Eikonal-based region growing for efficient clustering

Abstract

We describe an Eikonal-based algorithm for computing dense oversegmentation of an image, often called superpixels. This oversegmentation respects local image boundaries while limiting undersegmentation. The proposed algorithm relies on a region growing scheme, where the potential map used is not fixed and evolves during the diffusion. Refinement steps are also proposed to enhance at low cost the first oversegmentation. Quantitative comparisons on the Berkeley dataset show good performance on traditional metrics over current state-of-the art superpixel methods.

Robust superpixels using color and contour features along linear path

“…Moreover, contrary to most state-of-theart methods, SCALP is robust to noise, since it provides accurate and regular decompositions on the noisy part of the image. (Buyssens et al, 2014), ETPS (Yao et al, 2015) and LSC (Chen et al, 2017). SCALP obtains the most visually satisfying result with superpixels that adhere well to the image contours.…”

“…Most of the regular methods consider an initial regular grid, allowing to set the number of superpixels, and update superpixels boundaries while applying spatial constraints. Classical methods are based on region growing, such as Turbopixels (Levinshtein et al, 2009) using geometric flows, or eikonal-based methods, e.g., ERGC (Buyssens et al, 2014), while other approaches use graphbased energy models (Liu et al, 2011;Veksler et al, 2010). In Machairas et al (2015), a watershed algorithm is adapted to produce regular decompositions using a spatially regularized image gradient.…”

“…The gain over state-of-the-art methods is largely increased when computing superpixels on noisy images. Methods such as Buyssens et al (2014); Chen et al (2017); Liu et al (2011) obtain very degraded performances when applied to slightly noised images, while Van den Bergh et al (2012) is the only method that is robust to noise on all evaluated aspects. The state-of-the-art methods can indeed have very different behavior when applied to noisy images.…”

“…This aspect is expressed by the lower performances of CD over BR in the bottom part of Figure 15. The regularity is also degraded for all methods, except Buyssens et al (2014), that tends to generate more regular superpixels, failing at grouping homogeneous pixels. Finally, on the ASA metric, SCALP provides slightly higher results than SCALP+HC for these Figure 15: Comparison between the proposed SCALP framework and the state-of-the-art methods on contour detection (PR) and superpixel metrics (ASA, CD over BR and SRC) on the BSD test set.…”

Texture-Aware Superpixel Segmentation

“…Nevertheless, no other approach explicitly captures texture information, so TASP with default parameters still outperform state-of-the-art methods manually optimized for each dataset. mix-Stripes mix-Brodatz BSD Method ASA F ASA F ASA F SLIC [14] 0.7256 0.4048 0.7784 0.4607 0.9445 0.4706 ERGC [15] 0.6107 0.3717 0.7796 0.4677 0.9477 0.4571 ETPS [20] 0.7769 0.2953 0.7568 0.4354 0.9433 0.4710 LSC [18] 0.6979 0.3156 0.7908 0.4552 0.9503 0.4421 SNIC [16] 0.6659 0.3529 0.7662 0.4815 0.9410 0.4617 SCALP [19] 0.7307 0.3290 0.7977 0.4759 0.9499 0.4914 TASP 0.8706 0.4232 0.8139 0.4824 0.9503 0.4992 Table 2. TASP compared to the state-of-the-art methods.…”