ClusterFuG: Clustering Fully connected Graphs by Multicut

Authors: Ahmed Abbas, Paul Swoboda

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical evidence on instance segmentation on Cityscapes and clustering of Image Net datasets shows the merits of our approach.
Researcher Affiliation Academia 1MPI for Informatics, Saarland Informatics Campus, Germany 2University of Mannheim, Germany.
Pseudocode Yes Algorithm 1: GAEC (Keuper et al., 2015) (...) Algorithm 2: Dense GAEC (...) Algorithm 3: Incremental NN update
Open Source Code Yes 1Code available at https://github.com/aabbas90/ cluster-fug.
Open Datasets Yes We evaluate clustering of the Image Net (Deng et al., 2009) validation set containing 50k images. (...) We evaluate our method on the task of panoptic segmentation (Kirillov et al., 2019) on the Cityscapes dataset (Cordts et al., 2016).
Dataset Splits Yes We evaluate clustering of the Image Net (Deng et al., 2009) validation set containing 50k images. (...) from 500 images of the Cityscapes validation set.
Hardware Specification Yes We evaluate on NVIDIA A40 GPU with 48GB of memory. (...) All CPU algorithms are run on an AMD 7502P CPU with a maximum of 16 threads to allow for faster nearest neighbour search.
Software Dependencies No The paper mentions software like 'scikit-learn' and 'Mo Cov3' (Malkov & Yashunin, 2018; Johnson et al., 2019) but does not provide specific version numbers for any of its software dependencies.
Experiment Setup Yes For all multicut algorithms on all datasets we set the value of affinity strength αi in (11) to 0.4, preferring small clusters.