Min-Cut/Max-Flow Benchmark Instances for Computer Vision

Patrick M. Jensen and the Technical University of Denmark compiled this collection of min-cut/max-flow problem instances to benchmark algorithms. The dataset aggregates instances from multiple sources, including the University of Waterloo's repository and specific computer vision tasks like super-resolution and multi-object segmentation. It is released alongside the 2022 T-PAMI paper 'Review of Serial and Parallel Min-Cut/Max-Flow Algorithms for Computer Vision'.

Use Cases

• [Benchmarking min-cut/max-flow algorithms] based on [a wide selection of representative problem instances]
• [Evaluating segmentation algorithm performance] based on [sparse layered graph and multi-object surface fitting instances]
• [Comparing algorithm efficiency on dense vision problems] based on [super-resolution, texture restoration, and deconvolution instances]
• [Testing graph algorithm scalability] based on [instances provided in both DIMACS and a custom binary format]

Strengths

• [Provides a unified collection] from [multiple previously scattered or unavailable sources], addressing link rot
• [Includes instances] from [specific, cited computer vision papers] such as Verma & Batra (2012) and Jeppesen et al. (2020)
• [Offers data in two formats] including [the standard DIMACS format and a custom binary format]

Limitations

• [The dataset is explicitly noted as outdated] with a link provided to an updated version containing more problems
• [Specific row counts, file sizes, and column details are not provided in the available metadata]
• [Some original problem instances from the University of Waterloo] are [unavailable due to dead links]

Provenance

Source

Technical University of Denmark (compiled by Patrick M. Jensen), aggregating instances from University of Waterloo and specific computer vision publications.

Collection Method

Collected from a wide selection of sources to be representative, including public repositories and problem instances from published papers.

Time Range

Instances span publications from at least 2012 to 2020.