{"id":"05bd6014-9f01-4f6f-be63-c5acbe89d3bd","arxiv_id":"2505.16384","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"MAGE jointly predicts gaze direction and a device-independent point-of-gaze (PoGz) from a single RGB image, and uses a screen-free look-at-the-camera calibration module to fine-tune per-subject performance.","lead":"This paper proposes MAGE, a multi-task neural network that estimates both gaze direction and the point of gaze in 3D space from one RGB camera image. The method includes a screen-free personal calibration step, where users simply look at the camera lens, which could make gaze tracking practical for human-robot interaction.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 8's screen-free calibration labels use the face-bbox center as the gaze origin and omit kappa, injecting a ~3-8 degree systematic bias that exceeds the claimed 0.23 degree calibration gain; the IMRGaze validation cannot establish the calibration benefit.","rationale":"The paper's primary novelty is Easy-Calibration, and the validity of that module depends entirely on the correctness of the labels generated by Eq. 8. The reader's weakest_assumption identified exactly this issue, and I agree that it is the most load-bearing concern. A wrong calibration label definition would not necessarily invalidate the public-dataset SOTA claims, since those comparisons use standard gaze labels, but it would invalidate the IMRGaze-based demonstration that calibration improves accuracy. The absence of a statement about how IMRGaze ground-truth gaze is defined relative to the face-center proxy makes the reported 0.23 degree improvement impossible to interpret. Because the concern is concrete and testable, but not yet demonstrated to be fatal, the reader's CONDITIONAL verdict should stand: the paper should be accepted only if the authors can show that Eq. 8 labels are not biased relative to the evaluation protocol, or if the IMRGaze results are recomputed with corrected eye-gaze labels.","tokens_in":10416,"tokens_out":7159,"duration_ms":66969,"concrete_test":"On the IMRGaze calibration videos, or on any RGB-D recording with eye landmarks and camera intrinsics, compute for each frame the angular difference between (a) the Eq. 8 label from the face-bbox center and (b) the true label from the 3D midpoint of the eyes to the camera origin. If the median difference exceeds 1 degree, Eq. 8 is a biased gaze label; then rerun the calibration fine-tuning with labels (b) and compare test error. A material change in test accuracy would confirm that the reported calibration gain is an artifact of the face-center proxy and kappa omission.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section II-E, Eq. 8 defines the calibration ground truth as go = -[x_cs, y_cs, fs]^T, i.e., the vector from the face-bbox center projection to the camera origin. This is not the eye-gaze vector: the visual axis originates at the eye centers and includes the kappa angle, while the face-bbox center is offset by several centimeters in 3D. At a typical 60 cm working distance, a 3 cm lateral offset already gives about 2.9 degrees of angular error, and kappa adds roughly 2-5 degrees. These biases are larger than the reported calibration improvement (5.38 to 5.15 degrees, a 0.23 degree gain) and are comparable to the absolute error of 5.15 degrees. If the IMRGaze general-data labels are computed using the same face-center proxy, then the improvement is circular; if the general-data labels are true eye-gaze labels, fine-tuning with Eq. 8 should push predictions away from the evaluation labels. The paper does not state which origin definition is used for the IMRGaze ground truth, so the central claim for Easy-Calibration is not supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MAGE, a multi-task architecture that takes a single RGB image and a facial bounding box and predicts a 6-DoF gaze representation composed of gaze direction (gn and go) and a gaze point PoGz on the camera XY-plane. It introduces Easy-Norm, a normalization procedure that uses only the bounding box, and Easy-Calibration, a screen-free personalization module whose calibration labels are generated by having the subject look at the camera lens. Experiments on MPIIFaceGaze, EYEDIAP, and a new IMRGaze dataset report state-of-the-art or improved results, including a 0.23° calibration gain on IMRGaze.","tokens_in":10697,"tokens_out":8515,"duration_ms":69136,"significance":"If the claims held, MAGE would offer a practical contribution to HRI by providing both direction and position of gaze from a single RGB image with a lightweight calibration procedure, and the IMRGaze dataset could be a useful benchmark. The multi-task design is sensible, the Easy-Norm idea of avoiding head-pose/depth requirements is useful, and the ablation study gives a concrete account of the contribution of each branch. However, the paper does not ship code or a public release of IMRGaze, and the current evidence for the state-of-the-art and calibration claims is weakened by protocol inconsistencies and an unclear label-origin definition. The screen-free calibration idea is interesting but needs validation with explicitly consistent ground-truth definitions.","major_comments":[{"comment":"The calibration ground truth in Eq. (8) is the vector from the face-bounding-box center projection to the camera origin. The paper does not state whether the IMRGaze general-data ground-truth gaze vector is defined from the same face-center origin or from the eye centers. The Introduction defines gaze direction as originating from the ocular center, and public datasets such as MPIIFaceGaze and EYEDIAP use eye-based gaze vectors. If the IMRGaze evaluation labels are eye-based, then Eq. (8) labels carry a systematic bias from the face-eye offset and the kappa angle, which is likely larger than the reported 0.23° calibration improvement. If the evaluation labels are also face-center-based, then the reported calibration gain is for a nonstandard definition and cannot be compared with the public-dataset baselines. The authors must specify and justify the gaze-origin definition in both label sets and demonstrate that the calibration and evaluation labels are consistent.","section":"Section II-E (Eq. 8) and Section III-A"},{"comment":"The state-of-the-art claim relies on comparisons to numbers taken from other papers, including iTracker and AFF-Net quoted from the review [6]. There is no demonstration that the baselines were evaluated under the same normalization, input format (face vs. eyes), and cross-validation splits. The reported margins, e.g., 3.54° vs. 3.82° on MPIIFaceGaze, may be within protocol differences, and no error bars or statistical significance tests are reported. Please either re-evaluate all baselines under a unified protocol or restrict the claim to competitive performance with an explicit caveat that protocol details differ.","section":"Table II and Section IV-A"},{"comment":"The transformation of the gaze direction vector into the screen coordinate system is incorrect. In Eq. (3), g_s = R_s^c · g_o + t_s^c treats a direction vector as a point; a unit direction transforms as g_s = R_s^c · g_o without the translation term. If this equation is used for the PoG conversion that underlies the PoG errors in Table II, the reported PoG results are not trustworthy. This should be corrected and the PoG numbers recomputed.","section":"Section II-C (Eq. 3)"}],"minor_comments":[{"comment":"\"real-word HRI\" should be \"real-world HRI\".","section":"Abstract"},{"comment":"The gaze origin is defined as \"face center\" in the methodology, but the Introduction states that gaze direction originates from the ocular center. Please align the terminology and clarify which origin the model actually predicts.","section":"Section II"},{"comment":"The description of the IMRGaze general-data label computation is underspecified; please provide the exact formula used to convert RGB-D and MediaPipe outputs into the gaze-vector ground truth.","section":"Section III-A"},{"comment":"The notation involving O_PoGz and O_s_PoGz is confusing, and \"intersection point of the original gaze vector go\" should read \"intersection point of the gaze ray associated with go\".","section":"Figure 3 and Eq. (3)"},{"comment":"There is a missing space in \"λPoGz = 0and freezing\".","section":"Section III-D"},{"comment":"The paper does not state whether the IMRGaze dataset will be released; please clarify data and code availability for reproducibility.","section":"Section III-A"}],"recommendation":"major_revision","confidential_remarks":"The central risk is the unstated gaze-origin definition in the IMRGaze calibration and evaluation labels. I would ask the editor to require a precise statement of how the general-data ground-truth gaze vectors were computed and how they relate to Eq. (8). The Eq. (3) error is easy to fix but should be verified before the PoG numbers are used. The paper is within the journal's scope, but the SOTA claim needs a more controlled comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: MAGE is a sensible multi-task gaze estimation architecture, and the screen-free calibration idea is genuinely practical. But the calibration labels in Eq. 8 are based on the face bounding-box center rather than the eye, which injects a systematic bias that is probably larger than the reported calibration improvement. Read the IMRGaze results with that in mind.\n\nWhat's actually new: the paper combines gaze direction and PoGz (the intersection of the gaze ray with the XY-plane of the camera coordinate system) in a single RGB-only model, and it introduces Easy-Norm, a normalization that needs only the face bounding box rather than head pose and depth. The Easy-Calibration protocol—look at the camera lens, move your head, fine-tune the model—is a nice idea for HRI. The architecture is straightforward: two ResNet-18 branches plus a box encoder and linear decoders. It is described clearly enough to reimplement.\n\nThe public-dataset results are plausible. The multi-task ablation on EYEDIAP-Float and IMRGaze shows that removing the PoGz or go branches degrades accuracy, which supports the design. On MPIIFaceGaze and EYEDIAP the numbers beat the listed baselines, though the baselines are not all re-run under identical normalization and data splits—iTracker and AFF-Net numbers are quoted from a review paper, and there are no error bars. That is a common weakness in this subfield, but it means the SOTA claim is not as strong as the table suggests.\n\nThe bigger issue is Easy-Calibration. Eq. 8 defines the gaze ground truth as the vector from the projected face-bbox center to the camera origin. That is not the visual axis: the eye centers are offset from the face center, and the kappa angle adds another few degrees. At a typical 60 cm working distance, these biases are on the order of 3–8 degrees, far larger than the 0.23° gain from calibration reported in Table III. The paper does not state whether the IMRGaze general-data evaluation labels also use the face-center proxy. If they do, the calibration improvement is circular; if they do not, fine-tuning with Eq. 8 labels should push predictions away from the true eye-gaze labels. Either way, the current evidence does not support the calibration claim. This is a correctable problem, but it needs to be fixed before the paper is accepted.\n\nWho this is for: researchers in gaze-based HRI or assistive systems who want a screen-free personal calibration and a unified direction-plus-position output. It deserves a serious referee, but with the expectation of revision.\n\nMy recommendation: send it to review, ask for the code and IMRGaze labels (or at least a precise description of the origin definition used in the general-data ground truth), and require a re-run of baselines under identical conditions. The core architecture is sound; the calibration evidence is not yet.","headline":"MAGE has a sound multi-task gaze architecture and a practical screen-free calibration idea, but the calibration labels rely on a face-center proxy that likely swamps the reported gain—read the IMRGaze results with caution.","tokens_in":11260,"tokens_out":3377,"would_cite":false,"duration_ms":27156,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MAGE claims that complete 6-DoF gaze—direction plus position—can be predicted from a single RGB image and face bounding box, and that a screen-free calibration from roughly 50 frames adapts it to each user.","keywords":["gaze estimation","6-DoF gaze","point of gaze","multi-task learning","person-specific calibration","Easy-Norm","human-robot interaction","RGB-only gaze tracking"],"falsifier":"Run a controlled comparison on the same subjects between Easy-Calibration and a standard screen-based calibration: if the screen-based method yields materially lower gaze error, the Eq. 8 face-center/lens-center assumption is biased. A more direct check is to simulate an eye model with known kappa angle and verify that Eq. 8's ground-truth vector coincides with the simulated visual axis.","tokens_in":10205,"feed_emoji":"👁️","tokens_out":7122,"duration_ms":50534,"temperature":0.7,"pith_summary":"MAGE is a gaze-estimation network that takes a single RGB image plus a face bounding box and predicts both the 3D direction of gaze and where the gaze ray crosses the camera's XY-plane, a quantity the paper calls PoGz. Together these give the full six-degree-of-freedom gaze (origin and direction) needed for human-robot interaction. To handle person-to-person differences in eye anatomy, the paper adds Easy-Calibration, a screen-free procedure in which a user simply looks at the camera lens while moving his or her head; about 50 frames per person are enough to fine-tune the network. On MPIIFaceGaze, EYEDIAP, and the authors' new IMRGaze dataset, the method reports lower gaze-direction and point-of-gaze errors than the specialized direction-only and PoG-only baselines.","feed_headline":"One RGB image yields full 6-DoF gaze with a screen-free calibration","feed_subtitle":"MAGE predicts direction plus position and adapts to a user from about 50 frames, beating specialized gaze models.","key_machinery":"The load-bearing mechanism is Easy-Norm plus Easy-Calibration, coordinated through multi-task decoders. Easy-Norm standardizes camera intrinsics and rotates the camera z-axis toward the face center using only the facial bounding box, so the network sees a normalized face without requiring head pose or depth. Easy-Calibration generates ground-truth labels without a screen: while the subject gazes at the camera lens and moves the head, Eq. 8 defines the gaze vector as pointing from the projected face-center to the lens center, and these frames fine-tune the model. PoGz—the intersection of the gaze ray with the XY-plane of the camera coordinate system—is the device-independent positional output that lets the same model be evaluated on different screens and merges directional and positional supervision.","core_discovery":"The central claim is that gaze direction and point of gaze should be estimated jointly, and that the joint task can be made accurate across people with a bounding-box-only normalization and a screen-free personalization step. MAGE's Easy-Norm rotates the camera coordinate system so its optical axis points at the face center, using only the face bounding box, and then feeds the normalized face through separate gaze, pose, and box encoders. Decoders jointly output the normalized gaze vector, the original gaze vector, the intersection of the gaze ray with the camera's XY-plane (PoGz), the normalization rotation, and the face center, all supervised by a multi-task loss. On MPIIFaceGaze the paper reports 3.54° direction error and 32.73 mm point-of-gaze error; on EYEDIAP-Screen, 4.64° and 73.07 mm; on EYEDIAP-Float, 5.11°; and on its own IMRGaze dataset, 5.15° after Easy-Calibration with about 100 calibration frames.","pith_inferences":["PoGz's device independence suggests the model could be trained on multiple camera geometries and deployed on a new camera with no retraining; the paper does not test this cross-device transfer explicitly.","The calibration protocol could be folded into normal human-robot interaction, since moments when a user deliberately looks at the robot's camera are exactly the frames Eq. 8 needs; this would allow continuous, passive personalization.","Because the network also predicts the normalization rotation and gaze origin, a self-supervised consistency loss linking go, PoGz, and the rotation could further tighten predictions; the paper lists this as future work.","A testable extension is to apply Easy-Calibration with a kappa-angle correction to isolate how much of the reported 0.23° improvement is true individual-geometry adaptation rather than generic fine-tuning."],"forward_implications":["A depth sensor is unnecessary for complete 6-DoF gaze: one RGB camera plus a face detector supplies both direction and position.","Because PoGz is device-independent, the same trained model can serve both screen-pointer tasks and free-space human-robot interaction without retraining for each screen.","Easy-Calibration makes per-user adaptation practical outside the lab, requiring about 50 frames and roughly 17 seconds of video, with no screen-based annotations.","Supervising the model with PoGz improves its gaze-direction accuracy: removing the PoGz branch raises IMRGaze direction error from 5.38° to 5.83°."],"supporting_citations":[{"why":"Provides the MPIIFaceGaze full-face dataset and evaluation protocol used for the reported 3.54° and 32.73 mm results.","marker":"[14]"},{"why":"Provides the EYEDIAP dataset, including the Float and Screen subsets used for benchmark comparisons.","marker":"[24]"},{"why":"Review paper from which the iTracker and AFF-Net PoG comparison numbers in Table II are taken.","marker":"[6]"},{"why":"iTracker baseline and GazeCapture-based PoG approach that MAGE compares against on point-of-gaze error.","marker":"[12]"},{"why":"AFF-Net baseline that MAGE outperforms on both MPIIFaceGaze and EYEDIAP-Screen point-of-gaze error.","marker":"[13]"},{"why":"GazeTR transformer baseline that MAGE compares against on gaze-direction error across all datasets.","marker":"[21]"},{"why":"AGE-Net baseline, the previous best direction-only method that MAGE surpasses on public benchmarks.","marker":"[22]"},{"why":"L2CS-Net baseline, representing fine-grained appearance-based gaze direction estimation.","marker":"[23]"},{"why":"RCNN baseline that combines face and eye appearance cues for gaze direction estimation.","marker":"[10]"},{"why":"Prior EasyGaze3D method that the Easy-Calibration screen-free calibration approach extends.","marker":"[20]"}],"fun_headline_variants":["6-DoF gaze from one image, no screen needed for calibration","MAGE: joint gaze direction and point-of-gaze with easy screen-free calibration","One RGB image gives full gaze vector and target point, with quick personalization","Predict gaze direction and point together, calibrate without a screen","Screen-free calibration adapts gaze model to you in ~100 frames"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The screen-free calibration labels rely on the assumption that when a person looks at the camera lens, the gaze ray is exactly the line from the face bounding box center to the lens center; a large or systematic kappa angle, or a face-center proxy that does not match the true gaze origin, would bias the fine-tuning.","fun_headline_variants_meta":{"raw":{"variants":["6-DoF gaze from one image, no screen needed for calibration","MAGE: joint gaze direction and point-of-gaze with easy screen-free calibration","One RGB image gives full gaze vector and target point, with quick personalization","Predict gaze direction and point together, calibrate without a screen","Screen-free calibration adapts gaze model to you in ~100 frames"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001167,"raw_usage":{"total_tokens":4850,"prompt_tokens":990,"completion_tokens":3860,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":3778}},"tokens_in":606,"tokens_out":3860,"duration_ms":22589,"temperature":1.0,"reasoning_tokens":3778,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:01:46.261254+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled comparison on the same subjects between Easy-Calibration and a standard screen-based calibration: if the screen-based method yields materially lower gaze error, the Eq. 8 face-center/lens-center assumption is biased. A more direct check is to simulate an eye model with known kappa angle and verify that Eq. 8's ground-truth vector coincides with the simulated visual axis.","supporting_citations":[{"cited_title":"L2cs-net: Fine-grained gaze estimation in unconstrained environments,","cited_arxiv_id":null,"evidence_quote":"L2CS-Net baseline, representing fine-grained appearance-based gaze direction estimation."},{"cited_title":"Eyediap: A database for the development and evaluation of gaze estimation algorithms from rgb and rgb-d cameras,","cited_arxiv_id":null,"evidence_quote":"Provides the EYEDIAP dataset, including the Float and Screen subsets used for benchmark comparisons."},{"cited_title":"Appearance-based gaze estimation with deep learning: A review and benchmark,","cited_arxiv_id":null,"evidence_quote":"Review paper from which the iTracker and AFF-Net PoG comparison numbers in Table II are taken."},{"cited_title":"Eye tracking for every- one,","cited_arxiv_id":null,"evidence_quote":"iTracker baseline and GazeCapture-based PoG approach that MAGE compares against on point-of-gaze error."},{"cited_title":"Adaptive feature fusion network for gaze tracking in mobile tablets,","cited_arxiv_id":null,"evidence_quote":"AFF-Net baseline that MAGE outperforms on both MPIIFaceGaze and EYEDIAP-Screen point-of-gaze error."},{"cited_title":"Gaze estimation using transformer,","cited_arxiv_id":null,"evidence_quote":"GazeTR transformer baseline that MAGE compares against on gaze-direction error across all datasets."},{"cited_title":"Agent-guided gaze estimation network by two-eye asymmetry exploration,","cited_arxiv_id":null,"evidence_quote":"AGE-Net baseline, the previous best direction-only method that MAGE surpasses on public benchmarks."},{"cited_title":"Easygaze3d: Towards effective and flexible 3d gaze estimation from a single rgb camera,","cited_arxiv_id":null,"evidence_quote":"Prior EasyGaze3D method that the Easy-Calibration screen-free calibration approach extends."}],"review_version":1}