Tone mapping with Local Laplacian Filtering

Please click on the thumbnails to see the full-resolution images and use the back button to return to this page.

1. Input image

The original HDR image is available here. We resized it using Adobe Photoshop CS5. Below we show it tone mapped with a simple gamma curve for display purposes.

2. Our results

The α parameter controls how details are modified: α=0.25 amplifies detail the most, while α=1 keeps it unchanged.
The β parameter controls the balance between local and global contrast. Small values favor local contrast.
The σ_r parameter is fixed and equal to log(2.5).

	α = 0.25	α = 0.5	α = 0.75	α = 1
β = 0
β = 0.3
β = 0.6

3. Comparisons to previous work

Below we show results produced by two other tone mapping methods using code provided on the authors' webpages. For clearer comparison, we adjusted parameters for post-processing to match the same output dynamic range (100:1 contrast between the 99.5th and 0.5th percentiles) and color saturation (gamma corrected with an exponent of 1/2.2) as implemented by our own method. The treatment of the mid-tones may differ between the methods, which explains why the average brightness may differ.

3.1. Weighted least squares [Farbman et al. 2008]

The following results were produced using the method described in Edge-Preserving Decompositions for Multi-Scale Tone and Detail Manipulation, Farbman et al., ACM SIGGRAPH 2008 (webpage). We contacted the authors and used the fixed parameters that they suggested. We tested both their two-scale and their multi-scale method:

	neutral rendition λ=5, w = [1 0.45]	exaggerated rendition λ=0.1, w = [6 0.35]
two-scale method

	neutral rendition λ=[0.05 2 16] w = [1.2 1.2 1 0.45]	detailed rendition λ=[0.05 2 16] w = [2 1 0.4 0.25]	exaggerated rendition λ=[0.05 2 16] w = [6 0.6 0.6 0.12]
multi-scale method

This approach performs best for neutral renditions and detail enhancement up to an intermediate level. Exaggerated renditions similar to ours (α=0.25 and β=0) are more challenging. In many cases, especially in its multi-scale version, this method produces satisfying images with a high level of detail. For some images, however, rim halos and edge defects still appear. The code provided by Farbman and colleagues with scripts to generate these comparisons can be downloaded here.

3.2. Smooth gain maps [Li et al. 2005]

The following results were produced using the method described in Compressing and Companding High Dynamic Range Images with Subband Architectures, Li et al., ACM SIGGRAPH 2005 (webpage). The code available on-line includes a post-processing step not described in the original paper to "open up" the shadows. We did not apply this step here, to focus on the basic image processing. We compare to the parameter settings proposed by the authors, using an oversampled Haar wavelet decomposition:

	neutral rendition β = 0.6, α_A = 0.2 endrate = 0.6	detailed rendition β = 0.6, α_A = 0.2 endrate = 0.4

This approach is best suited for neutral renditions, where it performs well, but the detailed rendition often exhibits halo artifacts. We tried to adjust the parameters to increase the details even further, but could not achieve results similar to our exaggerated rendition without introducing unacceptable halos.