Skip to content

FaceSim3D – code🗿

Period Status Author
Feb, 2022 - Sep, 2024 in process Simon M. Hofmann

Last update

Analysis steps🗿

Representational Similarity Analysis (RSA)🗿

RSA is applied on similarity matrices of face-pairs in both viewing conditions (2D, 3D) that were computed on the following associated data:

Behavioral judgments🗿

Human judgments of face similarity from the triplet odd-one-out task.

VGG-Face activation maps🗿

Face images both in original and 3D-reconstructed form are fed into a pre-trained VGG-Face network. The resulting activation maps (of all layers) were used to compute similarity matrices.

Cognitive models🗿

Modeling human similarity judgments in both viewing conditions (2D & 3D). To this end, sparse and deep encoding models were used to predict human similarity judgments.

Human-aligned VGG-Face activation maps🗿

An adaptation of the deep neural network VGG-Face was aligned to human choices in the face similarity judgment task. Also, for this embedding model, similarity matrices were computed based on feature maps of the model.

Sparse embedding models: SPoSE & VICE🗿

Sparse models were developed for modeling human similarity judgments in similarity tasks, see Hebart et al. (2020), and Muttenthaler el al. (2022) for details. We used the SPoSE and VICE models to predict human similarity judgments and computed similarity matrices based on the model's predictions.

Physical face features🗿

Original face stimuli are from the Chicago Face Database (CFD).

CFD faces

Each face is associated with a large set of physical face features (e.g., length of the nose). That is, each face can be described by a high-dimensional feature vector. These vectors have been used to compute face similarity matrices.

FLAME and DECA dimensions🗿

The 3D-reconstructed faces have corresponding FLAME and DECA dimensions. These dimensions have been used to compute further similarity matrices.

Code structure🗿

The analysis code can be installed as a Python package facesim3d.

How to set up the research code🗿

See the section Analysis package for how to install the research code as a 🐍 Python package.

See the section API reference for more information on the analysis code of the study.