SLIC Superpixels Compared to State-of-the-Art Superpixel Methods

By Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurélien Lucchi, Pascal Fua, Sabine Süsstrunk. 2012

9,160 citations recorded by the source corpus.

Subfields: Computer Vision and Pattern Recognition, Media Technology

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Media Technology

Citations in the corpus

Learning Features by Watching Objects Move (2016, background): “Per-pixel saliency is then averaged over superpixels [1].”
Minding the Gaps for Block Frank-Wolfe Optimization of Structured SVMs (2016, methodology): “In addition to images and their pixel-level annotations, Kolesnikov et al. (2014) released15 oversegmentations (superpixels) of the images precomputed with the SLIC algorithm (Achanta et al., 2012) and the unary features of each superpixel computed similarly to the work of Lempitsky et al. (2011).”
Feedforward semantic segmentation with zoom-out features (2014, methodology): “In our experiments we use SLIC [1], but other methods that produce nearly-uniform grid of super-pixels might work similarly well.”
Interpretations are useful: penalizing explanations to align neural networks with prior knowledge (2019, methodology): “To obtain the binary maps of the patches for the skin cancer task, we ﬁrst segment the images using SLIC¡ a common image-segmentation algorithm (Achanta et al., 2012).”
Higher Order Conditional Random Fields in Deep Neural Networks (2015, methodology): “The Boundary IoU is slightly lower than the baseline, and this may be due to the fact that superpixels do not always align correctly with the edges of an object (the “boundary recall” of various superpixel methods are evaluated in [1]).”
Deep convolutional neural fields for depth estimation from a single image (2014, methodology): “We use SLIC [24] to segment the images into a set of non-overlapping superpixels.”
Unsupervised Learning of Edges (2015, methodology): “Specifically, we utilize SLIC super-pixels [1], which cover over 90% of all image edges, and match motion and superpixel edge pixels using bipartite matching (also used in BSDS evaluation) with a tolerance of 3 pixels.”
Connectionist Temporal Modeling for Weakly Supervised Action Labeling (2016, methodology): “We apply k-means clustering to frames in a way similar to SLIC [1] to over-segment the video.”
Deeper Depth Prediction with Fully Convolutional Residual Networks (2016, methodology): “[29], uses a MRF to infer depth from local and global features extracted from the image, while superpixels [1] are introduced in the MRF formulation in order to enforce neighboring constraints.”
Streaming Weak Submodularity: Interpreting Neural Networks on the Fly (2017, methodology): “Here, the features are superpixels determined by the SLIC image segmentation algorithm [Achanta et al., 2012] (regions from any other segmentation would also suffice).”
DeepSaliency: Multi-Task Deep Neural Network Model for Salient Object Detection (2015, methodology): “During the saliency regression process, the superpixel oversegmentation is carried out by the SLIC method [39] with the superpixel number N being 200.”
Discriminatively Trained Dense Surface Normal Estimation (2014, methodology): “Methods producing segments of large variation in size [20], capturing information over the right scale, may look visually more appealing, but the feature representations obtained using methods producing similar sized segments [22,23] may be more stable, and thus more discriminative.”
Efficient Joint Segmentation, Occlusion Labeling, Stereo and Flow Estimation (2014, methodology): “Our segmentation method also incorporates a length of boundary energy for shape regularization, as well as the evidence from the stereo and video pairs, which SLIC does not.”
SEEDS: Superpixels Extracted Via Energy-Driven Sampling (2012, methodology): “For instance, [12] reports a 5% tolerance margin for the overlap; and in [6] the boundaries are removed from the labeling before computing the UE.”