POSTER DETAILS
Voxelwise encoding model reveals 2D key points like representation in extrastriate body area
Giuseppe Marrazzo, Federico De Martino, Agustin Lage Castellanos, Maarten Vaessen, Beatrice De Gelder
Date / Location: Sunday, 10 July 2022 / S01-161
Add this poster event to your π
Add this poster event to your π
Aim: Currently, the extrastriate body area (EBA) is considered to be an object category area that represents body stimuli. However, our understanding of body posture shows a gap between processing of low-level features and high-level semantic categories. This mid-level gap might be filled by the role played by EBA in the processing of body postures. In this study, we used voxelwise encoding model to investigate the role of EBA in body processing. Methods: Twenty participants viewed body stimuli while 7T fMRI responses were recorded. The stimuli were generated using a variational autoencoder (VAE), via random sampling from the latent space parameters from 3 different viewpoints. The fMRI response was modeled using several features extracted from the stimuli, such as VAE representation (encoding/decoding layers, latent space), 2D/3D coordinates of key joints (kp2d/kp3d), pixel space (Gabor like representation). The fMRI predicted responses from each model were generated via banded ridge regression using crossvalidation. Results: Results show a pattern of responses across visual cortex with Gabor and kp2d model which best predict responses to our stimuli. Specifically, the Gabor representation shows higher prediction accuracy in early visual occipital area as opposed to the kp2d representation which shows higher prediction accuracy in high-level temporal areas such as EBA. Furthermore, kp2d model (viewpoint variant) shows higher accuracy in EBA than kp3d model (viewpoint invariant). Conclusion: These findings suggest that EBA codes for specific features of the body, which in the case of kp2d model, are the joints position, and that this representation is viewpoint variant.