GazeFS: Target-Centered Gaze-Trajectory Forecasting and Stabilization from Gaze-Head History
Yaozheng Xia, Zaiping Zhu, Bo Pang, Minghao Xie, Hui Li, Shaorong Wang, Sheng Li
cs.HC · cs.AI
Quick verdict
GazeFS forecasts target direction and phase from gaze-head history; rigorous internal OOF shows modest single-device gains, with external baselines missing.
Summary
The paper formulates gaze correction for target-centered AR/VR interaction as online forecasting: from a variable-length gaze–head history (gaze yaw–pitch and angular rates, head rates; seven channels), GazeFS predicts the next-frame target-center direction and a short-horizon Search/Focus probability, using target geometry only for training supervision. Labels come from 7,960 confirmation-anchored acquisition episodes recorded at 60 Hz from 30 HoloLens 2 participants. The architecture combines fine/coarse temporal patches, a shared causal Transformer, parallel trajectory and phase heads, and short-prefix refinement for cold starts. Under five-fold participant-grouped out-of-fold evaluation across three seeds, GazeFS reduces Focus-stage episode bias, within-episode dispersion, median error, and P90 error relative to raw hold by 0.182°, 0.257°, 0.109°, and 0.400° (participant-bootstrap 95% CIs excluding zero). Endpoint-free replay from empty history preserves the advantage, and future-suffix audits confirm zero leakage. Controls show history adds spatial information beyond explicit progress metadata; ablations attribute gains to phase-aligned coordinate training. Limitations: validation is internal to one corpus, coverage is one device and one segmented task, no published external spatial-correction baselines are compared, frame-to-frame residual motion increases (+0.0578°/frame), and the 12-participant closed-loop study confounds state reset with learned correction, showing only a narrow benefit over One Euro. No code or data release is stated in the main text.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A target-agnostic online model can jointly forecast the next target-center direction and a short-horizon Search/Focus phase from gaze–head history alone.
- Under 30-participant five-fold grouped OOF across three seeds, GazeFS reduces Focus-stage episode bias, within-episode dispersion, median error, and P90 target error relative to raw hold by 0.182°, 0.257°, 0.109°, and 0.400°, with participant-bootstrap 95% CIs excluding zero.
- Recent gaze–head history carries spatial information beyond explicit progress metadata (history-only beats Endpoint+Progress by 0.159°/0.132°/0.098°/0.303°), and the gains arise from phase-aligned coordinate training rather than the auxiliary phase classifier.
- The advantage persists in endpoint-free replay from empty history (phase BA/AUPRC 0.925/0.993), with future-suffix audits showing exactly zero leakage.
- The method improves centering at a temporal-smoothness cost (+0.0578°/frame residual motion), which the authors explicitly treat as a separate objective.
Strengths
Unusually thorough anti-leakage evaluation design
Future-suffix invariance is audited on all 60 learned OOF checkpoints with a global maximum absolute difference of exactly zero; endpoint-free replay from empty history (Table 6) and exact-prefix evaluation that never counts backward from a known endpoint (Table 5) jointly rule out endpoint/future information entering inference.
Statistical discipline
Paired effects use 5,000 participant-bootstrap resamples; the four primary Full-minus-Raw outcomes get max-|T| simultaneous intervals; metrics are per-episode, three-seed averaged, equal participant weighted. Honest nulls are reported rather than buried: the 32-frame prefix contrast is 0.000 [−0.044, 0.043] (Table 5) and the 32-frame cold-start superiority is explicitly called non-uniform.
Mechanism-isolating controls
Honest scoping of claims
Weaknesses
No external published baselines
Internal validation only
Narrow external validity
Closed-loop evidence is weak and confounded
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
GazeFS (arXiv 2609.03868, cs.HC/cs.AI) addresses head-mounted AR/VR gaze input, distinguishing two error types the introduction argues are often conflated: short-timescale fluctuation and persistent target-relative offset. Its three claimed contributions are (1) a formulation of gaze correction as online target-centered trajectory forecasting, (2) the GazeFS architecture — multi-scale history encoding with parallel direction and phase outputs — and (3) an evaluation protocol demonstrating improved Focus-stage centering without future or endpoint information. Notably, the paper scopes "stabilization" as target-relative centering and explicitly excludes temporal smoothness, then reports the smoothness cost anyway. The honesty of claim-scoping is a consistent strength.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is GazeFS and what problem does it solve?
GazeFS is a model for AR/VR gaze input that takes a short history of the user's gaze and head-motion signals and predicts (1) where the true target center lies, expressed as a head-local direction, and (2) whether the user is entering the final confirmation (Focus) phase of a target acquisition. Its goal is to fix the systematic offset between the measured gaze ray and the intended target — something low-pass filters cannot do, since smoothing still centers on the noisy measurement — rather than merely reducing jitter.
How much does GazeFS improve gaze-pointing accuracy?
On Focus (confirmation) frames, compared with simply holding the raw gaze direction, it reduces episode bias from 1.010° to 0.828°, within-episode dispersion from 0.777° to 0.520°, median target error from 1.016° to 0.906°, and P90 error from 1.814° to 1.415°, in a 30-participant cross-validated evaluation on HoloLens 2 data. Gains are concentrated in sustained Focus; the first ~250 ms after onset is actually worse in P90 error.
Does GazeFS need to know where the target is at runtime?
No. Target geometry is used only offline to build training labels and evaluation metrics. The paper runs several checks against information leakage: an endpoint-free replay that starts from an empty history, an audit confirming outputs are unchanged when future samples are replaced, and a prefix evaluation that never counts backward from a known endpoint.
Is GazeFS code or its dataset publicly available?
The reviewed main text does not state any code or data release; much of the methodology detail is deferred to a supplement, and the dataset is a private 30-participant HoloLens 2 collection. Reproducing exact numbers would therefore require reimplementing the method and collecting comparable data.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The Supplement (Secs. A–F), where several load-bearing details live — label-threshold sensitivity (B.3), full paired/simultaneous CI tables (A.4/A.7), the three-seed shuffle audit (A.2), and ONNX latency (C.8) — is not part of the provided text; those specific claims were taken on trust.
- Figures 1–3 are assessed from their captions only; the graphical content itself was not extractable.
- codeAvailable=false reflects the absence of any code/data availability statement in the provided text, not an explicit statement that code is unavailable.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.