Everyday Eye Contact Detection Using Unsupervised Gaze Target Discovery
summary
Combines an appearance-based gaze estimator with unsupervised gaze target discovery that mounts one camera to each candidate object, projects estimated gaze onto the camera plane, and uses density clustering to pick the cluster nearest the camera as positive. Trains an eye-contact classifier on a 4,096-d face feature after automatically labeling samples. Evaluates on 14 participants × 3 objects (tablet, display, clock) with 3 display camera positions and a head-mounted set (3 participants, 28 interactions), annotates 5,000 test frames, reports results with MCC score and shows best performance on tablets and robustness across camera positions.
AI-written summary · about
chapters (10)
- 0:00Introduction and motivation
- 1:11Related work limitations
- 3:21Proposed unsupervised approach
- 8:02Method implementation details
- 8:22Evaluation datasets
- 10:59Baselines and results
- 13:37Clustering and labeling examples
- 14:40Unsupervised clustering results
- 15:26Performance across objects
- 16:28Demo and conclusion
related talks
19mGazeProjector: Accurate Gaze Estimation and Seamless Gaze Interaction Across ...
19mAggreGaze: Collective Estimation of Audience Attention on Public Displays
20mBuilding a Personalized, Auto-Calibrating Eye Tracker from User Interactions
17mEyeGrip: Detecting Targets in a Series of Uni-directional Moving Objects Using Optokinetic ...
21mSelf-Calibrating Head-Mounted Eye Trackers Using Egocentric Visual Saliency
20mEye-tracking and AI: Unveiling the Limitations and Synergies
18mOrbits: Gaze Interaction for Smart Watches using Smooth Pursuit Eye Movements
48mEye Tracking (a Deep Learning Approach)