Comments on the "Kinship Face in the Wild" Data Sets

Area of Science:

• Computer Vision
• Biometrics
• Artificial Intelligence

Background:

• The Kinship Face in the Wild datasets are widely used benchmarks for kinship verification.
• Recent studies highlight potential biases within these datasets affecting algorithm evaluation.

Purpose of the Study:

• To critically evaluate the suitability of the Kinship Face in the Wild datasets for kinship verification research.
• To demonstrate the impact of image properties, specifically cropping, on algorithm performance.

Main Methods:

• A simple scoring method based on image similarity was employed.
• The method calculated the distance of chrominance averages in the Lab color space.
• No specific kin features or training were utilized.

Main Results:

• Performance comparable to state-of-the-art kinship verification methods was achieved using the simple scoring approach.
• The results indicate that image similarity, derived from cropping, significantly influences performance.
• This suggests the datasets may not accurately reflect true kinship verification capabilities.

Conclusions:

• The Kinship Face in the Wild datasets should be used with caution in kinship verification research due to inherent biases.
• The findings necessitate a re-evaluation of current benchmarks and methodologies in kinship verification.
• Further research should focus on developing less biased datasets and algorithms that capture genuine kinship cues.