Apparent Personality Prediction from Facial Images A Literature-Based Study of Computer Vision Approaches

June 22, 2026

This thesis studies apparent personality prediction from static images derived from the ChaLearn First Impressions benchmark. The task is formulated as supervised regression from RGB images to the official segment-level Big Five labels. These labels represent observer-attributed apparent personality impressions and are not treated as measurements of stable or self-reported personality traits. The work proposes a reproducible pipeline that converts the original video benchmark into a static-image experimental setting. The pipeline preserves the link between each extracted image, its source segment, the official labels, and the data split. To reduce information leakage, a source-video-level split is constructed. The main representation uses full frames resized to 224 × 224 pixels, while a parallel face-crop representation is evaluated as an input variant. Three pretrained convolutional architectures are compared: ResNet50, EfficientNetB0, and ConvNeXt-Tiny. Models are trained at image level, whereas the main evaluation is performed at segment level by averaging predictions across the images sampled from each segment. All trained models improve over a training-label mean baseline. The strongest single run obtains a test MAE of 0.1019, and a post-training ensemble reduces test MAE to 0.0981. The results indicate that static images retain relevant signal for apparent personality prediction, although the setting is not equivalent to the official ChaLearn challenge protocol.

Authors in the community:

• 

  J

  João Antonio Martins Adorno

  ist1101215

Supervisors of this institution:

• 

  J

  João Fernando Cardoso Silva Sequeira

  ist13132

electrical-engineering-electronic-engineering-information-engineering - Electrical engineering, electronic engineering, information engineering

eng - English

Instituto Superior Técnico