Training pitch angle boundary models for InterFaceGAN using automated labeling
In short, the project can:
- Generate fake fotos using StyleGAN
- Estimate face rotation and label samples with the help of img2pose
- Train and test boundaries with InterFaceGAN for rotating the pitch angle of generated samples
Shown above: A random sample being rotated using 5 boundaries. Each trained with a filtered set of samples based on an absolute angle threshold (first to last row shows thresholds of 0, 5, 10, 15 and 20 respectively)
Contents
Notebook
pitch_boundary_creation.ipynb contains the iPython notebook used for generating the boundaries. Click here to open in Google Colab. When testing it is recommended to decrease the NUMBER_OF_IMAGES variable to a lower number (for example 100) to test that everything works as expected. The generation of samples and pose estimations are very time-consuming processes for a high number of samples.
Data directory
The data/ dir contains multiple npy files from a run of the notebook, generating 10 000 samples:
| Name | Content |
|---|---|
| data/random_samples/ | Contains two files for the z and w latent space codes for all generated samples |
| data/boundaries/ | Contains the boundaries generated with specified threshold |
| data/face_pitch_scores.npy | Array with labels (0 or 1) for positive/negative angles |
| data/face_pitch_latents_* | Corresponding latent vectors for each score label |
| data/all_face_angles.npy | All estimations containing rotation angles and translations for each sample |
face_pitch_* files and all_face_angles.npy all have the same lengths, and indices correspond to each other.