BROWSE
Detail
Conference Paper
GazeFlow: Diverse Driver Gaze Synthesis via Conditional Flow Matching
- Authors
- Yonghyun Kim, Woojin Kim, Hyunsuk Kim, Eun Hye Jang, Daesub Yoon
- Issue Date
- 2026-07
- Citation
- ACM SIGGRAPH 2026, pp.1-3
- Publisher
- ACM
- Language
- English
- Type
- Conference Paper
- Abstract
- Modeling diverse human gaze behavior is a fundamental challenge in behavioral simulation and character animation for interactive graphics and digital humans. In safety-critical applications such as Driver Monitoring Systems (DMS) under the European New Car Assessment Programme (Euro NCAP) 2026, existing synthesis approaches produce deterministic outputs that lack the stochastic diversity of real drivers. To overcome this limitation, we cast gaze-motion synthesis as a stochastic coverage problem anchored to a measured human intra-condition diversity baseline of Dhuman = 1.94, defined as mean pairwise trajectory distance under identical conditioning, computed from 59, 405 condition-matched subject pairs across 689 scenario groups. A conditional flow matching model over 5D gaze-head trajectories recovers \(93\%\) of this baseline at our human-calibrated operating point (σ = 0.8). A strong conditional variational autoencoder (CVAE) baseline, under matched conditioning, saturates at \(41\%\) of human diversity regardless of latent scale. The sampling noise scale σ further enables continuous fidelity–diversity trade-off control: at σ = 0.6 the model achieves absolute position error (APE) within \(7\%\) of the CVAE while covering \(64\%\) of human variability.
- KSP Keywords
- Absolute position, Character animation, Driver monitoring, Flow matching, Gaze behavior, Gaze synthesis, Human gaze, Interactive graphics, Matching model, Monitoring system, Motion synthesis
This work is distributed under the term of Creative Commons License (CCL)
(CC BY)
(CC BY)
