acceptodds
Under review as a conference paper at ICLR 2027

EdgeGaze: Efficient End-to-End Multi-Person Gaze Target Detection

Abstract

Gaze target detection on edge devices requires identifying where every person in a scene is looking under tight memory and latency budgets. Existing two-stage methods rely on external head detection and repeated per-person gaze estimation, while recent end-to-end models can grow to hundreds of millions of parameters. We introduce EdgeGaze, a compact end-to-end model built on EdgeCrafter that jointly detects multiple people and predicts their gaze targets in a single forward pass, with computation nearly independent of scene crowding. Gaze is strongly shaped by scene structure: people often look at other people, salient regions, and meaningful objects. We exploit this through Scene Prior Guidance, using saliency, object, and head priors to learn a Prior-Guided Scene Representation (PGSR) during training. PGSR is fused with person queries for gaze prediction, while the external prior models are discarded at inference. We further introduce a Unified Gaze Distribution (UGD), which places in-frame gaze locations and the out-of-frame outcome in a single normalized prediction space. EdgeGaze achieves state-of-the-art performance on GazeFollow and VideoAttentionTarget, with strong zero-shot transfer to GOO-Real and ChildPlay. On GazeFollow, EdgeGaze-M achieves 0.960 AUC and 0.798 mAP with only 26.7M parameters and 4.9 ms latency, compared with 0.929 AUC and 0.626 mAP for the 862M-parameter GazeHTA. EdgeGaze-N further reduces the model to 8.5M parameters and 3.1 ms. These results show that accurate multi-person gaze target detection can be compact, fast, and fully end-to-end, making it well suited for edge deployment. Code and models will be released; demo: https://anonymous-edgegaze.pages.dev/

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
gaze target detection, gaze following, auxiliary supervision

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_XbqFOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes EdgeGaze, an end-to-end multi-person gaze target detector built on the EdgeCrafter set-based detector. Each of N = 20 queries predicts a head box, a confidence score and a gaze distribution. Two additions are proposed: Scene Prior Guidance, where saliency, object and head maps supervise the shared encoder features during training to form a "Prior-Guided Scene Representation" (PGSR) fused with the person queries, and a Unified Gaze Distribution (UGD), a single softmax over the 64×64 gaze grid plus one out-of-frame bin. The paper reports good results on GazeFollow and VideoAttentionTarget, zero-shot transfer to GOO-Real and ChildPlay, and 3.1–4.9 ms latency with less parameters than Gaze-LLE.

Critical strengths
  • Efficiency is re-measured under a single protocol for all methods, including frozen backbones and the head detector required by two-stage methods.
  • The paper is transparent and honest about annotation incompleteness, wrong experiment estting and reports the strict detection mAP.
  • A compact end-to-end model is a useful practical direction, and the dataset statistic in Fig. 2 is a nice motivation.
Critical weaknesses
  • The evaluation protocol uses test-set information, and the reported numbers are not valid test results. This avoid the normal usage of test set in machine learning. All previous methods like Gaze-LLE would never do that.

    • Checkpoints are selected on the test split: "For each training run, we select the checkpoint with the highest mAP on the corresponding test split and report all metrics from that checkpoint" (L830-831). This applies to the main results and to every ablation.
    • The effect is visible in the paper itself. Table A5 evaluates the full model "at the final epoch" and gives VAT mAP 0.8211, L2 0.0569, AUC 0.9763, whereas Table 1 (test-selected checkpoint) reports 0.824, 0.056, 0.977. This gap is as large as most of the component ablation effects (Tables 3-4).
    • Predictions are assigned to annotations with Hungarian matching whose evaluation cost includes the gaze term with weight 100 and the out-of-frame term (Table A1, L787), and this single assignment is shared across AUC, L2, out-of-frame AP and mAP (L824-827). If the gaze cost is computed against the annotated gaze point, as Table A1 implies, the evaluation selects among 20 queries the one whose gaze is closest to the ground truth, using a weight 20× larger than the box term. This is an oracle selection, not a detection result. It would also explain why the VAT Out. AP is 0.999-1.000 for all variants, far above two-stage methods (≤0.951).
    • It is not stated whether the confidence threshold (L753-754) is applied before matching, nor how annotated people with no matching detection are counted in AUC/L2. If missed or unmatched people are excluded, the hardest cases are removed from the evaluation.
  • The comparison between end-to-end and two-stage methods is not like-for-like, and the SOTA claims (L38-39, L102-103, L377, L482) do not hold.

    • Two-stage methods produce one prediction per GT head; end-to-end methods are scored after the matching above. The pattern in Table 1 reflects this: all end-to-end methods have lower AUC than recent two-stage methods but much lower L2 (e.g. Tonini et al. 0.922 AUC / 0.072 L2 vs. ViTGaze 0.949 / 0.105). These distances are not on the same scale.
    • Even with this protocol, EdgeGaze is not SOTA on GazeFollow: PAGE reports 0.966 AUC vs 0.960 (L283), and the L2 advantage depends on the protocol issue above.
    • L412-413 claims the zero-shot setting is "stricter" than two-stage methods that receive GT heads. With gaze-based matching, it is not. A 26.7M model trained only on GazeFollow halving the in-domain ChildPlay L2 of Gaze-LLE ViT-L (0.053 vs 0.101, Table 2) needs to be verified under a protocol where matching uses head boxes only.
    • A fair comparison is easy to run, since the head detector is already used for the latency measurements: evaluate two-stage methods on detected heads, and evaluate EdgeGaze with box-only matching, thresholded predictions, and missed heads counted as errors.
  • The proposed components do not explain the results.

    • Without any scene prior (Table 3a, first row), the model already obtains 0.9592 AUC / 0.0584 L2 / 0.7857 mAP, i.e. above every baseline in Table 1. The full Scene Prior Guidance adds +0.0008 AUC and −0.0021 L2.
    • On VAT (Table A5), the no-prior model has higher AUC than the full model (0.9764 vs 0.9763) and the same Out. AP (0.9997). On GazeFollow, Lsal+Lobj without Lhead has the best AUC (0.9601).
    • UGD vs separate prediction (Table 4) changes AUC by 0.0015 and Out. AP from 0.9982 to 0.9997, and worsens VAT L2 (0.0545 → 0.0561). Out. AP is saturated and cannot discriminate between the two.
    • All differences come from single runs with checkpoints selected on test mAP. Notably, the full model stands out mainly in mAP (the selection criterion), while the other metrics are nearly flat. Multiple seeds and variance are needed.
    • The gains therefore appear to come mainly from the pretrained EdgeCrafter detector (COCO + Objects365, App. A.1) and the evaluation protocol, not from PGSR or UGD.
  • The "edge" and "independent of crowding" claims are not supported.

    • All latencies are measured on a desktop RTX 2080 Ti with TensorRT FP16 (L369). No embedded device (e.g. Jetson, mobile NPU/CPU), memory footprint or power measurement is reported, despite the title and L45 "well suited for edge deployment".
    • Computation is "nearly independent of scene crowding" (L31-32, L74-75, L101) because the model has a fixed cap of N = 20 queries (L303). It cannot output more than 20 people, yet Fig. A2 plots it up to K = 32 heads. Performance as a function of the number of people (and beyond 20) should be reported.
  • Limited novelty. The set-prediction formulation is HGTTR applied to the EdgeCrafter detector. Saliency/object/head auxiliary supervision follows prior work on joint gaze-saliency modeling (Chong et al., 2018) and object-aware gaze detection (Tonini et al., 2023; GaTector). UGD is a softmax with one additional "out" bin. Combined with the small ablation effects, the technical contribution is thin for top conferences (ICLR/NeurIPS/CVPR) but it's fine with flagship conferences or application centric conferences like ECCV, WACV, IEEE FG or BMVC.

Some minor issues:

  • Several numbers and statements are inconsistent across the main text and the appendix:

    • Table 4 (UGD row) and Table A5 (all three priors) describe the same full EdgeGaze-M configuration, yet report different VAT results: AUC 0.9771 vs 0.9763, L2 0.0561 vs 0.0569, mAP 0.8240 vs 0.8211. The paper should state which number is the actual result (see the checkpoint-selection comment above).
    • L317 "10 epochs of fine-tuning on VideoAttentionTarget" → Table A1 reports that only 20% of the training batches are used per epoch, i.e. about 2 full epochs.
    • L742 "EdgeGaze achieves the reported GazeFollow performance with just 10 training epochs" → per L830-831, the reported checkpoint is the one with the best test mAP, which is not necessarily the epoch-10 checkpoint.
  • Incorrect or inconsistent citations:

    • L364 states that Gaze-LLE uses DINOv3. Gaze-LLE uses a frozen DINOv2 encoder, while PAGE and GazeAnywhere are based on DINOv3. Please check which model was used for the parameter and latency measurements of Gaze-LLE.
    • TextGaze's parameter count is said to include "Qwen2.5-VL (Bai et al., 2025)" (L326), but Bai et al. (2025) is the Qwen3-VL technical report (L542-544).
    • Zong et al. 2023a and 2023b are the same paper (L661-666).
  • Missing comparisons: GazeDETR (L154) and GaTector+ (L190) are end-to-end / head-free methods discussed in the related work, and both appear in Fig. A1, but neither is compared in Tables 1-2. Head detection quality (AP or recall on annotated heads) is also never reported, although it is a core output of an end-to-end detector.

  • Several analyses and figures are presented selectively:

    • Fig. 1 plots Pareto fronts using Avg. Dist., the metric most affected by the matching protocol, and omits PAGE, which is in Table 1. With AUC as the accuracy axis and PAGE included, the margin over Gaze-LLE/GazeMoE shrinks to ≤0.002 AUC and PAGE is more accurate (0.966), which does not support "a substantially better accuracy–efficiency trade-off" (L109-110).
    • App. F.1 (L1024-1025) reports that all three priors give the best mAP and L2 after fine-tuning, but omits that the same model ranks 5th of 8 in VAT AUC (0.9763 vs 0.9773, Table A5).
    • Fig. A3 is selected by construction: the six cases are those where "only the full model places its gaze estimate within 0.06 of the annotation" (L1029-1030). It cannot serve as evidence for PGSR.
    • UGD is claimed to contribute to zero-shot transfer (L124-127), but it is not ablated on GOO-Real or ChildPlay, and no out-of-frame metric is reported on ChildPlay even though it provides out-of-frame labels.
  • The GazeHTA-style mAP excludes unmatched predictions, so it has no false positives and is closer to a matched-instance accuracy than an AP. The strict mAP (0.474, L910) should be reported in Table 1, with the same protocol applied to the baselines.

  • GOO-Real: the evaluated copy has ten images fewer than the standard split (L808-809), and AUC uses a 5×5 one-hot grid (L817). Please confirm the baseline numbers in Table 2 were obtained under the same definition.

  • Fig. 2 / App. C.2: L86 states that gaze falls "on another head", but C.2 (L851-852) counts gaze points "inside the head boxes". Please clarify whether the gazer's own head box is excluded.

  • The authors should tone down the language:

    • L109-110 "consistently lies on the Pareto frontier, offering a substantially better accuracy–efficiency trade-off" → see the comment on Fig. 1 above.
    • L415-416 "the learned scene and gaze representations generalize beyond the training distribution rather than overfitting to a particular benchmark" → two zero-shot datasets evaluated under a different protocol from the baselines do not support this.
    • L470-471 "provides a cleaner and more effective prediction space" → UGD changes GazeFollow AUC by 0.0015 and worsens VAT L2.
    • L485 "making end-to-end gaze understanding practical for edge deployment" → no edge device is evaluated.
    • L743 "indicating faster optimization" → the comparison is uncontrolled, as acknowledged in L744-746.
Questions for authors
  • The topic is interesting and important but the experiment design is not rigorous and has many fake claims. Each main argument above is followed by supporting evidence. Please address each main argument and the evidence behind it.
  • Please re-report Tables 1-4 with checkpoints selected on a held-out validation split (or the final epoch), with matching based on head boxes only, confidence thresholding before matching, and missed annotated people counted as errors. Please also report the number of matched vs. total annotated people.
  • Please evaluate at least Gaze-LLE and Sharingan on detected heads under the same end-to-end protocol, and report latency/memory on an actual edge device.
  • If there is no space in rebuttal windows, all minor issues can be skipped.
17 comments
AuthorsOpenReview ↗

# Response (1/14) to Reviewer XbqF

Thank you for your feedback. Due to the limited time remaining and the numerous questions, these are our final responses for the discussion period.

Question: The evaluation protocol uses test-set information, and the reported numbers are not valid test results. This avoid the normal usage of test set in machine learning. All previous methods like Gaze-LLE would never do that.

  • Checkpoints are selected on the test split: "For each training run, we select the checkpoint with the highest mAP on the corresponding test split and report all metrics from that checkpoint" (L830-831). This applies to the main results and to every ablation.
  • The effect is visible in the paper itself. Table A5 evaluates the full model "at the final epoch" and gives VAT mAP 0.8211, L2 0.0569, AUC 0.9763, whereas Table 1 (test-selected checkpoint) reports 0.824, 0.056, 0.977. This gap is as large as most of the component ablation effects (Tables 3-4).
  • Predictions are assigned to annotations with Hungarian matching whose evaluation cost includes the gaze term with weight 100 and the out-of-frame term (Table A1, L787), and this single assignment is shared across AUC, L2, out-of-frame AP and mAP (L824-827). If the gaze cost is computed against the annotated gaze point, as Table A1 implies, the evaluation selects among 20 queries the one whose gaze is closest to the ground truth, using a weight 20× larger than the box term. This is an oracle selection, not a detection result. It would also explain why the VAT Out. AP is 0.999-1.000 for all variants, far above two-stage methods (≤0.951).
  • It is not stated whether the confidence threshold (L753-754) is applied before matching, nor how annotated people with no matching detection are counted in AUC/L2. If missed or unmatched people are excluded, the hardest cases are removed from the evaluation.

Response: Checkpoint selection. Gaze-LLE's official GazeFollow script loads the test split (lines 44–47) and tracks the best epoch by Min. L2 (lines 107–111) [1]. PAGE likewise loads the GazeFollow test split (lines 94–102) and tracks the best Min. L2 epoch (lines 231–233) [2]. Object-aware GTD uses the GazeFollow test annotations for validation (lines 67–73) and saves checkpoints by validation AUC (lines 8–13) [3].

We could not locate a public implementation released by the original HGTTR authors. The public third-party reproduction (README line 3) uses the GazeFollow and VAT test annotations for validation and selects checkpoints by that split's AUC.

Our submission selected checkpoints by test mAP. We retrained EdgeGaze-M with held-out validation splits and selected checkpoints on validation. Both settings use the original scoring protocol on the test splits. The comparison also reflects the change in training data.

Dataset / checkpoint selectionAUC ↑mAP ↑Avg. L2 ↓Min. L2 ↓Out. AP ↑
GazeFollow / test (submitted)0.96000.79820.05630.0271—
GazeFollow / validation0.95920.78940.05710.0276—
VAT / test (submitted)0.97710.82400.0561—0.9997
VAT / validation0.97560.80980.0598—0.9998
AuthorsOpenReview ↗

# Response (1/14) to Reviewer XbqF

Hungarian matching. GazeFollow and VAT provide gaze annotations for only a subset of the people in a scene (Appendices B.2 and D.1). To score the annotated head–gaze pairs, we use one Hungarian assignment per image between the queries and the available annotations. The assignment minimizes the joint cost in Table A1, which uses the annotated gaze and in/out labels. Under the matched-instance mAP protocol, correct predictions for unannotated people do not incur false-positive penalties from missing labels. GazeHTA [4] likewise calls its matcher in the evaluator (lines 95–113) and solves a Hungarian assignment (lines 156–161) during test scoring (lines 119–130). The unofficial HGTTR implementation also matches predictions to test annotations for gaze AUC and L2, with a ground-truth gaze heatmap cost.

Confidence threshold and AUC/L2. We agree that Appendix A.2 should clarify where the confidence threshold is applied. It filters inference outputs. For Table 1, Hungarian matching uses all 20 final decoder predictions and assigns one to each eligible ground-truth instance in the evaluated GazeFollow and VAT splits. A valid in-frame annotation contributes to AUC and L2 even if its assigned head prediction has low confidence or poor localization. We report head recall under head-only matching in Response (2/14).

[1] Ryan, F., et al. "Gaze-LLE: Gaze Target Estimation via Large-Scale Learned Encoders." CVPR, 2025. Paper.

[2] Ye, Z., et al. "PAGE: Towards Practical Human-level Gaze Target Estimation." arXiv:2607.04860, 2026.

[3] Tonini, F., et al. "Object-aware Gaze Target Detection." ICCV, 2023.

[4] Lin, Z.-Y., et al. "GazeHTA: End-to-end Gaze Target Detection with Head-Target Association." ICRA, 2025.

AuthorsOpenReview ↗

# Response (2/14) to Reviewer XbqF

Question: The comparison between end-to-end and two-stage methods is not like-for-like, and the SOTA claims (L38-39, L102-103, L377, L482) do not hold.

  • Two-stage methods produce one prediction per GT head; end-to-end methods are scored after the matching above. The pattern in Table 1 reflects this: all end-to-end methods have lower AUC than recent two-stage methods but much lower L2 (e.g. Tonini et al. 0.922 AUC / 0.072 L2 vs. ViTGaze 0.949 / 0.105). These distances are not on the same scale.
  • Even with this protocol, EdgeGaze is not SOTA on GazeFollow: PAGE reports 0.966 AUC vs 0.960 (L283), and the L2 advantage depends on the protocol issue above.
  • L412-413 claims the zero-shot setting is "stricter" than two-stage methods that receive GT heads. With gaze-based matching, it is not. A 26.7M model trained only on GazeFollow halving the in-domain ChildPlay L2 of Gaze-LLE ViT-L (0.053 vs 0.101, Table 2) needs to be verified under a protocol where matching uses head boxes only.
  • A fair comparison is easy to run, since the head detector is already used for the latency measurements: evaluate two-stage methods on detected heads, and evaluate EdgeGaze with box-only matching, thresholded predictions, and missed heads counted as errors.

Response: Comparison protocol. We agree that Table 1 compares different evaluation settings. Two-stage gaze predictors are conditioned on supplied head boxes. EdgeGaze must localize heads and is scored after ground-truth-assisted Hungarian matching (Appendix B.2). Localization and association introduce additional sources of error, and gaze-assisted assignment affects AUC and L2. We report both settings side by side, following HGTTR (Table 1) [1] and GazeHTA (Table I) [2].

GazeFollow comparison. We supplied the released Gaze-LLE model with head boxes predicted by a dedicated YOLOv5m detector, which achieved 99.10% head recall. Both methods use the same confidence floor and head-only matching at IoU ≥ 0.5. AUC/L2 use matched heads. The mAP denominator includes all annotated heads, so misses count as failures.

MethodMatched/GT headsHead recall ↑AUC ↑mAP (head-only) ↑Avg. L2 ↓Min. L2 ↓
EdgeGaze-M (validation-selected)4695/478298.18%0.93200.70220.12980.0680
YOLOv5m + Gaze-LLE [3]4739/478299.10%0.95840.78360.09880.0413

Table 1 and PAGE. The submitted RTX 2080 Ti timings include the external head detector for two-stage methods (Section 4). Table 1 marks PAGE's 0.966 GF AUC as the best result. PAGE has 1,718.5 M parameters and takes 107.2 ms on that GPU, compared with 26.7 M and 4.9 ms for EdgeGaze-M.

ChildPlay zero-shot evaluation. In Table 2, ground-truth head boxes give the two-stage baselines a localization and association advantage. EdgeGaze must detect the heads, and its gaze scores can reflect errors in both detection and association. We computed the submitted L2 of 0.053 after ground-truth-assisted Hungarian matching, following GazeHTA's gaze-assisted association convention (matcher lines 129–134) [2]. We also evaluated the GazeFollow-validation-selected model with confidence filtering and head-only association.

ModelMatched/GT headsHead recall ↑AUC ↑L2 ↓
EdgeGaze-M (zero-shot)24322/2493597.54%0.94010.1160

Head recall covers all annotated heads in the evaluated frames. AUC/L2 use matched in-frame annotations. The common-protocol Gaze-LLE comparison above is on GazeFollow.

[1] Tu, D., et al. "End-to-End Human-Gaze-Target Detection With Transformers." CVPR, 2022. Section 4.3 and Table 1.

[2] Lin, Z.-Y., et al. "GazeHTA: End-to-end Gaze Target Detection with Head-Target Association." ICRA, 2025. Section IV-E and Table I.

[3] Ryan, F., et al. "Gaze-LLE: Gaze Target Estimation via Large-Scale Learned Encoders." CVPR, 2025. Paper.

Reviewer_CPr5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

EdgeGaze adapts the compact set-based EdgeCrafter detector to predict head boxes and gaze targets for multiple people in one forward pass. Its Scene Prior Guidance uses offline saliency, object, and head targets to supervise a shared scene representation, which is combined with person-specific query features. A Unified Gaze Distribution jointly normalizes spatial gaze locations and an out-of-frame outcome. Three model sizes are evaluated on GazeFollow and VideoAttentionTarget, with transfer to GOO-Real and ChildPlay. The paper reports 0.960 AUC and 0.798 mAP on GazeFollow for the 26.7M-parameter M variant, with 4.9 ms TensorRT FP16 engine latency on an RTX 2080 Ti.

Critical strengths
  1. Predicting heads and gaze jointly with a fixed query budget, while using external experts only for training, directly addresses the inference cost of multi-person gaze estimation. The paper measures inference parameters and computation across methods under an explicit accounting convention, including required external components. Even though the measurements are desktop-GPU engine timings rather than complete edge-device measurements, they provide useful evidence for the efficiency of the architecture.

  2. The paper evaluates multiple model sizes, two training benchmarks, and two transfer datasets, and provides all eight combinations of the scene-prior losses. It also explicitly documents incomplete head annotations and shows how alternative treatments change mAP. These are useful diagnostics, although the evaluation choices discussed below prevent me from accepting the headline accuracy conclusions as currently established.

Critical weaknesses
  1. Checkpoint selection uses test labels. Appendix B.2, p. 16, states that each training run selects the checkpoint with the highest mAP on the corresponding test split and reports all metrics from that checkpoint. This makes the reported test scores part of model selection rather than an independent generalization assessment. It also affects comparisons among ablations, especially where improvements are small. The final-epoch VAT results in Table A5 are useful additional evidence, but do not resolve the selection policy for the main tables or the selection of their source checkpoints. Validation-selected or prespecified-checkpoint results are needed before the accuracy claims can be assessed.

  2. Target-dependent matching complicates both person-specific accuracy and cross-method comparability. Table A1 specifies a gaze coefficient of 100 for evaluation matching, and Appendix B.2 says that a single Hungarian assignment is shared across AUC, distance, out-of-frame AP, and mAP, with unmatched predictions excluded from the latter. Matching candidate people using their ground-truth gaze can preferentially select a prediction with a favorable gaze answer, rather than first establishing person identity from head geometry. Excluding all unmatched predictions also removes the usual penalty for duplicate or otherwise extraneous predictions. Partial annotation is a real problem, and the paper is transparent about this protocol, but it does not establish equivalence to conventional head-conditioned gaze metrics or detection AP. This is particularly consequential for the comparison to two-stage methods supplied with a fixed ground-truth head.

Questions for authors
  1. Can you provide the main results and key ablations using a checkpoint selected solely on a validation split, or a training schedule fixed without consulting test performance? Please also confirm that neither GOO-Real nor ChildPlay labels were used to select the source checkpoint for the zero-shot results. Do the accuracy and ablation conclusions persist?

  2. Can you associate predictions with annotated people using only head identity or geometry, without ground-truth gaze or in/out labels in the assignment, and then compute gaze metrics? Please report head recall and account for duplicates and misses, using documented ignore regions for genuinely unannotated people rather than dropping every unmatched prediction. How do EdgeGaze and the principal baselines compare under this common protocol?

2 comments
AuthorsOpenReview ↗

# Response (1/2) to Reviewer CPr5

Thank you for your feedback. Due to the limited time remaining and the numerous questions, these are our final responses for the discussion period.

Question:

Checkpoint selection uses test labels. Appendix B.2, p. 16, states that each training run selects the checkpoint with the highest mAP on the corresponding test split and reports all metrics from that checkpoint. This makes the reported test scores part of model selection rather than an independent generalization assessment. It also affects comparisons among ablations, especially where improvements are small. The final-epoch VAT results in Table A5 are useful additional evidence, but do not resolve the selection policy for the main tables or the selection of their source checkpoints. Validation-selected or prespecified-checkpoint results are needed before the accuracy claims can be assessed.

Can you provide the main results and key ablations using a checkpoint selected solely on a validation split, or a training schedule fixed without consulting test performance? Please also confirm that neither GOO-Real nor ChildPlay labels were used to select the source checkpoint for the zero-shot results. Do the accuracy and ablation conclusions persist?

Response: We selected the submitted checkpoints by test mAP (Appendix B.2). We retrained with validation holdouts from GazeHTA's repository lists [7]. These cover 3,997 GF images (the paper reports 2,697 samples) and six VAT shows (Driving Miss Daisy, Breaking Bad, My Dinner with Andre, Star Wars, The View, and Veep). GF [1] and VAT [2] lack official validation splits. We selected EMA checkpoints by validation hit rate, requiring head IoU ≥ 0.5 and either in-frame L2 < 0.15 or a correct out-of-frame prediction. VAT started from the selected GF checkpoint.

Test scoring follows the original protocol. The comparison also reflects the changed training splits. Δ = validation − test before rounding.

Dataset/selectionAUC ↑mAP ↑Avg. L2 ↓Min. L2 ↓Out. AP ↑
GF, test (submitted)0.96000.79820.05630.0271—
GF, validation0.95920.78940.05710.0276—
GF, Δ-0.0008-0.0088+0.0009+0.0005—
VAT, test (submitted)0.97710.82400.0561—0.9997
VAT, validation0.97560.80980.0598—0.9998
VAT, Δ-0.0015-0.0141+0.0037—+0.0001

We selected both zero-shot source checkpoints using GF only, with test selection in the submission and validation selection here.

GF source selectionGOO AUC ↑GOO L2 ↓ChildPlay AUC ↑ChildPlay L2 ↓
Test (submitted)0.94660.08530.96510.0532
Validation0.94360.09390.96480.0543
Δ-0.0030+0.0087-0.0003+0.0010

On GF, Gaze-LLE [3] and PAGE [4] track the best test Min. L2 (Gaze-LLE split and best epoch, PAGE split and best epoch). Object-aware GTD [5] and a public third-party reproduction of HGTTR [6] select by test AUC (Object-aware GTD split and checkpoint rule, HGTTR reproduction split and checkpoint rule).

[1] Recasens, A., et al. "Where Are They Looking?" NeurIPS, 2015.

[2] Chong, E., et al. "Detecting Attended Visual Targets in Video." CVPR, 2020.

[3] Ryan, F., et al. "Gaze-LLE: Gaze Target Estimation via Large-Scale Learned Encoders." CVPR, 2025.

[4] Ye, Z., et al. "PAGE: Towards Practical Human-level Gaze Target Estimation." arXiv:2607.04860, 2026.

[5] Tonini, F., et al. "Object-aware Gaze Target Detection." ICCV, 2023.

[6] Tu, D., et al. "End-to-End Human-Gaze-Target Detection With Transformers." CVPR, 2022.

[7] Lin, Z.-Y., et al. "GazeHTA: End-to-end Gaze Target Detection with Head-Target Association." ICRA, 2025. Official GF validation list.

AuthorsOpenReview ↗

# Response (2/2) to Reviewer CPr5

Question:

Target-dependent matching complicates both person-specific accuracy and cross-method comparability. Table A1 specifies a gaze coefficient of 100 for evaluation matching, and Appendix B.2 says that a single Hungarian assignment is shared across AUC, distance, out-of-frame AP, and mAP, with unmatched predictions excluded from the latter. Matching candidate people using their ground-truth gaze can preferentially select a prediction with a favorable gaze answer, rather than first establishing person identity from head geometry. Excluding all unmatched predictions also removes the usual penalty for duplicate or otherwise extraneous predictions. Partial annotation is a real problem, and the paper is transparent about this protocol, but it does not establish equivalence to conventional head-conditioned gaze metrics or detection AP. This is particularly consequential for the comparison to two-stage methods supplied with a fixed ground-truth head.

Can you associate predictions with annotated people using only head identity or geometry, without ground-truth gaze or in/out labels in the assignment, and then compute gaze metrics? Please report head recall and account for duplicates and misses, using documented ignore regions for genuinely unannotated people rather than dropping every unmatched prediction. How do EdgeGaze and the principal baselines compare under this common protocol?

Response: We evaluated EdgeGaze-M and YOLOv5m + Gaze-LLE on the same GazeFollow images and annotations, using released Gaze-LLE weights [1]. We applied a head-score floor of 0.001 before one-to-one matching in score order at head IoU ≥ 0.5. AUC/L2 use matched heads. The reported mAP is the head–gaze hit rate over all annotated people, with misses counted as failures.

MethodMatched/GT headsHead recall ↑AUC ↑mAP (head-only) ↑Avg. L2 ↓Min. L2 ↓
EdgeGaze-M (validation-selected)4,695/4,78298.18%0.93200.70220.12980.0680
YOLOv5m + Gaze-LLE4,739/4,78299.10%0.95840.78360.09880.0413

EdgeGaze-M had 87 missed heads, 28,077 duplicate boxes and 62,868 extra boxes. YOLOv5m + Gaze-LLE had 43 misses, 221 duplicates and 150,016 extra boxes. We counted duplicates and extras separately from the hit rate. Extra boxes may include unannotated people. We retained them in the counts because the 8,842 detector-generated ignore regions lack manual verification.

Our submitted mAP excludes unmatched predictions, following GazeHTA's matched-instance scoring (evaluator lines 36–113) [2]. Our run of the public third-party HGTTR reproduction [3] yielded mAP 0.3403 on GF and 0.2931 on VAT. The original paper reports 0.547 and 0.514 with ResNet-50.

[1] Ryan, F., et al. "Gaze-LLE: Gaze Target Estimation via Large-Scale Learned Encoders." CVPR, 2025. Paper.

[2] Lin, Z.-Y., et al. "GazeHTA: End-to-end Gaze Target Detection with Head-Target Association." ICRA, 2025.

[3] Tu, D., et al. "End-to-End Human-Gaze-Target Detection With Transformers." CVPR, 2022. Paper, Table 1; unofficial reproduction.

Reviewer_s47UOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents EdgeGaze, a compact end-to-end framework for multi-person gaze target detection built on EdgeCrafter. It jointly detects heads and predicts their gaze targets in a single forward pass. The method introduces Scene Prior Guidance, which uses saliency, object, and head supervision during training to learn a prior-guided scene representation, and a Unified Gaze Distribution (UGD), which jointly normalizes in-frame gaze locations and an out-of-frame outcome. The model achieves a strong accuracy-efficiency trade-off on standard benchmarks and transfers well to unseen datasets, with its main variant reaching 0.960 AUC and 0.798 mAP on GazeFollow using 26.7M parameters and 4.9 ms inference latency.

Critical strengths
  1. EdgeGaze achieves state-of-the-art results on GazeFollow and other benchmarks.
  2. Saliency, object, and head priors provide intuitive complementary supervision.
  3. The computational cost is relatively small.
Critical weaknesses
  1. The methodological contribution is relatively incremental. The efficient architecture is largely inherited from EdgeCrafter, while scene-prior supervision follows a familiar auxiliary-supervision strategy. UGD is simple and useful, but its conceptual novelty is limited.
  2. The ablation study seems not obvious. The ablation results do not clearly establish the contribution of each proposed component. Table 3 (b) lacks the results of removing all components.
  3. The ablation study cannot answer why the proposed method achieves so good results. It seems that the improvement of all proposed components is small. This raise the question: which component makes the results good.
  4. The title of this paper is “multi-person gaze target detection”. However, the reviewer can not see the specific motivation or dataset contribution of this point.
Questions for authors
  1. Which component makes the results good?
  2. Whether the latency of all models is tested on the same device with same input?
  3. What is the role of multi-scale memory?
3 comments
AuthorsOpenReview ↗

# Response (1/3) to Reviewer s47U

Thank you for your feedback. Due to the limited time remaining and the numerous questions, these are our final responses for the discussion period.

Question: The methodological contribution is relatively incremental. The efficient architecture is largely inherited from EdgeCrafter, while scene-prior supervision follows a familiar auxiliary-supervision strategy. UGD is simple and useful, but its conceptual novelty is limited.

Response: We agree that EdgeGaze is assembled from simple parts. The efficient set-prediction backbone and encoder are inherited from EdgeCrafter, as the paper states. Scene Prior Guidance uses saliency, object, and head targets only as training losses, so no external predictor runs at inference. The Unified Gaze Distribution normalizes the in-frame locations and the out-of-frame outcome with a single softmax. EdgeGaze-M predicts heads and gaze targets in one forward pass with 26.7 M parameters. It takes 4.9 ms on an RTX 2080 Ti (Table 1) and 9.8 ms on Jetson AGX Orin (Response (3/3)). The cost of two-stage methods grows with the number of detected heads (Section 4).

Question: The ablation study seems not obvious. The ablation results do not clearly establish the contribution of each proposed component. Table 3 (b) lacks the results of removing all components.

Response: As suggested, we trained a query-only baseline without Scene Prior Guidance, scene/interaction fusion, or UGD. It predicts gaze with a spatial heatmap head and a separate in/out classifier. The detector, multi-scale encoder and memory, and query initialization are retained.

For this new comparison, both variants use the same GazeFollow training split, detector initialization, and ten-epoch schedule. We evaluate the predeclared final-epoch EMA checkpoints using all four Table 3 metrics.

SettingAUC ↑mAP ↑Avg. L2 ↓Min. L2 ↓
Query-only (separate heads)0.95090.79130.05100.0411
EdgeGaze-M (complete)0.95940.77940.05800.0287

The complete model has higher AUC and lower Min. L2; the query-only baseline has higher mAP and lower Avg. L2. The results show a metric-dependent trade-off in this comparison. Since this baseline changes the gaze-specific design as a whole, it does not isolate individual component effects. We will include this comparison alongside Table 3(b) and make this scope explicit.

AuthorsOpenReview ↗

# Response (2/3) to Reviewer s47U

Question:

The ablation study cannot answer why the proposed method achieves so good results. It seems that the improvement of all proposed components is small. This raise the question: which component makes the results good.

Which component makes the results good?

Response: With the same EdgeCrafter initialization, enabling the three prior losses in Table 3(a) raises GazeFollow mAP from 0.7857 to 0.7982 (+1.59%) and lowers Min. L2 from 0.0286 to 0.0271 (−5.24%). In the fixed-epoch GazeFollow comparison, AUC is higher with prior losses and mAP is higher without them. UGD combines spatial gaze and the out-of-frame outcome in a single distribution and raises VAT mAP from 0.8125 to 0.8240, with L2 increasing from 0.0545 to 0.0561 (Table 4).

Question: The title of this paper is “multi-person gaze target detection”. However, the reviewer can not see the specific motivation or dataset contribution of this point.

Response: In classrooms and child–caregiver interactions, gaze predictions for several people are needed from the same image. A two-stage pipeline runs a head detector and then a per-person gaze computation (Section 4). At a fixed query budget N, EdgeGaze encodes the image once and decodes up to N head–gaze candidates in parallel. With the same checkpoint and the submitted scoring protocol, increasing N from 20 to 50 lowers GazeFollow Avg. L2 from 0.05628 to 0.04847.

We augment existing training images with offline saliency, object, and head targets derived from expert predictors and available head annotations (Appendix C). Inference uses only the input image.

AuthorsOpenReview ↗

# Response (3/3) to Reviewer s47U

Question: Whether the latency of all models is tested on the same device with same input?

Response: For the submitted benchmark, all methods use the same RTX 2080 Ti and the same first 100 GazeFollow test images, resized to each method's native resolution. We time TensorRT FP16 engine execution at batch size one after 50 warm-up iterations (Section 4 and Appendix E.1). Two-stage methods process all 599 detected heads.

We also timed seven models on the same NVIDIA Jetson AGX Orin using TensorRT 10.4 in FP16 at batch size 1. EdgeGaze-N, -S and -M take 5.2, 7.5 and 9.8 ms, respectively. Tonini et al., HGTTR, GazeHTA and Sharingan take 11.1, 12.8, 111.7 and 22.5 ms, respectively.

Question: What is the role of multi-scale memory?

Response: The multi-scale memory is inherited from EdgeCrafter, where features at different scales support detecting objects of different sizes. It contains image features at strides 8, 16, and 32 that the decoder reads through cross-attention. EdgeGaze's detection task focuses on human heads. Our ablation shows that reading all three scales is unnecessary to achieve the gaze accuracy reported below.

We re-trained EdgeGaze-M on GazeFollow with the decoder reading either all three scales or just one, using the same training recipe and final-epoch checkpoint rule. The multi-scale encoder and query initialization were kept unchanged. The metrics follow Table 3.

Decoder memoryAUC ↑mAP ↑Avg. L2 ↓Min. L2 ↓
Strides 8 + 16 + 32 (default)0.95940.77940.05800.0287
Stride 8 only0.95860.78020.05850.0288
Stride 16 only0.95950.78840.05790.0275
Stride 32 only0.95870.78150.05870.0285

Stride 16 alone is slightly better on all four gaze metrics. These results show that a single decoder memory scale is sufficient in this setting. Multi-scale features remain in the encoder. We will clarify this point and add the table to the appendix.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.