ATI Radeon 8500 is limited to one level of dependent texture reads. T texture coordinates per-pixel operations decoded texel encoded texture map

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Adaptive Texture Maps in 2 Dimensions WO L EVELS OF THE D ATA R EPRESENTATION:  Index data (upper level): o Each cell/texel of a coarse grid corresponds to one data block. o Each cell/texel specifies coordinates and scaling factors of the corresponding data block.  Packed data (lower level): o All data blocks packed into one uniform grid/texture. T

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Adaptive Texture Maps in 2 Dimensions WO S TEPS OF S AMPLING A DAPTIVE T EXTURES:  Read index data and calculate coordinates for the second step.  Read and interpolate actual texture data from packed data. T (s,t) (s’,t’)

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Adaptive Texture Maps in 2 Dimensions WO P ASSES OF A F RAGMENT S HADER P ROGRAM: T texture coordinates, constants, primary and secondary color sampling and routing arithmetics (dep.) sampling and routing arithmetics temporary registers fragment color 1st pass 2nd pass

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Adaptive Texture Maps in 2 Dimensions WO S TEPS OF G ENERATING A DAPTIVE T EXTURES:  Separate downsampling of each data block: o Downsampling is repeated for each data block until some error threshold is reached. o Scale factors are stored in index data. o Special treatment of block boundaries (cont. interpol.!)  Packing of downsampled data blocks: o Simple, non-optimal packing algorithm is sufficient. o Coordinates of packed blocks are stored in index data. T

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Second Part of the Talk WO A PPLICATIONS (V ARIANTS) :  Volume Rendering (3D Texture Maps) (Data from the Stanford volume data archive.)  Light Field Rendering (4D Texture Maps) (Data from the Standford light fields archive.) T

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Application: Volume Rendering WO K INDS OF T EXTURE- B ASED V OLUME R ENDERING : object-aligned axis-aligned (2d textures) (3d textures) T

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Application: Volume Rendering WO I NTERNAL 3D T EXTURE M APS F OR A DAPTIVE 3D T EXTURE M APS : index data: 32 3 cells packed data: voxels (32 3 cells×16 3 voxels/cell = voxels) T

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Application: Volume Rendering WO K INDS OF A RTIFACTS :  Sampling error in data set.  Discontinuous boundaries between data blocks because of fixed-point arithmetics in fragment shader programs. T

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Application: Volume Rendering WO S TEPS TO V ECTOR Q UANTIZATION :  Use few (256), tiny data blocks (2 3 voxels).  Use nearest-neighbor interpolation in packed data. (See “Texture Compression” in EUROGRAPHICS 2002, Tutorial T4.) T

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Application: Light Field Rendering WO P AIRS OF C OORDINATES : (Levoy, Hanrahan: Light Field Rendering, SIGGRAPH ‘96.) T

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Application: Light Field Rendering WO- L EVEL D ATA R EPRESENATION :  Each data block covers several values of s and t, one value of u and all values of v. index data one data block T

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Application: Light Field Rendering WO T RILINEAR I NTERPOLATIONS F OR O NE Q UADRILINEAR I NTERPOLATION :  Trilinear interpolation of L(floor(u), v, s, t).  Trilinear interpolation of L(ceiling(u), v, s, t).  Linear interpolation of results with weights ceiling(u) - u and u - floor(u) gives L(u, v, s, t). T

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Application: Light Field Rendering WO E XAMPLES : T

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Conclusions WO Q UESTIONS :  How useful are adaptive texture maps? o Depends very much on the texture data. o Very useful for data with strongly varying resolution. o Also useful for data with large empty regions. o Vector quantization is useful for volume rendering.  Should they be implemented in hardware? o No, because programmable graphics hardware will soon be good enough. T

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 Future Work WO A REAS :  Exploiting new graphics hardware: o Floating-point precision, o Combination with other per-pixel computations, o Deeper hierarchies with more dependent texture reads.  Exploring more fields of application: o Normal maps, environment maps, shadow maps, … o BRDFs, multi-dimensional transfer functions, … T

Martin Kraus and Thomas Ertl, Adaptive Texture Maps Graphics Hardware 2002 Saarbrücken September 1-2, 2002 WO M ORE T HINGS TO S AY :  Thank you!  Any questions? T