acceptodds
Under review as a conference paper at ICLR 2027

Fusing Complementary Multi-view Feature for Screen-based Eye Tracking

Abstract

Current multi-view gaze estimation remains limited by existing datasets, insufficient exploitation of complementary cross-view information, and evaluation focused primarily on average gaze error. We address these limitations through a more systematic study of multi-view gaze estimation. First, we introduce PrismGaze, a new dataset with over three million images, capturing continuous head-pose variation for the same gaze targets. Second, we propose PrismFusion, a multi-view feature fusion framework based on region partitioning, which masks complementary image regions across views during training to encourage effective cross-view information integration. Third, we develop a broader evaluation framework that examines the effects of camera number and placement, target location, viewing depth, and unseen viewpoints. Our experiments show that the primary benefit of multi-view gaze estimation comes from compensating for poorly observed views with cameras providing more favorable viewpoints. PrismFusion remains robust to changes in viewing depth and unseen viewpoints, while achieving strong calibration-free performance on seen views. Together, our dataset, method, and evaluation provide a more comprehensive foundation for studying multi-view gaze estimation in realistic settings.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Eye Tracking, Multi-view

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_LXP8OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies multi-view gaze estimation using PrismGaze, a dataset of over three million images that captures continuous head-pose variation for the same gaze targets. It proposes PrismFusion, which masks complementary image regions across views during training to encourage cross-view feature integration. The evaluation examines camera configuration, target location, viewing depth, and unseen viewpoints. The results suggest that well-positioned cameras can compensate for poorly observed views, while PrismFusion remains robust to depth changes and unseen viewpoints.

Critical strengths

This paper is easy to follow and well-written. The studied Multiview gaze estimation is an interesting and important problem.

Critical weaknesses

In this work, the authors claimed three contributions including one collected dataset, one multi-view feature fusion framework and one broader evaluation framework. Yet the above claimed contributions are insufficient for an acceptance of top-tier conferences with the following concerns. The first one is the collected dataset. The authors need to report the comparisons with the existing Multiview dataset such as Lian et al. 2019 published in TNNLS 2019. They need to report the comparisons in number of images, participants, to support their collected dataset’ advantages and necessity. The second one is the proposed method. The multi-view feature fusion framework might be lack of contributions and novelty, with an incremental combination of existing modules, such as shared encoder, cross-view fusion. They even used AI to draw the framework figure. The third contribution is the broader evaluation framework. However, the claimed evaluation framework is just ablation studies. It’s over-claimed.

Questions for authors

see weakness.

Reviewer_YdSFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces PrismGaze, a synchronized multi-camera dataset collected from 24 participants who vary head pose while fixating on predefined targets, containing approximately 3.24 million raw images. It proposes PrismFusion, which combines full-image and local face/eye features with dual-token cross-view attention. During training, Region Partition distributes visible face-grid regions across retained cameras to encourage complementary feature use. Experiments compare methods on PrismGaze and report results on ETH-XGaze, then analyze camera number, placement, target location, and viewing distance. PrismFusion reports 0.94 cm error in the eight-camera PrismGaze setting versus 1.33 cm for the adapted DVGaze baseline, and 1.63 cm when Region Partition is removed. A separate 12-participant study evaluates distance changes and per-participant, per-distance affine calibration.

Critical strengths
  1. Capturing multiple head poses for the same fixation target addresses a meaningful weakness of sparse target-pose sampling. The analysis also goes beyond a single average error to examine camera quality, camera composition, spatial target effects, and a separate participant cohort. The reported finding that additional cameras do not invariably help is practically informative and appropriately visible in the ETH-XGaze results.

  2. Removing Region Partition while keeping the stated backbone configuration fixed increases error from 0.94 to 1.63 cm in Table 3. Together with the within-dataset comparisons in Table 1, this is concrete evidence that the training intervention improves the reported PrismGaze setting. It warrants further investigation even though the specific explanation in terms of complementary cross-view integration is not yet fully isolated.

Critical weaknesses
  1. The separate-cohort experiment does not support a broad robustness conclusion without further explanation and selection controls. Table 5 reports uncalibrated PrismFusion error of 9.62 cm at 60 cm, compared with 0.94 cm on the main test split at approximately the same viewing distance. These are different cohorts and protocols, so the discrepancy is not proof of an implementation error, but it is a major generalization gap that the paper does not explain. The best practical results additionally use a separate affine calibration for each participant and distance. Section 5.7 selects the two-camera pair with the lowest mean error across the three evaluated distances, without stating an independent selection set. This raises an additional concern about optimistic camera selection. The current results establish benefits under calibration more clearly than calibration-free robustness to new users and distances.

  2. The evidence does not yet cleanly identify the proposed fusion mechanism or establish a reproducible external comparison. The no-partition ablation is valuable, but there is no matched control for independent regional masking or an equally strong generic masking regularizer, and the separate feature/weight-token design is not isolated. Attention-error correlations are consistent with the proposed explanation but do not establish that complementary information use causes the gain. Table 2 also explicitly combines published ETH-XGaze results from differing protocols, while its own 'Within' and 'Person Specific' settings are not adequately defined. The manuscript omits key training and baseline-adaptation details, and Section 4.3 names DINOv3 ViT-S while Figure 2 labels the full-image backbone ViT-B. These gaps make it difficult to determine which parts of the improvement would survive a controlled reproduction.

Questions for authors
  1. GWhat changes between the main PrismGaze test and the separate 60 cm cohort explain the 0.94-to-9.62 cm increase? Please specify the checkpoint, preprocessing, coordinate convention, target distribution, and participant protocol used in both. Were the two-camera pair and any refinement choices selected on independent validation data? If not, what are the separate-cohort results with a pair fixed without access to those evaluation errors?

  2. Does Region Partition still improve accuracy over independent masks matched for retained area and camera dropout, and over simple averaging with the same encoders? Please clarify whether dropout is held fixed in Table 3 and quantify the contribution of the separate feature/weight-token attention. These comparisons would determine whether complementary cross-view partitioning adds value beyond regularization and backbone choice.

  3. Please define the ETH-XGaze 'Within' and 'Person Specific' protocols, including calibration or adaptation data, coordinate conversion, and aggregation across camera pairs. Also provide exact PrismGaze subject counts and per-subject results, the actual DINO variant, training settings, and baseline adaptations. Can the strongest comparator be evaluated under the identical ETH-XGaze protocol? The current quoted results are explicitly not a controlled ranking.

Reviewer_Zg7fOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies multi‑view gaze estimation for screen‑based eye tracking. It claims that it construct a large‑scale multi‑view gaze dataset with over 3 M synchronized images. It also introduce PrismFusion, a multi‑view feature fusion framework with a novel region‑partition training strategy

Critical strengths

(1) The proposed dataset looks very useful with such a big size.

(2) The authors openly discuss dataset constraints (lab environment, no glasses, synchronized multi‑camera hardware cost), which helps readers understand the boundary of the work.

Critical weaknesses

(1) No analysis and visualization for the proposed dataset. Although the authors claim that they propose a very large dataset. However, I do not see any analysis about the dataset. I recommend the authors to provide the distribution of head pose, gender, age and so on. Meanwhile, some visulization images should be provided. I am also wondering if the dataset will be released publically.

(2) Poor method visualization. At stage 1 (Region partition), the input images looks like compeletly AI-generated face images. Why noy just use a real face image from your dataset?

(3) The proposed method is not very novel. It looks like the proposed method is just a three branch network with a cross-attention machnism, which is quite common in the gaze estimation area.

(4) The role of the Figure 1. Figure 1 occupies a large area in the main paper with limited information. I recommend to put the figure in the supplymental material.

Questions for authors

See Weaknesses.

1 comment

Since no response is provided, I keep my rating of reject.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.