search

mapooon / BlendFace

[ICCV 2023] BlendFace: Re-designing Identity Encoders for Face-Swapping https://arxiv.org/abs/2307.10854
Other
170 stars 7 forks source link

visualization #6

Closed lizian1611 closed 9 months ago

lizian1611 commented 1 year ago

Figure 9 in your paper visualizes the images of ArcFace and BleedFace. If convenient, could you please provide some details? I really appreciate my interest in this aspect, thank you!

mapooon commented 11 months ago

Sorry for the late reply.

We conduct the following procedure:

  1. We compute identity similarity between input and reference. Let this score be a.
  2. We mask a region of the input and compute identity similarity to reference. Let this score be a. This measures the importance of the masked region in terms of identity. We define the importance as a-b.
  3. We slide the masked region and compute id sim in the same manner as step 2. We iterate this process for all the regions of the input.
  4. We change the mask size and do step 2 and 3. We adopt sizes of {16, 24, 32, 40, 48, 56}.
  5. We average the importance for all the regions over the masks and visualize it as a heatmap in figure 9.