Superpixels

1 Dec

SLIC Superpixels Compared to State-of-the-art Superpixel Methods, Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Süsstrunk. PAMI12. Website.

Achanta et al. analyze the requirement for superpixels, provide reasonable targets and performance measures, and compare a few of the leading approaches for superpixels (including Felzenszwalb and Huttenlocher’s approach and normalized cuts from Berkeley). They propose a new approach termed “simple linear iterative clustering” (SLIC). SLIC starts by defining cluster centers essentially in a grid over the image (with some smart initialization), assigning all remaining pixels to one of the cluster centers, recomputing the centers, and iterating. L2 distance over color and position is used to assign pixels to clusters. Only nearby cluster centers are checked for each pixel to keep things efficient.. I’m omitting some details, but that’s the gist of the algorithm. Couldn’t ask for anything simpler!

The result is a very fast algorithm (fraction of a second per image) and results are visually pleasing. Examples of superpixels at different scales (taken from author website):

54082_combo210088_combo302003_combo

The superpixels are fairly regular, compact, and provide good boundary recall (qualitative results are provided). Essentially all you could ask for from superpixels!

Source code (GPL unfortunately — people please stop using GPL so code can be used by all — consider BSD or LGPL) is provided on the project website. This is definitely THE algorithm I’ll use if ever I use superpixels in my code. Nice.

A great example of a simple algorithm that outperforms much more complex and computationally expensive approaches. (Note: similar ideas were proposed by my colleagues Larry Zitnick and Sing Bing Kang in their IJCV07 paper as the authors note). These ideas weren’t presented at a computer vision conference (indeed papers of this form often have trouble getting into CV conferences due to some misguided notion of “novelty”) — which is unfortunate because it may not get as much exposure as it deserves.

Going home tonight and will be thinking about applications for superpixels in sliding window object detection…

Advertisements

One Response to “Superpixels”

  1. Royi Avital May 10, 2015 at 11:06 pm #

    Looks very good.
    Thank You for the information.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s