Adapting Appearance-Based Gaze Estimation to Narrow-Range, Long-Duration Screen Viewing

1. INTRODUCTION

• Prolonged screen viewing combines gaze concentrated near the screen center, limited head motion, and sessions long enough for eye-tracker calibration drift.The authors note that these conditions are largely absent from training data for current gaze models.Source: §1 · p. 1
• Because the recording-mean predictor is already accurate in narrow ranges, low mean angular error alone does not establish model robustness.The paper motivates evaluating more descriptive metrics alongside angular error.Source: §1 · p. 1
• Differential estimation was the only frame-wise system to significantly outperform the constant predictor while improving gaze metrics.The study evaluates synchronized facial video and eye tracking under a subject-independent protocol, with behaviorally grounded fixation- and saccade-level metrics.Source: §1 · p. 1

2. DATASET

• The PRECOG recordings include facial video at 195 Hz and EyeLink gaze at 250 Hz, averaging 75 minutes per participant across 190 hours.The original data collection included 146 college-aged adults completing two cognitive tasks.Source: §2 · p. 1
• Both the Stroop and sentence-processing tasks used the same camera configuration and display, with gaze predominantly near screen center and limited head movement.Sessions were long enough for participant pose and eye-tracker calibration to potentially drift; EyeLink annotations also provided fixation, saccade, and blink intervals.Source: §2 · p. 1

3.1. Benchmark Models

• Six published configurations span full-face and eye-region inputs, convolutional and transformer architectures, and static and temporal prediction.Each was retrained on PRECOG data while retaining its native input representation, gaze parameterization, and training procedure; two matched custom systems regress absolute gaze or gaze changes.Source: §3.1 · p. 2

3.2. Differential Gaze Estimator

• The differential estimator uses a weight-shared Direct CNN to encode two crops from the same recording, then maps the embedding difference to gaze displacement.Predicting changes allows additive recording-specific offsets to cancel from the training target.Source: §3.2 · p. 2
• At inference, the model reconstructs absolute gaze relative to a labeled anchor and averages predictions across anchors.The temporal variant contextualizes a five-frame window spanning 3.2 or 6.4 seconds with a bidirectional GRU before differential prediction.Source: §3.2 · p. 2

4.1. Pre-processing

• Temporal alignment and quality exclusions yielded 140 recordings from 100 participants, totaling approximately 115 hours and 80.4 million video frames.Camera-clock drift was corrected against the nominal frame rate and residual offsets through blink alignment; the common test set contained 144,743 samples from 105 recordings and 76 participants.Source: §4.1 · pp. 2–3

4.2. Training Protocol

• The differential encoder was initialized from an ImageNet-pretrained ResNet-18 and trained for four epochs with AdamW and a smooth-L1 displacement loss plus an amplitude penalty.Published models were retrained on the same subject partitions using their respective training objectives and procedures.Source: §4.2 · p. 3

4.3. Test-time Calibration

• Every model received 20 labeled frames per recording, distributed across the session; differential calibration used leave-one-anchor-out predictions to prevent label leakage.Anchors could occur later than the reconstructed query, so the primary protocol targets offline reconstruction or online tracking with distributed calibration phases.Source: §4.3 · p. 3

4.4. Evaluation protocol

• Evaluation measured great-circle angular error, within-recording yaw and pitch correlations, and fixation- and saccade-level displacement agreement.Saccades shorter than 1° were excluded, recordings with fewer than 20 eligible saccades were omitted, and participant-level comparisons used bootstrap intervals and corrected tests.Source: §4.4 · p. 3, Table 1 · p. 3, Table 2 · p. 3

5.1. Performance under narrow-range viewing

• The differential estimator alone among static models beat the constant predictor, reaching 1.616° angular error and the strongest yaw and pitch correlations, 0.636 and 0.447.Published estimators stayed near the 1.780° label-informed recording mean, with error ratios of 0.995–1.019.Source: §5.1 · p. 3

5.2. Matched-budget anchor comparison

• With the same 20 anchors, differential estimation reached 1.765° using anchor-based reconstruction and 1.616° after calibration, versus 2.991° for the direct CNN.Relative to the CNN, the differential model’s participant-level error reduction was 0.275°; Table 3 reports recording-macro mean angular error across 105 recordings.Source: Table 3 · p. 4, §5.2 · p. 4

5.1. Performance under narrow-range viewing

• Temporal context improved both model families without removing differential estimation’s advantage; its 6.4 s variant had the lowest angular error and highest yaw correlation, significantly better than its static counterpart.Gaze360-LSTM with seven frames also significantly beat the constant predictor, at an error ratio of 0.954.Source: §5.1 · p. 4

5.2. Matched-budget anchor comparison

• Frontloaded anchors increased average error across all systems by 32.1%.Source: §5.2 · p. 4

5.3. Alignment to event annotations

• Against Gaze360-LSTM, differential fixation correlations rose from 0.629 yaw and 0.533 pitch to 0.715 and 0.548, while median directional error fell from 35.5° to 31.2°.Figure 2 depicts centered fixation positions, predicted versus reference saccade amplitude, and displacement-angle distributions; saccades under 1° were excluded.Source: Table 4 · p. 4, §5.3 · p. 4, Figure 2 · p. 4
• With 6.4 s of temporal context, differential estimation further reduced median saccade directional error to 30.7°; its paired comparisons remained significant against published configurations.The event analysis covered 69,754 fixations and 15,111 saccades of at least 1°; Table 4 summarizes agreement across recordings with at least 20 saccades.Source: Table 4 · p. 4, §5.3 · p. 4
• Angular-error rankings and saccade agreement were only weakly aligned, so movement-level measures distinguish eye tracking from regression toward the center of the observed range.Source: §5.3 · p. 4

• Differential estimation achieved the lowest angular error and strongest fixation- and saccade-level agreement, with further gains from temporal context.The complete differential pipeline reduced error by 46% relative to the uncalibrated direct regressor.Source: §6 · p. 4

• The study analyzed previously collected human-subject data; the original study had institutional review board approval and participants gave written informed consent.No new data were collected for this work.Source: §7 · p. 5