Quickshift++: Provably Good Initializations for Sample-Based Mean Shift

Authors: Heinrich Jiang, Jennifer Jang, Samory Kpotufe

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We then show strong clustering performance on real datasets as well as promising applications to image segmentation.Despite the simplicity of our approach, we show that Quickshift++ considerably outperforms the popular density-based clustering algorithms, while being efficient.5. Simulations, 6. Image Segmentation, 7. Clustering Experiments
Researcher Affiliation Collaboration 1Google Research, Mountain View, CA 2Uber Inc, San Francisco, CA 3Princeton University, Princeton, NJ.
Pseudocode Yes Algorithm 1 MCores (estimating cluster-cores) and Algorithm 2 Quickshift++
Open Source Code Yes Code release is at https://github.com/google/quickshift.
Open Datasets Yes Datasets Used: Summary of the datasets can be found in Figure 8. Seeds, glass, and iris are standard UCI datasets (Lichman, 2013) used for clustering... We also used a small subset of MNIST (Le Cun et al., 2010) for our experiments.
Dataset Splits No The paper runs experiments on various real datasets and scores against groundtruth, but it does not explicitly provide details about train/validation/test splits, percentages, or sample counts for reproduction.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models or other system specifications used for running the experiments.
Software Dependencies No The paper mentions general software like 'scikit-learn' but does not provide specific version numbers for any key software components or libraries required to replicate the experiments.
Experiment Setup Yes Quickshift++ settings were fixed at k = 20, β = 0.7 for all the datasets (Figure 3). β = 0.9 is reasonable across a wide range of images, β was fixed to this value for segmentation here. (Section 6). Quick Shift was set with bandwidth 10 and Quickshift++ was set with k = 300 and β = 0.9. (Figure 4 caption). β = 0.3 for all but one of the datasets (Section 7). The procedures were tuned in their respective essential hyperparameter: k-means number of clusters, DBSCAN epsilon, MCores k, mean shift bandwidth, quick shift bandwidth, Quickshift++ k. (Figure 9 caption)