{"id":"fe4d3312-70ca-488b-a8fb-ce09e12a4527","arxiv_id":"2502.02307","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Curated MAE pre-training on normalized, pose-balanced face images improves gaze estimation generalization across datasets, outperforming semantic pre-training and prior domain-generalization methods.","lead":"A new gaze-estimation model pre-trains a vision transformer on 1.6 million face images, with no gaze labels, using masked autoencoding, and then fine-tunes on small labeled gaze datasets. It reports consistently lower gaze-direction errors on unseen datasets than existing models, suggesting a recipe for better cross-domain gaze estimation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"XGaze-Dense pre-training data may inflate XGaze test columns because it is unlabeled target-domain data, not a truly unseen domain.","rationale":"The reader's weakest assumption identifies the right area, but my read is partial rather than full agreement: the paper explicitly excludes the 60 XGaze-Dense identities from the XGaze Test set, so this is not direct test-identity leakage. The sharper issue is unsupervised target-domain exposure: pre-training has seen unlabeled renderings of 75% of the ETH-XGaze identity pool and the same capture/reconstruction domain before the model is evaluated on XGaze Test. That weakens the 'unseen dataset' interpretation of every XGaze column in Tables 2, 5, and possibly 4. The non-XGaze cross-dataset results are substantial and would support the central claim even without the XGaze columns, which is why the appropriate disposition remains CONDITIONAL rather than REJECT. The missing ablation is cheap and decisive: remove XGaze-Dense from pre-training and measure the change in XGaze Test error. No repeated-seed error bars are reported, further limiting how strongly the XGaze improvements can be interpreted. The code/model release helps verification but does not by itself resolve the target-domain exposure question.","tokens_in":21082,"tokens_out":9020,"duration_ms":81061,"concrete_test":"Retrain UniGaze with the full pipeline but remove XGaze-Dense from the pre-training corpus while keeping FFHQ-NV and all other synthetic/real components; rerun the leave-one-dataset-out protocol (Table 5) and the cross-dataset runs whose columns include XTest (Tables 2b-2e). If the XGaze Test error degrades substantially (for example, more than 10% relative to the full-corpus model) while the non-XGaze columns stay roughly flat, the XGaze columns are inflated by pre-training exposure to the target dataset rather than by generic pre-training quality. Independently, compare the subject IDs in the [73] within-dataset split used in Table 4 with the 60 XGaze-Dense identities to confirm that no test identity overlaps.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that pre-training improves generalization to unseen gaze domains. For every XGaze column, that condition is questionable. Section 3.1 and Table 1 add XGaze-Dense, 267,160 unlabeled novel-view renderings of the 60 ETH-XGaze training identities, to the pre-training corpus. Section 4.2 notes that these 60 subjects are excluded from the XGaze Test set, so direct test-identity overlap is avoided. However, in the cross-dataset runs of Tables 2b-2e and in the leave-one-dataset-out protocol of Table 5, the XTest column evaluates ETH-XGaze after pre-training has already seen 75% of that dataset's identity pool and the same multi-camera/reconstruction pipeline. The statement in Section 3.1 that XGaze-Dense is 'used without gaze labels' and 'equivalent to a generic facial dataset' is the load-bearing step: it is not generic, because it supplies the target dataset's appearance and pose domain. Section 4.4 also inherits the [73] split for within-dataset ETH-XGaze without verifying that its test identities are disjoint from the 60 XGaze-Dense subjects. Since XTest shows some of the largest gains (11.29 vs 16.31 for ResNet-50 in Table 5), the 'unseen dataset' claim for the XGaze columns is not established without a no-XGaze-Dense ablation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UniGaze, a gaze estimation approach that pre-trains a ViT encoder with Masked Autoencoding on a curated corpus of about 1.6 million real and synthetic face images, then fine-tunes on labeled gaze datasets. The authors report consistent cross-dataset improvements over ImageNet pre-trained ViTs, CNN baselines, and prior domain-generalization methods on MPIIFaceGaze, GazeCapture, EYEDIAP, Gaze360, and ETH-XGaze. They also introduce leave-one-dataset-out and joint-dataset evaluation protocols and provide ablations on pre-training data size, composition, normalization, and pose diversity.","tokens_in":21375,"tokens_out":3978,"duration_ms":36190,"significance":"If the reported results hold, the paper provides a valuable practical recipe for gaze-estimation pre-training: use normalized face crops, balanced head-pose distributions, and high identity diversity, rather than generic semantic pre-training. The two proposed evaluation protocols are useful benchmarks for the field. The paper also releases source code and models, which supports reproducibility. The main empirical claim that pre-training improves generalization to unseen gaze domains is, however, weakened by the inclusion of XGaze-Dense in the pre-training corpus, because that dataset is derived from the same identity pool and acquisition pipeline as the ETH-XGaze evaluation set.","major_comments":[{"comment":"The pre-training corpus includes XGaze-Dense, 267,160 novel-view renderings of the 60 ETH-XGaze training subjects, created with the same multi-view reconstruction pipeline used to build ETH-XGaze. The statement in §3.1 that this data is 'equivalent to a generic facial dataset' is not justified: it is unlabeled target-domain data for every XGaze evaluation column in Tables 2, 4, 5, and 6. Since the central claim is generalization to unseen datasets, and the XGaze columns show some of the largest gains (e.g., 16.31 to 11.29 in Table 5), the paper must provide a no-XGaze-Dense ablation to show that the improvement on XGaze is not due to domain or identity memorization.","section":"§3.1, Table 1"},{"comment":"The within-dataset ETH-XGaze evaluation follows the train/test split from 3DGazeNet [73], but the paper does not verify that the test identities in that split are disjoint from the 60 subjects used to generate XGaze-Dense. If they overlap, the within-dataset XGaze number (3.96) is inflated. The authors should either verify disjointness explicitly or re-run the within-dataset evaluation using the standard 60/20 subject split.","section":"§4.4, Table 4"},{"comment":"All results are reported as single runs without error bars, confidence intervals, or repeated-seed variance. Given that some margins are small (e.g., Table 4: MPIIFaceGaze 4.07 vs. 4.0, Gaze360 9.44 vs. 9.6), the claim that UniGaze 'consistently' and 'significantly' outperforms baselines is not statistically supported. Adding at least three seeds with standard deviations for the main comparisons would substantially strengthen the paper.","section":"Tables 2-6 and Figure 3"},{"comment":"The comparison with AGG, CLIP-Gaze, LG-Gaze, and Gaze-BAR relies entirely on published numbers from works that may use different pre-processing pipelines. The paper re-implements only ResNet-18 and PureGaze, and the supplementary notes minor discrepancies even for those. The claim of 'surpassing SOTA domain generalization methods' is therefore not yet established for the non-re-implemented methods. The authors should either re-implement all compared methods with the same pre-processing or clearly qualify which comparisons are direct and which are indirect.","section":"Table 3"}],"minor_comments":[{"comment":"In the sentence 'FaRL-B [96] ... pre-trained on 20 MLAION-Face samples', 'MLAION' appears to be a typo for 'M LAION' (i.e., 20 million LAION-Face samples).","section":"§3.2"},{"comment":"The caption defines 'XTest' only in the table footnote; please define it in the main text at first use, and clarify that XGaze Test refers to the 20-subject held-out subset.","section":"§4.2, Table 2"},{"comment":"The sentence 'For GazeCapture and Gaze360, it is the same as defined in Sec. 4.1' should read 'they are the same as defined in Sec. 4.1' to avoid ambiguity about which split is being referenced.","section":"§4.4"},{"comment":"The vertical axis label 'Percentage Change from Baseline (%)' is ambiguous; it would be clearer to state that it is the relative error reduction with respect to the 0% (ImageNet pre-trained) model.","section":"Figure 3"},{"comment":"The statement '0% refers to the ImageNet pre-trained ViT-L' is clear, but the paper should also state whether the 0% baseline undergoes the same fine-tuning protocol as the pre-trained models; otherwise the comparison may not isolate the effect of pre-training.","section":"§4.7.1"},{"comment":"The hyperparameter ranges for color jitter and the probability of grayscale are given only in the supplementary; please also state them briefly in the main text for completeness, since these are part of the proposed pre-training recipe.","section":"Supplementary Table 3"}],"recommendation":"major_revision","confidential_remarks":"The XGaze-Dense leakage concern is the main barrier to acceptance; it strikes directly at the 'unseen domain' claim. I would also check whether the authors adequately differentiate their contribution from the concurrent work [33] that reports MAE failure for gaze estimation, since the paper cites it but does not discuss the apparent contradiction in depth. The two proposed protocols are a genuine contribution and the released code/model are a plus."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth reading. It gives a practical recipe for MAE pre-training for appearance-based gaze estimation and shows that off-the-shelf semantic pre-training (DINO, MoCo, FaRL) transfers poorly to gaze, while a curated pipeline—normalized face alignment, balanced head pose, identity diversity, and a mix of real/synthetic/novel-view data—consistently beats ImageNet-initialized ViTs on cross-dataset tests. The two new evaluation protocols (leave-one-dataset-out and joint-dataset) are useful, and the supplementary ablations (data composition, color jitter, patch normalization) are more thorough than most papers in this space. They also re-implemented some baselines to control for preprocessing, which is honest.\n\nThe main soft spot is the XGaze-Dense contamination. Section 3.1 adds 267k unlabeled novel-view renderings of the 60 ETH-XGaze training subjects to the pre-training corpus. The paper says this is \"equivalent to a generic facial dataset,\" but that is not right: it supplies the target domain's identity pool and capture pipeline. The XGaze Test column in Tables 2, 5, and 6 therefore measures generalization to a domain the pre-training has already seen—not an unseen domain. The paper excludes those 60 subjects from the test set, so direct identity leakage is avoided, but the domain overlap remains, and the supplementary Table 2 shows that adding novel-view data yields the largest gains in pose-heavy settings. Without an ablation that removes XGaze-Dense from pre-training, the XGaze columns are not interpretable as evidence for the \"unseen dataset\" claim. Section 4.4 also uses the [73] split for within-dataset ETH-XGaze without checking identity overlap with those 60 subjects; that split may draw test identities from the same pool.\n\nThe rest of the evidence is more solid: the non-XGaze test columns (MPIIFaceGaze, GazeCapture, EYEDIAP, Gaze360) show gains where the pre-training data does not include the target domain. No error bars or repeated runs are reported, which is minor for a pre-training study but worth noting. The SOTA comparison mixes cited and re-implemented numbers, but they re-implemented the key baselines and the improvement is large enough that preprocessing differences likely do not explain it.\n\nBottom line: the recipe and the negative results are a genuine contribution, and the paper deserves a serious referee. The referee needs to require a no-XGaze-Dense ablation and a verification of the Section 4.4 split. If those come back clean, the paper is a solid acceptance; as is, treat the XGaze columns with suspicion and rely on the other test sets for the central claim.","headline":"A solid empirical recipe for MAE pre-training in gaze estimation, but the XGaze test columns are likely inflated: the pre-training set includes novel-view renderings of the same 60 ETH-XGaze identities used for fine-tuning.","tokens_in":21878,"tokens_out":2957,"would_cite":true,"duration_ms":25634,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Masked-autoencoder pre-training on 1.6 million curated in-the-wild face images gives gaze-estimation models lower error on every unseen dataset in the paper's battery, while pre-training designed for semantic tasks fails to help.","keywords":["gaze estimation","self-supervised pre-training","masked autoencoder","vision transformer","cross-domain generalization","domain generalization","face representation learning","head pose distribution"],"falsifier":"Re-run the pipeline with XGaze-Dense removed from pre-training, or replaced by novel-view renders of 60 identities disjoint from ETH-XGaze, and compare the XGaze test error on the strict 20-subject test split used in the supplementary. If the XGaze scores stay about the same, the claim that pre-training learns generalizable face geometry survives; if they degrade substantially, part of the gain comes from identity overlap. A complementary check is to list the test subjects in the Table 4 within-dataset split and verify that none of them belong to the 60 identities rendered into XGaze-Dense.","tokens_in":20870,"feed_emoji":"👁️","tokens_out":15721,"duration_ms":127999,"temperature":0.7,"pith_summary":"This paper claims that self-supervised pre-training can substantially improve how well gaze-estimation models transfer to unseen environments — but only when the pre-training is built for gaze's geometric nature rather than for semantic understanding. The authors pre-train a Vision Transformer with Masked Autoencoders (a method that hides random image patches and trains the network to rebuild them) on roughly 1.6 million in-the-wild face images, curated so that faces are normalized on the same geometry downstream gaze models expect, head poses are diverse yet balanced, and identity count is high. Across three evaluation protocols — training on one dataset and testing on others, leaving one dataset out of training, and pooling all datasets into one training set — the resulting UniGaze model reports lower gaze error than CNNs, plain ViTs, semantic self-supervised models, and specialized domain-generalization methods on five public gaze benchmarks. The reason to care: if the recipe holds, gaze systems can be deployed in new environments without collecting new labeled data, and the pre-training principles become a transferable template for other geometric regression tasks.","feed_headline":"Pretraining on 1.6M faces lifts gaze accuracy across five datasets","feed_subtitle":"No new labels: a masked-autoencoder recipe on curated faces beats specialized gaze-domain methods on five benchmarks.","key_machinery":"The load-bearing mechanism is Masked Autoencoder (MAE) pre-training on a Vision Transformer: 75% of image patches are randomly masked and the encoder–decoder is trained to reconstruct the masked pixels in per-patch normalized space, so the encoder learns facial structure from roughly 1.6 million images without any gaze labels. Around this sits a curated corpus spanning over 260,000 identities — real video and web data (CelebV-Text, VFHQ, VGGFace2) plus synthetic and novel-view-rendered faces (FaceSynthetics, SFHQ-T2I, FFHQ-NV, XGaze-Dense) — all resized to 224x224 through the standard gaze normalization pipeline, with head pose estimated by landmark detection and perspective-n-point to filter extreme angles and measure pose balance. The paper's three discovered prerequisites — normalized input space, balanced wide head-pose coverage, and identity diversity — are what the ablation studies show to be doing the work. A second contribution is procedural: the leave-one-dataset-out and joint-dataset evaluation protocols measure whether one model can serve all domains, which single-dataset cross-evaluation cannot.","core_discovery":"UniGaze's central claim is that masked autoencoding on a large, deliberately curated collection of faces — rather than merely more data or a bigger model — is what unlocks cross-domain gaze estimation. The paper shows that off-the-shelf semantic pre-training (DINO, MoCo-v3, FaRL) and vanilla MAE trained on small gaze data fail to transfer, sometimes doing worse than a plain ResNet-50, and that increasing ViT model size without the right pre-training also fails. Its ablations isolate three load-bearing ingredients: pre-training on normalized face crops that keep the spatial alignment used by downstream gaze models; wide but balanced head-pose distributions so all viewing angles are represented; and over 260,000 distinct identities so facial appearance varies. With these ingredients, UniGaze-H reports the lowest errors in nearly every cross-dataset cell of its tables (for instance 5.57 degrees on ETH-XGaze-to-MPIIFaceGaze versus a best prior of 5.91), and the paper introduces leave-one-dataset-out and joint-dataset protocols as more realistic measures of generalization. The authors' conclusion is that large-scale pre-training helps gaze estimation only when the pre-training task, the input normalization, and the data distribution are aligned with the geometry of gaze.","pith_inferences":["Because each ablation changes pose balance or identity diversity together with total data volume, the paper leaves open which of its three ingredients dominates; a controlled set of synthetic faces with identity count fixed and pose range widened would isolate the mechanism.","The same recipe plausibly transfers to neighbouring geometric tasks — head-pose estimation, gaze target detection, or 3D face alignment — where semantic pre-training has similarly underperformed; the paper only demonstrates the effect for gaze direction regression.","The pre-training corpus is partly procedural (novel-view and diffusion-generated faces), pointing to a scaling path the paper only hints at: procedurally generating faces with controllable pose and identity could grow the corpus well beyond 1.6 million images, shifting the bottleneck from labeled data to rendering compute."],"forward_implications":["Practitioners can expect accuracy gains on unseen cameras, identities, and lighting conditions without new labeled gaze data, since UniGaze improves on nearly every cross-dataset transfer direction reported.","The benefit scales with pre-training data: error drops consistently as the corpus grows from 25% to 100% of 1.6 million images, with diminishing returns past 75%, making curated data collection a predictable route to better gaze models.","Model size pays off only when paired with gaze-appropriate pre-training: an ImageNet-pretrained ViT-Huge often loses to ResNet-50, while UniGaze-Huge is best in most settings, so scaling alone is not a substitute for the right recipe.","The two proposed protocols, leave-one-dataset-out and joint-dataset training, give the field a standard way to measure the practical goal of one gaze model serving all domains.","Pre-training cannot compensate for a narrow label range at fine-tuning time, since models trained on MPIIFaceGaze still fail on wide-pose datasets, so label diversity in the downstream stage remains essential."],"supporting_citations":[{"why":"Supplies the Masked Autoencoder pre-training framework that UniGaze adapts to face images.","marker":"[29]"},{"why":"Defines the data-normalization pipeline whose aligned input space is the first prerequisite for effective pre-training.","marker":"[90]"},{"why":"ETH-XGaze is both a main benchmark and the source of the 60 subjects re-rendered into the XGaze-Dense pre-training set.","marker":"[92]"},{"why":"Provides the novel-view-synthesis method used to generate FFHQ-NV and XGaze-Dense pose-expanded pre-training images.","marker":"[61]"},{"why":"Supplies the within-dataset ETH-XGaze split used in Table 4 and the 3DGazeNet synthetic-view baseline UniGaze is compared against.","marker":"[73]"},{"why":"FaRL is the face-centric pre-training baseline whose failure to improve gaze motivates the gaze-specific recipe.","marker":"[96]"},{"why":"DINO represents the semantic self-supervised pre-training class that the paper shows does not transfer to gaze estimation.","marker":"[10]"},{"why":"PureGaze is a state-of-the-art domain-generalization method that UniGaze must outperform in the cross-dataset comparison.","marker":"[13]"}],"fun_headline_variants":["Masked autoencoding on 1.6M faces improves cross-domain gaze","Why more data fails: normalized crops and pose balance are key for gaze pretraining","1.6M faces pretraining: masked autoencoding outperforms specialized gaze models","Cross-dataset gaze leaps: UniGaze reveals what makes pre-training work","From semantic to spatial: MAE on curated faces beats gaze-specific baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported ETH-XGaze scores assume that pre-training on XGaze-Dense — novel-view renderings of the same 60 subjects used in the joint-dataset fine-tuning and possibly in the within-dataset split — teaches general face geometry instead of memorizing those particular faces; if that assumption fails, the XGaze columns in the main tables overstate the generalization gain.","fun_headline_variants_meta":{"raw":{"variants":["Masked autoencoding on 1.6M faces improves cross-domain gaze","Why more data fails: normalized crops and pose balance are key for gaze pretraining","1.6M faces pretraining: masked autoencoding outperforms specialized gaze models","Cross-dataset gaze leaps: UniGaze reveals what makes pre-training work","From semantic to spatial: MAE on curated faces beats gaze-specific baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001113,"raw_usage":{"total_tokens":4654,"prompt_tokens":983,"completion_tokens":3671,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":3568}},"tokens_in":599,"tokens_out":3671,"duration_ms":21500,"temperature":1.0,"reasoning_tokens":3568,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T12:35:45.845224+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the pipeline with XGaze-Dense removed from pre-training, or replaced by novel-view renders of 60 identities disjoint from ETH-XGaze, and compare the XGaze test error on the strict 20-subject test split used in the supplementary. If the XGaze scores stay about the same, the claim that pre-training learns generalizable face geometry survives; if they degrade substantially, part of the gain comes from identity overlap. A complementary check is to list the test subjects in the Table 4 within-dataset split and verify that none of them belong to the 60 identities rendered into XGaze-Dense.","supporting_citations":[{"cited_title":"Masked autoencoders are scalable vision learners","cited_arxiv_id":null,"evidence_quote":"Supplies the Masked Autoencoder pre-training framework that UniGaze adapts to face images."},{"cited_title":"Re- visiting data normalization for appearance-based gaze esti- mation","cited_arxiv_id":null,"evidence_quote":"Defines the data-normalization pipeline whose aligned input space is the first prerequisite for effective pre-training."},{"cited_title":"Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze variation","cited_arxiv_id":null,"evidence_quote":"ETH-XGaze is both a main benchmark and the source of the 60 subjects re-rendered into the XGaze-Dense pre-training set."},{"cited_title":"Learning-by-novel-view-synthesis for full-face appearance- based 3d gaze estimation","cited_arxiv_id":null,"evidence_quote":"Provides the novel-view-synthesis method used to generate FFHQ-NV and XGaze-Dense pose-expanded pre-training images."},{"cited_title":"3dgazenet: Generalizing 3d gaze estimation with weak- supervision from synthetic views","cited_arxiv_id":null,"evidence_quote":"Supplies the within-dataset ETH-XGaze split used in Table 4 and the 3DGazeNet synthetic-view baseline UniGaze is compared against."},{"cited_title":"General facial representation learning in a visual-linguistic manner","cited_arxiv_id":null,"evidence_quote":"FaRL is the face-centric pre-training baseline whose failure to improve gaze motivates the gaze-specific recipe."}],"review_version":1}