# DAVIDSTUTZ

03rdFEBRUARY2018

Liu et al. propose intrinsic manifold SLIC (IMSLIC) based on their earlier work [1]. While the authors introduce the theoretical background of performing SLIC on a 2-manifold in $\mathbb{R}^5$ (i.e. coordiantes + color), the overall algorithm is very similar to the original SLIC [2] algorithm. Essentially, stretching factors $\lambda_1$ and $\lambda_2$ are used to define the map
$\Psi(p, c) = (\lambda_1 p, \lambda_2 c)$
where $p$ is the spatial position and $c$ the color of a pixel. Then, SLIC is applied using a geodesic distance instead of Eucliean distance. Additionally, the local search area around each seed is augmented using the stretch factors – which are computed per seed. In contrast to SLIC, IMSLIC additionally uses random initialization on the 2-manifold. Overall, the proposed algorithm produces qualitatively reasonable superpixels, see Figure 2, and is shown to outperform SLIC – in total, the authors compare 11 superpixel algorthms. The proposed method has several advantages over the original SLIC algorithm: it generates exactly the desired number of superpixels, the initialization adapts automatically to the image content and it performs slightly better. However, I also want to note that the superpixels – in practice, i.e. Figure 1 – look very similar.