| Filter Properties | Our approach | EAW Fattal [2009] |
WLS Farbman et al. [2008] |
PLBF Adams et al. [2009] |
CTBF Yang et al. [2009] |
| Real-time performance | Yes | Yes | No | No | Yes |
| True color filtering | Yes | Yes | Yes | Yes | No |
| Performance independent of the filtering parameters | Yes | Yes | Yes | No (uses downsampling) |
No (uses quantization) |
| Continuous control over the kernel's size | Yes | No (only powers-of-two) |
Yes | Yes | Yes |
| Control over the kernel's shape | Yes | Yes | No | Yes | Yes |
Our edge-aware filters are the first to achieve real-time, parameter-independent performance when working on color images at arbitrary scales. Having all of these properties concurrently can lead to significant improvements on applications that rely on edge-aware operators.
Real-time performance enables instantaneous feedback for users in interactive environments, providing friendlier and easier-to-use tools for the public at large. This performance also makes possible on-the-fly processing of high-resolution videos, which are becoming more prominent in TV broadcasting and on the Web. Furthermore, since our filters have a running time independent of their parameters, our approach does not impose any parameter-space constraint due to excessive computational overhead.
Other properties such as true color filtering are important to guarantee high-quality filtering results, since processing color information independently should introduce artifacts around the edges [Tomasi and Manduchi 1998]. Furthermore, continuous control over the kernel's size (i.e., σs and σr) allows for more freedom when performing edge-preserving smoothing, since it enables operating at arbitrary scales. The images below exemplify the importance of this property when filtering a photograph, shown in (a).
Images (b) to (e) show that our filter manages to continuously smooth image regions while preserving strong edges. Notice how details inside the lamp and on the wall are progressively smoothed, while the overall image structure is preserved. In comparison, images (f) to (i) show that the edge-avoiding wavelets filter (EAW) [Fattal 2009], which restricts filtering to only a few scales, does not manage to remove some high-frequency details from the image, such as details inside the lamp and on the wall. For these examples, edge-preserving smoothing with EAW was achieved by scaling the detail coefficients, for each decomposition level, by 0.6^(max_depth - level).