As a benchmark for comparison, we recommend using the pre-specified training/testing split, which was generated randomly and independently for 5-fold cross validation. Finally, the mean accuracy and the receiver operating characteristic (ROC) curves will be used for evaluation.

Following the example of the LFW benchmark, we also design the following three different settings to evaluate the performance of different kinship verification algorithms:

  • Unsupervised setting: No labeled kin relation information is used.
  • Image-restricted setting: Only the given kin relation information is used in the training splits.
  • Image-unrestricted setting: The identity information of the person is available to potentially form additional negative pairs in the training splits.

Accuracy and ROC curves for various methods on these setting are available on results page.

Go To Top