Multi-frame Labeled Faces Database
Benchmark
This benchmark was created for the scientific purposes while this database was published and initial
results were done by our team. These results are considered as a fundamentals for this research
area and they are shown in leader-board table. There are 2,500 image sequences in the test set and
all resulting images are compared with ground truth images and evaluated using several metrics. Mean
of the values for given metric is used for representation as a single value result. Overall score is
then computed as a sum of normalized metrics results. The lowest value is considered as better, in
case observed metric where higher value is better (SSIM, PSNR), the normalized value is computed as
1 - norm(x). Following metrics are computed with min and max
values for normalization:
- Mean Square Error (MSE). Normalization min = 0, max =
500.
- Structural Similarity (SSIM). Normalization min = 0, max
= 1; (score = 1 -norm(x)).
- Peak Signal to Noise Ratio (PSNR). Normalization min = 0,
max = 50; (score = 1 -norm(x)).
- Sharpness difference (CPBD). Normalization min = 0, max =
1.
- Face feature extraction rate (FR failed) - using
well-known FaceRecognition
framework. This metric says how many predicted images failed to
extract face features needed for face recognition (using CNN model). There is no mean for
this metric. Normalization min = 0, max = 2500.
All metrics and overall score are computed automatically and detailed results are available via
score hyperlink in leaderboard table. There is .json file containing means of all metrics and their
normalized values. CSV file with all measurements will be send by email.