HAGI: Head-Assisted Gaze Imputation for Mobile Eye Trackers
summary
A diffusion-based imputation model uses time-aligned head orientation (rotation + translation) and observed gaze to reconstruct missing gaze segments, training on full trajectories from AMAR and testing on Nymeria, Ego-Exo4D, and HOT3D. It uses a head-movement feature extractor, a hybrid fusion module, and a tailored loss, operates on 5 s windows, and handles 10–90% data loss, multi-modal diffusion, reduces mean angular error by up to 22%, matches gaze velocity distributions better (lower JS divergence), but requires GPU and is not real-time.
AI-written summary · about
chapters (10)
- 0:00Introduction and motivation
- 1:07Missing data in eye tracking
- 2:33Introducing Hagi approach
- 4:17Dataset and evaluation setup
- 5:41Results and comparisons
- 6:46Visualizations and extreme cases
- 7:26Head rotation vs translation
- 7:49Limitations and future work
- 8:32Summary and internship call
- 8:53Q&A session
related talks
38mWhere are we looking? Predicting human gaze using deep networks.
48mEye Tracking (a Deep Learning Approach)
16mEye-Trace: Segmentation of Volumetric Microscopy Images with Eyegaze
06mMultimodal Joint Head Orientation Estimation in Interacting Groups via Proxemics and Interaction ...
19mWhat Can Be Predicted from Six Seconds of Driver Glances?
08mSo Predictable! Continuous 3D Hand Trajectory Prediction in Virtual Reality
32mHeart to Heart and Eye to Eye with Deep Fakes
15mC-Face: Continuously reconstructing facial expressions by deep learning contours of the face ...