Extended Berkeley Segmentation Benchmark

The Berkeley Segmentation Dataset and Benchmark is a standard benchmark for image segmentation that includes multiple ground truth segmentations per image, as shown in Figure 1. While the benchmark includes some standard metrics such as precision, recall and F1-score, these are usually not appropriate to evaluate superpixel segmentations. As part of my bachelor thesis, I extended the benchmark to include several superpixel metrics such as undersegmentation error, achievable segmentation error, compactness and explained variation. The extended benchmark can be found on GitHub:

Figure 1: An example image with multiple ground truth segmentations.

Extended Berkeley Segmentation Benchmark on GitHub extended-bsds