| A Fast Method for Inferring High-Quality Simply-Connected Superpixels|
Superpixel segmentation is a key step in many image processing and vision tasks. Our recently-proposed connectivity- constrained probabilistic model  yields high-quality superpixels. Seemingly, however, connectivity and parallelized inference cannot coexist. Thus, the implementation from  is serial, hence slow. The contributions of this work are as follows. First, we show that effective parallelization is in fact possible. This leads to a fast GPU implementation that scales gracefully with both the number of pixels and number of superpixels. Second, we show that the superpixels are improved by replacing the fixed and restricted spatial covariances from  with unrestricted Bayesian estimates. Quantitative evaluation on public benchmarks shows the proposed method outperforms the state-of-the-art.
People Involved: Oren Freifeld, Yixin Li, John W. Fisher III
[ Paper ] [ Code (with wrappers for either Python, C++ or Matlab) ] Timing comparisons with other methods (for our method, we used the python wrapper; the C++ wrapper is slightly faster): Both the abscissa and ordinate are scaled logarithmically. After gSLIC, the proposed method is the fastest. Quantitatively, however, the proposed method outperforms all methods; see below. Quantitative results: the abscissa is scaled logarithmically. The proposed method performs best excepting the second plot. There, CC-GMM slightly outperforms it due to computationally expensive (see the figure above) split-and-merge moves. Interestingly, via the improved spatial-covariance modeling, the proposed method outperforms CC-GMM-SM on the other 3 indices. A visual-comparison example: Additional examples in higher resolution: [ original (file size: ~38MB) ] [ slightly compressed (file size: ~5MB) ]