{"id":"edec5ad3-8d90-4da5-9194-4d63845750a3","arxiv_id":"2412.08074","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"EM-Net, a MobileNetV3-based gaze estimator with a Swin-style attention branch and an EM refinement module, reports angular-error gains of 0.08 to 0.24 degrees over GazeNAS-ETH using 50% training data.","lead":"EM-Net is a lightweight gaze-estimation model that pairs a MobileNetV3 backbone with a self-attention branch and an expectation-maximization refinement stage. On three public datasets it reports angular errors slightly lower than GazeNAS-ETH while training on only half the data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim rests on an uncontrolled comparison: EM-Net's 0.08–0.24° margins over GazeNAS-ETH are computed against published numbers generated with 100% data and no ETH-XGaze pretraining, so the '50% data' advantage is unverified.","rationale":"This is the single load-bearing concern because the central contribution—data efficiency while beating SOTA—is entirely empirical. The EM module is underspecified (Algorithm 1 never defines Y, Z, θ, or iteration count), but even a fully specified module would only matter if the comparison were valid. The comparison is invalid as presented because it mixes protocols: EM-Net uses 50% interval-sampled data and ETH-XGaze pretraining, while baseline numbers are taken from original papers using 100% data and their own pipelines. The margins are tiny and unaccompanied by error bars. A same-protocol rerun is feasible because GazeNAS-ETH has public code. If the rerun confirms a real margin, the paper would be conditionally acceptable; if not, the headline claim fails. The reader's weakest assumption identified the same issue, so my assessment agrees. No mathematical circularity or authorial intent is invoked.","tokens_in":12287,"tokens_out":2468,"duration_ms":24624,"concrete_test":"Obtain the released GazeNAS-ETH implementation and retrain it on exactly the odd-indexed 50% subsets of MPIIFaceGaze, Gaze360, and RT-Gene used by EM-Net, with the same ETH-XGaze pretraining, input resolution, and evaluation protocol (including the same train/test split). Run at least three seeds and report mean±std angular error per dataset. If GazeNAS-ETH's mean error is within one standard deviation of EM-Net's reported 3.88°/10.29°/6.27° on any dataset, the claimed accuracy/data-efficiency advantage is not demonstrated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is the 2% accuracy improvement over GazeNAS-ETH using half the training data. The evidence is Table 2, whose baseline entries are explicitly 'the data in the original paper' (§4.3). EM-Net's protocol differs in two material ways: it uses only odd-indexed 50% of each training set (§4.2) and it initializes from ETH-XGaze pretraining (§4.1). Neither change is applied to the baselines. The reported margins are 0.08°, 0.24°, and 0.13°—small enough to be within normal run-to-run variance for gaze networks, and no error bars or multiple seeds are provided. Without a same-protocol re-run of GazeNAS-ETH, the claim that EM-Net beats it with half the data is not established. The ablation (Table 6) and GAM comparison (Table 4) are self-consistent but do not validate the headline comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EM-Net, a lightweight gaze estimation network built on an improved MobileNetV3 backbone and augmented with a Global Attention Mechanism (GAM) and an Expectation Maximization (EM) module. The authors claim that EM-Net achieves lower angular errors than GazeNAS-ETH and other baselines on MPIIFaceGaze, Gaze360, and RT-Gene while using only 50% of the training data, and that it is robust to Gaussian noise. The experiments report single-run angular errors, parameter counts, FLOPs, and inference times, along with ablation studies and attention-replacement comparisons.","tokens_in":12375,"tokens_out":7981,"duration_ms":74900,"significance":"If the central claims were properly validated, the paper would contribute a lightweight gaze estimation model with a notably low parameter count (2.93M) and computational cost (0.31 GFLOPs), and would demonstrate an interesting data-efficiency property via the EM module. However, the current evidence is not convincing: the headline comparison against published baseline numbers is uncontrolled, the EM module is described only at the level of generic EM equations without a concrete instantiation, and all quantitative results lack error bars or repeated-seed runs. The ablation and attention comparisons show plausible trends, but the small differences (as low as 0.08°) could easily be within run-to-run variance.","major_comments":[{"comment":"The central claim of a 0.08°–0.24° improvement over GazeNAS-ETH using half the training data is based on baseline angular errors cited from the original papers, not from re-running those models under EM-Net's protocol. EM-Net differs in two material ways: it uses only odd-indexed 50% of the training samples (§4.2) and it initializes from ETH-XGaze pretraining (§4.1). Without a same-protocol re-run of GazeNAS-ETH (or at least training GazeNAS-ETH on the same 50% subset with the same pretraining), the comparison is not controlled, and the claimed margins are small enough that they could be explained by training noise or the pretraining initialization. The '50% data' advantage is therefore unverified.","section":"Section 4.3, Table 2"},{"comment":"The EM module is not actually specified as a network layer. Algorithm 1 states the textbook EM update equations for a generic probabilistic model P(Y,Z|θ), but the paper does not define what Y, Z, and θ are in the context of the feature maps f∈R^{4×4×540}, nor how the E-step and M-step are computed on those tensors. The line 'θ(i) = Conv1×1' is not a meaningful initialization of model parameters for a probability model. Consequently, the module is not reproducible, and the assertion that it 'retains the characteristic that the EM algorithm is good at processing unobserved variables' (§4.3) has no concrete algorithmic content.","section":"Section 3.2, Algorithm 1"},{"comment":"All angular errors in the paper are reported without standard deviations, confidence intervals, or the number of repeated runs. This is particularly problematic because the central comparison yields differences as small as 0.08° (MPIIFaceGaze) and the GAM-vs-SE/CBAM and ablation differences in Tables 4 and 6 are 0.10°–0.74°. Without error bars, one cannot distinguish the proposed method's reported improvements from run-to-run variation, especially given the known variance of gaze estimation results across training runs.","section":"Sections 4.3, 4.4, 4.6 (Tables 2, 4, 6)"}],"minor_comments":[{"comment":"In the E-step, Q(θ, θ(i+1)) should be Q(θ, θ(i)); the same index correction applies to the maximization step.","section":"Algorithm 1"},{"comment":"The percentage improvement for Gaze360 is reported as 2.2%, but 0.24° relative to 10.52° is approximately 2.28%; the rounding should be made consistent.","section":"Section 4.3"},{"comment":"Clarify whether the odd-indexed sampling is applied only to the training split or to the entire dataset, and specify how the test set is constructed; this affects the comparability of the reported errors.","section":"Section 4.2"},{"comment":"State explicitly whether Gaussian noise is added to the test images as well as the training images in the robustness experiments, and whether the same noise level is used for both.","section":"Section 4.5"},{"comment":"The GAM appears to combine SE-like channel attention with Swin Transformer's shifted-window self-attention; the novelty relative to the Swin Transformer block should be discussed, and a direct comparison with a standard Swin block would be more informative than the SE/CBAM comparison alone.","section":"Section 3.1.2"},{"comment":"The text states that inference times are computed under the same hardware conditions, but it is unclear whether the baseline times are re-measured or taken from their original papers; the subsequent sentence saying parameters and FLOPs are cited from original papers adds ambiguity.","section":"Table 3"},{"comment":"The paper contains numerous English grammar and missing-article errors (e.g., 'the proposed Global Attention Mechanism(GAM) is added to extract features related to gaze estimation to improve the model's ability'); a careful proofread is needed.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses an interesting direction but the current evidence is not sufficient for publication. The reader's reject recommendation is defensible; I lean toward major_revision because the central flaws—uncontrolled baseline comparison and underspecified EM module—are technically fixable with additional experiments and a rewritten method section, assuming the authors have access to the baseline code and compute. The small reported gains make it especially important to add error bars or multiple seeds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper's real content is EM-Net, a lightweight gaze estimator that combines an improved MobileNetV3 with a global attention branch (GAM) and an EM module. The GAM is a hybrid of SE channel gating and Swin shifted-window attention; the EM module is meant to make features robust to occlusion and noise. The ablation runs show both pieces help, and the receptive-field visualizations support the attention claim. So there is a plausible architecture contribution here.\n\nThe central claim, however, is not supported as stated. Table 2 lists EM-Net and baselines with EM-Net using 50% of the training data and ETH-XGaze pretraining, while the baseline numbers are taken from the original papers (§4.3: 'the angle errors of the models included in Table 2 all refer to the data in the original paper'). That means GazeNAS-ETH and the others are not re-run under the authors' protocol. The margins are 0.08°, 0.24°, and 0.13° — small enough to be noise for this task — and no error bars or seeds are given. Without a same-protocol re-run, the '50% data' advantage is unverified.\n\nThe EM module itself is also under-specified. Algorithm 1 just rewrites the textbook EM update; the observed variable, hidden variable, parameters, and iteration count are never defined. That leaves the mechanism unfalsifiable from the text, and no code or weights are provided to check the reported parameters.\n\nI want to be fair: the paper does not oversell against Gaze-Swin, and the robustness experiment with Gaussian noise is a genuine attempt to show where the EM module might help. The ablation table is internally consistent.\n\nFor a revision, the authors must re-run at least GazeNAS-ETH under their exact sampling and pretraining, report multiple seeds, and either release code or make the EM module concrete. If they cannot re-run the baselines, the headline comparison should be replaced with an honest reference table. As it stands, I would not cite it, but the idea is not dead. A serious referee could push the authors to fix the protocol gaps, so I would send it to review with a request for major revision rather than desk-reject it.","headline":"A plausible architecture with an uncontrolled head-to-head: the 50%-data claim rests on published baselines not retrained under the same protocol.","tokens_in":13017,"tokens_out":3136,"would_cite":false,"duration_ms":29910,"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":"This paper proposes EM-Net, a lightweight gaze estimator that combines a global attention mechanism with an expectation-maximization module, and claims it beats the GazeNAS-ETH baseline on three public benchmarks using only half the…","keywords":["gaze estimation","expectation maximization","global attention mechanism","lightweight network","shifted-window self-attention","data efficiency","MPIIFaceGaze","Gaze360"],"falsifier":"Re-run GazeNAS-ETH under exactly EM-Net's training protocol—the same odd-indexed 50% split, the same ETH-XGaze initialization, the same preprocessing—and evaluate with the same angular-error metric. If GazeNAS-ETH's error then equals or beats EM-Net's on MPIIFaceGaze, Gaze360, or RT-Gene, the paper's central comparison collapses.","tokens_in":11937,"feed_emoji":"👀","tokens_out":7940,"duration_ms":68037,"temperature":0.7,"pith_summary":"This paper tries to show that a lightweight gaze-estimation model can be accurate without large datasets or large models. EM-Net is built on a slimmed MobileNetV3 backbone; a proposed Global Attention Mechanism (GAM) uses shifted-window self-attention to enlarge the receptive field and fuse spatial and channel information, while an Expectation-Maximization (EM) module refines the CNN features by iteratively estimating latent structure. The paper's central experimental claim is that, trained on only 50% of the training samples of MPIIFaceGaze, Gaze360, and RT-Gene, EM-Net reduces angular error relative to GazeNAS-ETH by 2.02%, 2.2%, and 2.03%, respectively. If this holds, accurate gaze estimation becomes cheaper to train and easier to deploy on resource-limited devices.","feed_headline":"Lightweight gaze model cuts error with half the data","feed_subtitle":"EM-Net combines global attention with expectation maximization to beat a stronger baseline on three gaze benchmarks.","key_machinery":"The load-bearing machinery is the pairing of a Global Attention Mechanism (GAM) with an Expectation-Maximization module inserted between a MobileNetV3-derived backbone and gaze regression. GAM reuses the shifted-window and masking idea from Swin Transformer to exchange information across spatial windows and across channels, explicitly enlarging the effective receptive field. The EM module first projects the 960-channel feature map down to 540 channels with a 1x1 convolution, then iterates the classic E-step/M-step update on the feature distribution: it computes the expected log-likelihood over hidden variables given current parameters, then re-estimates parameters by maximization. This iteration is what the paper claims lets the network deal with occluded or noisy facial regions and generalize from fewer samples.","core_discovery":"The paper's central discovery is that a carefully combined attention-and-EM architecture can maintain or improve gaze accuracy while halving the training data. EM-Net reports angular errors of 3.88° on MPIIFaceGaze, 10.29° on Gaze360, and 6.27° on RT-Gene, improvements of 2.02%, 2.2%, and 2.03% over GazeNAS-ETH, with 2.93M parameters and 0.31 GFLOPs. The authors attribute the gains to GAM, which captures long-distance dependencies through shifted-window multi-head self-attention and masks, and to the EM module, which iterates between an expectation step and a maximization step over the feature map so that occluded or noisy input regions are handled statistically. In ablation, removing either GAM or the EM module costs a similar amount of accuracy, and removing both raises the Gaze360 error by 0.74° (7.2%).","pith_inferences":["Because the reported margins over GazeNAS-ETH are only 0.08° to 0.24°, a controlled re-run under identical training splits, pretraining, and preprocessing could easily reverse the ranking; the 'data efficiency' conclusion would be stronger if the baselines had been retrained on the same 50% subset.","The EM module, as implemented, is a feature-refinement block rather than a full generative model fit; ablating the number of EM iterations (including a zero-iteration pass) would show whether the improvement comes from the EM principle or simply from the extra computation.","If the generalization benefit is real, it should become more pronounced as the training fraction drops; testing at 25% and 10% data would provide a direct, falsifiable extension.","GAM borrows shifted-window self-attention from the Swin design, so the comparison against SE and CBAM mainly shows that some form of windowed self-attention helps; swapping GAM for a standard Swin attention module at equal FLOPs would isolate what the GAM-specific arrangement adds."],"forward_implications":["Deploying gaze estimation on edge devices becomes more feasible: EM-Net's 2.93M parameters and 0.31 GFLOPs are an order of magnitude smaller than many transformer-based competitors while maintaining competitive accuracy.","Training cost drops: using half the training data of three standard benchmarks directly halves annotation and processing effort for this task.","The EM module's robustness to Gaussian noise (angle errors degrade only roughly 0.2–0.7° on MPIIFaceGaze and 0.4–0.9° on Gaze360) suggests the architecture is suited to real-world images with sensor noise or occlusion.","The comparison implies that receptive-field expansion (via GAM) and latent-feature refinement (via EM) are roughly equally important, since ablating either one costs a similar amount of accuracy."],"supporting_citations":[{"why":"Supplies the state-of-the-art baseline whose published angular errors EM-Net claims to beat while using half the data.","marker":"[33]"},{"why":"Larger hybrid CNN-transformer baseline whose accuracy and much higher parameter/FLOP counts frame EM-Net's efficiency comparison.","marker":"[34]"},{"why":"Transformer-based gaze method compared on all three datasets; EM-Net claims lower errors with 74.3% fewer parameters and 83.1% fewer FLOPs.","marker":"[35]"},{"why":"Dataset (and full-face estimation approach) that provides one of the three benchmarks and the evaluation protocol.","marker":"[20]"},{"why":"Dataset for the second benchmark, with in-the-wild images and the published baseline numbers.","marker":"[31]"},{"why":"Third benchmark dataset, including occluded-eye samples relevant to the EM module's robustness claim.","marker":"[32]"},{"why":"Backbone architecture that EM-Net slimmed down; supplies the base convolutional structure and h-swish design.","marker":"[17]"},{"why":"Source of the shifted-window multi-head self-attention design reused in GAM, including the masking step.","marker":"[28]"},{"why":"Large pretraining dataset used to initialize EM-Net before fine-tuning on each benchmark.","marker":"[30]"}],"fun_headline_variants":["Gaze estimation gets lighter and sharper with EM-Net","Half the data, better gaze accuracy with EM-Net","EM-Net: lean gaze model beats baselines with fewer samples","Attention plus EM trims gaze error and data needs","Lighter gaze model, half the data, better results"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim depends on the assumption that the GazeNAS-ETH numbers quoted from its original paper are directly comparable to EM-Net's results, even though EM-Net trains on a different 50% subsample, uses ETH-XGaze pretraining, and applies its own preprocessing.","fun_headline_variants_meta":{"raw":{"variants":["Gaze estimation gets lighter and sharper with EM-Net","Half the data, better gaze accuracy with EM-Net","EM-Net: lean gaze model beats baselines with fewer samples","Attention plus EM trims gaze error and data needs","Lighter gaze model, half the data, better results"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000533,"raw_usage":{"total_tokens":2558,"prompt_tokens":929,"completion_tokens":1629,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":1547}},"tokens_in":545,"tokens_out":1629,"duration_ms":12481,"temperature":1.0,"reasoning_tokens":1547,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:15:27.349822+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run GazeNAS-ETH under exactly EM-Net's training protocol—the same odd-indexed 50% split, the same ETH-XGaze initialization, the same preprocessing—and evaluate with the same angular-error metric. If GazeNAS-ETH's error then equals or beats EM-Net's on MPIIFaceGaze, Gaze360, or RT-Gene, the paper's central comparison collapses.","supporting_citations":[{"cited_title":"Searching efficient neural architecture with multi-resolution fusion transformer for appearance-based gaze estimation,","cited_arxiv_id":null,"evidence_quote":"Supplies the state-of-the-art baseline whose published angular errors EM-Net claims to beat while using half the data."},{"cited_title":"Gaze-swin: Enhancing gaze estimation with a hybrid cnn-transformer network and dropkey mechanism,","cited_arxiv_id":null,"evidence_quote":"Larger hybrid CNN-transformer baseline whose accuracy and much higher parameter/FLOP counts frame EM-Net's efficiency comparison."},{"cited_title":"Gaze estimation using trans- former,","cited_arxiv_id":null,"evidence_quote":"Transformer-based gaze method compared on all three datasets; EM-Net claims lower errors with 74.3% fewer parameters and 83.1% fewer FLOPs."},{"cited_title":"It’s written all over your face: Full-face appearance-based gaze estimation,","cited_arxiv_id":null,"evidence_quote":"Dataset (and full-face estimation approach) that provides one of the three benchmarks and the evaluation protocol."},{"cited_title":"Gaze360: Physically unconstrained gaze estimation in the wild,","cited_arxiv_id":null,"evidence_quote":"Dataset for the second benchmark, with in-the-wild images and the published baseline numbers."},{"cited_title":"Rt-gene: Real-time eye gaze estimation in natural environ- ments,","cited_arxiv_id":null,"evidence_quote":"Third benchmark dataset, including occluded-eye samples relevant to the EM module's robustness claim."},{"cited_title":"Searching for mobilenetv3,","cited_arxiv_id":null,"evidence_quote":"Backbone architecture that EM-Net slimmed down; supplies the base convolutional structure and h-swish design."},{"cited_title":"Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze varia- tion,","cited_arxiv_id":null,"evidence_quote":"Large pretraining dataset used to initialize EM-Net before fine-tuning on each benchmark."}],"review_version":1}