Session-Adaptive Bias Correction and Avatar Pre-Distortion for Shared Attention in VR

Master Thesis open

Project overview

In our recent VR dyad study, 41.7 percent of the variance in Receivers’ attention-decoding error was systematic: a smooth, largely target-dependent bias field. A static correction recovered part of this bias on aggregated targets but backfired on individual trials because per-trial noise dominates. The study therefore proposes, but does not build, two systems: a session-adaptive probabilistic corrector that accumulates evidence about a Receiver’s bias and falls back to no correction when data are sparse, and avatar pre-distortion, where the system renders the Sender as attending to a pre-shifted location so the Receiver’s own bias lands their estimate on the truth. This project builds and evaluates them.

Project motivation

Systematic-but-noisy error is exactly the regime where naive correction hurts and probabilistic correction shines, a lesson established for mid-air pointing [1]. Pre-distortion additionally has a perceptual budget: avatar gaze and pose can only be redirected so far before users notice, and recent detection-threshold work quantifies that budget [2]. Combining both lines into a working attention-correction system, with honest uncertainty handling, is a strong systems-plus-evaluation thesis, and the parent study’s dataset (3,631 trials with full bias fields) provides everything needed for offline development before any new data collection.

Project goals

  1. Offline corrector. Develop and cross-validate a per-Receiver, uncertainty-aware bias model on the existing dataset (for example a Gaussian-process or hierarchical Bayesian field over target direction), quantifying when correction helps versus harms as calibration data accumulate.
  2. Online evaluation. Integrate the best corrector into the Unity environment in both variants (post-hoc correction of estimates and avatar pre-distortion within perceptibility thresholds) and evaluate with a live dyad study.

You will

  • Perform a literature review on offset correction and gaze redirection
  • Develop probabilistic correction models on the existing dataset (Python)
  • Integrate the corrector into the existing Unity VR environment
  • Run an evaluation study and quantify decoding-accuracy gains
  • Summarize your findings in a thesis and present them to an audience
  • (Optional) co-write a research paper

You need

  • Strong communication skills in English
  • Solid Python and machine-learning or Bayesian-modeling skills
  • Good knowledge of Unity

References

  1. Mayer, S., Schwind, V., Schweigert, R., & Henze, N. (2018). The effect of offset correction and cursor on mid-air pointing in real and virtual environments. CHI 2018. https://doi.org/10.1145/3173574.3174227
  2. Schott, D., et al. (2025). Estimating Detection Thresholds of Being Looked at in Virtual Reality for Avatar Redirection. CHI 2025. https://doi.org/10.1145/3706598.3714041
  3. Bailenson, J. N., Beall, A. C., Loomis, J., Blascovich, J., & Turk, M. (2005). Transformed Social Interaction, Augmented Gaze, and Social Influence in Immersive Virtual Environments. Human Communication Research. https://doi.org/10.1111/j.1468-2958.2005.tb00881.x
  4. The parent study on avatar-mediated shared attention in virtual reality; details are available from the advisor on request.

Keywords: VR, Bias Correction, Bayesian Modeling, Avatar Pre-Distortion, Gaze Redirection

Interested in this topic? Reach out through your university student email address via the contact form, with a short motivation, your transcript of records and, if available, a CV.

← All topics