Accurancy

[vittoria-thomasini]

Hi ChenglongMa,

How you calculate the accuracy ? Are you using some study as reference for this ?

[ChenglongMa]

Hi @vittoria-thomasini,

Thanks for your question.

This is our original evaluation metric.

At the very beginning, we used a weighted Delta E (CIE 2000) ($\Delta E_{00}$) distance metric (see here) to measure the accuracy:

$$c^*=\underset{c\in C}{\arg \min} \sum_{d \in D}{\rho_d \cdot \Delta E_{00}(d,c)},$$
where

• $C$ is the set of color Categories in the specified palette. $c \in C$.
• $D$ is the set of detected Dominant colors in the specified portrait/image. $d \in D$.
• $\rho_d$ is the percentage of the skin area that the color $d$ occupies.

It measures the distance between the assigned color palette and the dominant colors in the face area. Values are between 0 and 100. A lower value indicates a more accurate classification.

However, sometimes the word Distance is not easy for non-technical users to understand, so I changed it to Accuracy.
Accordingly, the formula becomes:

$$accuracy = 100 - c^*$$.

Then, the values are still between 0 and 100, but the bigger the better.

Thanks!

We also explained this metric in our paper.