SSIM/MSE and Arrest Characteristics

In this markdown I collect a set of correlation plots to check for potential relationships between the quality of the GAN projected mugshots and the predicted outcome. The hypothesis is this:

What we are testing With these plots I am probing the GAN. Essentially I am looking for indications on whether or not the ability of the GAN to produce a realistic mugshot correlates with underlying characteristics which we may care about.

• If image quality (SSIM or MSE) is correlated with our predicted probability then:
  • There may be structurally important features which the GAN is not adequately learning
  • This would make it difficult to curate the images we use for editing
  • Ideally we want to make sure that we have the ability to ‘move’ into any important direction without loosing image quality

Note These plots are an initial probe based on existing information. They do not check for more minute detail like hair color, tattoos, color of clothing, etc. These are meant as a sort of macro view.

With this in mind we ideally want to see no obvious correlation between image quality and any of these characteristics. Below I present plots for:

1. Overall predicted probability by SSIM and MSE
2. SSIM and felony flag
3. SSIM and finial arrest flag
4. SSIM and bond characteristics (these are potentially less important)
5. SSIM and demographics (these I believe are worth a look !)

None of these indicate any obvious correlations indicating that image quality is not ‘structurally’ lower across different demographic characteristics or arrest types.

---

---

Figure 1: SSIM and MSE with linked image level predicted probability

Figure 2: SSIM and Primary Arrest Characteristics

Figure 3: SSIM and bond characteristics

Figure 4: Demographics and additional arrest-type characteristics