Spatial Tone Mapping in High Dynamic Range Imaging Zhaoshi Zheng
High Dynamic Range Imaging Conventional digital representation of images in computers can not produce the color range in natural world So the techniques for capturing, storing, and reproducing realistic scene developed – HDRI Source: apw
Tone Mapping Most display device can not accommodate the color range of HDR Map HDR to LDR: tone reproduction, or tone mapping Spatial: Operators work directly on pixels Others: Frequency and Gradient domain
Spatial Tone Mapping Operators Demonstrate inherited data level parallelism Good candidates for multi-threaded implementation on modern GPUs Global: same mapping curve for all pixels Local: mapping curve for each pixel depends on its neighbor pixels
Tumblin-Rushmeiser Brightness- Preserving Operator
Tumblin-Rushmeiser Brightness- Preserving Operator (Implementation)
Results (Image Comparison) Left: Noop, direct pixel value in hdr file; Right: TR tone mapping operator, w/ scale factor equals 256 and saturation equals 0.8, produced by CPU
GPU Naïve Implementation
Results (Image Comparison) Left: Original CPU code released with the book, using double precision; Middle: Slightly modified (on data structure and file IO) CPU code, using single precision; Right: GPU code, using single precision.
Results(Execution Time)
Result(Speedup)
Future work Optimization Vectorize computation of luminance Compute Log Average on GPU (tradeoffs) More operators 3~5 other operators, aiming at speed up
Reference Eric Reinhard, Greg Ward, Sumanta Patanaik and Paul Debevec, “High Dynamic Range Imaging”