{"id":"747374f7-8bb6-49db-bcca-dc4a7424e92d","arxiv_id":"2411.18061","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Multitask-Gaze reports gaze-estimation error reductions of 1.7% on Gaze360 and 2.7% on MPIIFaceGaze over SUGE with sharply reduced compute.","lead":"This paper presents a lightweight neural network for gaze estimation that combines factorized convolutions, attention, and multi-task regression heads. The authors report lower angular error than the SUGE baseline on the MPIIFaceGaze and Gaze360 datasets while using about 75% fewer parameters and 87% fewer FLOPs.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline accuracy comparison to SUGE is not verifiable because Section III-A omits the evaluation protocol, and the reported margins are small enough to be explained by protocol or run-to-run variance.","rationale":"The reader's conditional verdict is well calibrated. The central claim is empirical and rests entirely on Table I, whose rows are assumed to be comparable. The paper provides no protocol details in Section III-A, no code, and no repeated-run statistics, so the comparison cannot be reproduced from the manuscript. I did not find a stronger internal inconsistency: the UC parameter reduction arithmetic is approximately correct for standard factorized convolutions, the ablation anomaly is explicitly acknowledged as possible overfitting, and the individual modules are plausible. The missing protocol is the single most load-bearing problem because the claimed improvements are small (0.11° and 0.18°) and could easily be reversed under a different preprocessing, split, or training schedule. The concrete test I propose would settle the question by directly re-running the two models under one shared protocol; if the margins survive, the central claim stands, and if not, the paper should be revised to a weaker claim about efficiency rather than accuracy. I therefore recommend keeping the reader's CONDITIONAL verdict rather than upgrading or rejecting.","tokens_in":7172,"tokens_out":5424,"duration_ms":47970,"concrete_test":"Obtain SUGE's official evaluation configuration and the released Multitask-Gaze implementation, then re-run both models under an identical protocol: same input resolution, same face-crop and normalization, same train/validation/test splits, same pretraining data, and at least five random seeds. Report mean angular error and standard deviation. If SUGE's MPII or Gaze360 error changes by more than about 0.1° under that protocol, or if Multitask-Gaze's margin over SUGE does not exceed the run-to-run noise, the headline 1.71%/2.75% claim is not established. In addition, re-run the Table II '-ALL' ablation on MPII with multiple seeds to test whether the full model's 3.90° versus 3.84° gap is stable or just a single-run artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the Table I comparison: Multitask-Gaze beats SUGE on MPIIFaceGaze (3.90° vs 4.01°) and Gaze360 (10.33° vs 10.51°) while using 2.822M parameters and 0.243G FLOPs. For this claim to hold, every row in Table I must have been produced under the same evaluation protocol. Section III-A gives only 'ETH-XGaze is used as the pre-training dataset'; it does not report input resolution, face-crop and normalization, train/validation/test splits, batch size, epochs, optimizer, learning-rate schedule, or the exact SUGE configuration used to fill in the baseline row. The margins are small: 0.11° on MPII and 0.18° on Gaze360. These are comparable to typical benchmark run-to-run variance, and Table II shows an internal warning: the full model has higher MPII error than the stripped '-ALL' model (3.90° vs 3.84°), which the authors attribute to overfitting. That makes the MPII margin over SUGE especially fragile. Without code or a detailed protocol, the claimed 1.71% and 2.75% improvements cannot be independently confirmed. The architecture itself is coherent, but the empirical comparison is underspecified at exactly the point where it must be exact.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Multitask-Gaze, a lightweight gaze-estimation network built on MobileNetV3. It replaces depthwise convolutions with unidirectional convolutions, inserts spatial-and-channel attention modules, replaces global average pooling with a Global Convolution Module, and adds a multi-task regression head with separate yaw/pitch predictors and a joint predictor. The claimed result is that Multitask-Gaze outperforms the SUGE baseline on MPIIFaceGaze (3.90° vs 4.01°) and Gaze360 (10.33° vs 10.51°) while reducing parameters by 75.5% and FLOPs by 86.88%. The paper also reports ablations and receptive-field visualizations.","tokens_in":7439,"tokens_out":5904,"duration_ms":49711,"significance":"The parameter and FLOP reductions versus SUGE are large and, if the accuracy numbers are reproducible, would be practically useful for real-time gaze estimation. The architectural ideas are clearly described and the module-level motivation is reasonable. However, the paper provides no code, no evaluation protocol details, no repeated-run statistics, and no public artifacts that would allow the central empirical claim to be checked. The work is not circular: it is an empirical architecture comparison, but that comparison currently rests on underspecified single-run numbers.","major_comments":[{"comment":"The central comparison to SUGE is not verifiable because the evaluation protocol is missing. The implementation details state only that ETH-XGaze is the pretraining dataset and that MAE loss is used; they do not specify input resolution, face cropping and normalization, train/validation/test split, optimizer, learning-rate schedule, batch size, epochs, pretraining procedure, or how the SUGE baseline numbers in Table I were produced. The margins being claimed (0.11° on MPIIFaceGaze and 0.18° on Gaze360) are small enough that they could easily be explained by protocol differences or run-to-run variance. Please provide the full protocol, code or a detailed configuration, and either error bars over multiple seeds or a statistical comparison.","section":"III-A, Table I"},{"comment":"The ablation study contradicts the claim that each proposed module improves accuracy. On MPIIFaceGaze the full model has the highest error (3.90°), while removing MRM, SCA, GCM, or all modules decreases the error to 3.88°, 3.86°, 3.88°, and 3.84°, respectively. The text attributes this to slight overfitting, but no regularization details, validation protocol, or variance estimates are reported. As written, Table II does not support the paper's contribution claims on MPIIFaceGaze; the authors should either demonstrate that the differences are meaningful (e.g., with multiple runs and an actual overfitting analysis) or revise the claims.","section":"III-C, Table II"}],"minor_comments":[{"comment":"The improvement percentages are swapped. MPIIFaceGaze improves by (4.01−3.90)/4.01 ≈ 2.74%, while Gaze360 improves by (10.51−10.33)/10.51 ≈ 1.71%; the abstract states the reverse.","section":"Abstract"},{"comment":"'Gazet-Swin' should be 'Gaze-Swin'.","section":"III-B"},{"comment":"When comparing to MobileNetV3, the FLOPs actually increase from 0.233G to 0.243G; the text mentions only the 32.9% parameter reduction and should state this trade-off explicitly.","section":"III-B"},{"comment":"The text says different weight parameters are assigned to yaw and pitch, but A1 and B1 are shared by both; please clarify whether a single weight pair is intended.","section":"II-C, Eq. (2)"},{"comment":"The Avg-pooling row in Table III is identical to the -GCM row in Table II. If this is intentional because GCM replaces the pooling layer, say so explicitly and explain what additional information the table provides beyond the ablation.","section":"III-D, Table III"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is under-specified; the main risk is not circularity but reproducibility. I would consider acceptance only after the protocol, code, and multi-run statistics are supplied. The venue may also want to verify the SUGE baseline configuration and the MobileNetV3 adaptation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is an incremental engineering contribution that combines known components — factorized convolution, spatial/channel attention, conv-based pooling replacement, multi-task heads — into a MobileNetV3-style gaze estimator. It reports slightly better angular error than SUGE on two benchmarks with far fewer parameters and FLOPs. That claim is plausible but not verifiable from the manuscript as written.\n\nWhat’s actually new: not much. Eq. (1) is exactly the standard asymmetric factorization of a k×k convolution into 1×k and k×1. Calling it “Unidirectional Convolution” is a rename. SCA and GCM are also close to existing attention and pooling-replacement ideas. The paper does honestly show that removing modules can improve MPII error (Table II), which it attributes to overfitting; that’s the kind of detail a less careful paper would hide.\n\nThe soft spots are the usual ones for this type of submission, but they matter because the headline claim depends on them. Section III-A gives almost no evaluation protocol: no input resolution, no face-crop/normalization, no train/val/test split, no optimizer or schedule, no batch size or epochs, no repeated runs or error bars. The margins over SUGE are 0.11° on MPII and 0.18° on Gaze360 — the same scale as run-to-run variance. The MPII ablation anomaly weakens the claim further: the full model is worse there than the stripped “-ALL” model. Without code or a detailed protocol, Table I is just a table, not evidence.\n\nThat said, the architecture is coherent and the writing is reasonably clear. The parameter/FLOP reductions are large and likely real, since factorized convolutions and lightweight attention do reduce cost. The RT-Gene numbers, where the model is competitive with MobileNetV3 and better than most baselines, are interesting even if the SUGE comparison is fragile.\n\nWho is this for? Someone working on edge gaze estimation might find the specific configuration useful as a baseline. The paper deserves a serious referee only if the authors provide code and full training details; as submitted, it is incomplete. I would send it to review with a request for a code release and an expanded protocol section, and I would not cite the numbers until they are reproducible.\n\nRecommendation: conditional acceptance path — invite revision with code and protocol, or desk reject in its current form. I lean toward engaging rather than rejecting outright, because the architecture is simple and the cost savings are plausible.","headline":"Incremental lightweight gaze network with a plausible but under-specified accuracy claim; worth a careful referee only if the authors release code and protocol.","tokens_in":7987,"tokens_out":1666,"would_cite":false,"duration_ms":15366,"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":"A 2.8M-parameter gaze estimator with factorized convolutions reports lower angular error than SUGE on two benchmarks","keywords":["gaze estimation","lightweight network","unidirectional convolution","attention mechanism","multi-task regression","MobileNetV3","angular error","model compression"],"falsifier":"Reproduce both models under one fixed protocol with identical ETH-XGaze pretraining, input resolution, face crop, training schedule, and test splits, and check whether Multitask-Gaze still shows lower mean angular error than SUGE on Gaze360 and MPIIFaceGaze; if the gap vanishes, the central accuracy claim collapses.","tokens_in":6942,"feed_emoji":"👀","tokens_out":12082,"duration_ms":96625,"temperature":0.7,"pith_summary":"The paper tries to show that a lightweight gaze-estimation network can avoid the usual accuracy penalty of small feature-channel counts. Its Multitask-Gaze model, built on MobileNetV3 bneck blocks, replaces square convolutions with Unidirectional Convolution (UC), adds Spatial and Channel Attention (SCA), swaps global pooling for a Global Convolution Module (GCM), and couples yaw/pitch regression in a Multi-task Regression Module (MRM). The reported result is lower mean angular error than the SUGE baseline on Gaze360 (10.33° vs 10.51°) and MPIIFaceGaze (3.90° vs 4.01°) at 2.822M parameters and 0.243G FLOPs. If this holds, accurate gaze tracking becomes practical on devices that cannot carry the 11M-parameter, 1.83G-FLOP heavier models.","feed_headline":"Gaze model cuts 75% of parameters and still beats heavier nets","feed_subtitle":"A 2.8M-parameter model posts lower angular error than SUGE on two gaze benchmarks.","key_machinery":"The load-bearing object is Unidirectional Convolution (UC): replacing a square convolution kernel $K_H\\times K_W$ with two perpendicular one-dimensional convolutions, $1\\times K_W$ followed by $K_H\\times 1$. This factorization cuts parameters and FLOPs from $K_H K_W C_{in} C_{out}$ to $(K_H+K_W)C_{in} C_{out}$ per output position, while stacking such layers produces a receptive field comparable to or larger than the original square kernel. Around UC, the paper adds Spatial and Channel Attention (SCA) to reweight gaze-relevant features, a Global Convolution Module (GCM) that replaces pooling with convolution-based global fusion, and a Multi-task Regression Module (MRM) whose three MLP branches predict yaw, pitch, and the joint pair, fused with weights $A_1=B_1=0.5$.","core_discovery":"The paper's central claim is that the small-channel bottleneck of lightweight backbones can be addressed architecturally rather than by scaling up the network. In Multitask-Gaze, each MobileNetV3 bneck block replaces its depthwise or standard convolution with UC, a factorization of a $K\\times K$ kernel into a $1\\times K$ followed by a $K\\times 1$ convolution, so the receptive field grows without a proportional rise in parameters or FLOPs; SCA modules inserted after the 3rd, 6th, and 9th bneck blocks reweight spatial and channel features; GCM performs global fusion without pooling loss; and MRM predicts yaw and pitch both separately and jointly, combining the three MLP outputs with equal weights. Evaluated after ETH-XGaze pretraining, the model reports angular errors of 10.33° on Gaze360, 3.90° on MPIIFaceGaze, and 6.35° on RT-Gene, with a receptive-field visualization showing that UC stacks match or exceed the field of standard $5\\times 5$ or $7\\times 7$ convolutions at lower cost. The paper takes these numbers as evidence that the four modules jointly recover the performance that lightweight networks normally lose.","pith_inferences":["If the mechanism generalizes, the UC+SCA+GCM recipe should transfer to other lightweight backbones; a cross-backbone study on MobileNetV2-style blocks would test that directly.","Because no variance or repeated-run statistics are reported, the gap between Multitask-Gaze and SUGE could be within run-to-run noise; re-running with multiple seeds would establish the systematic margin.","Editorial note: the abstract's improvement percentages (1.71% for MPIIFaceGaze, 2.75% for Gaze360) are swapped relative to Table I, so readers should rely on the table's angular errors when citing the gains.","The plug-and-play framing implies GCM and SCA could be dropped into non-gaze regression heads; testing them on head-pose or depth regression would clarify which component carries the improvement."],"forward_implications":["Deploying Multitask-Gaze uses 75.5% fewer parameters and 86.88% fewer FLOPs than SUGE while keeping or improving angular error, so embedded and mobile gaze trackers can use the same accuracy at lower cost.","Replacing global average pooling with GCM lowers angular error on all three datasets (for example 10.33° vs 10.73° on Gaze360), indicating that pooling-free global fusion is a workable head design.","Removing MRM raises angular error on Gaze360 and RT-Gene but slightly lowers it on MPIIFaceGaze, so the multi-task yaw–pitch coupling helps most on the larger benchmark sets.","A 3-layer $[7\\times 1; 1\\times 7]$ UC stack can replace a 4-layer $5\\times 5$ standard stack with 71.25% fewer parameters and 19.54% fewer FLOPs, a recipe reusable in other bneck-style blocks."],"supporting_citations":[{"why":"Supplies the MobileNetV3 bneck backbone whose convolutional blocks UC modifies.","marker":"[13]"},{"why":"Defines the SUGE baseline whose angular error, parameters, and FLOPs set the comparison.","marker":"[26]"},{"why":"ETH-XGaze is the pretraining dataset used before evaluation on the three benchmarks.","marker":"[15]"},{"why":"Gaze360 provides one benchmark dataset and the Gaze360 baseline angular errors in Table I.","marker":"[16]"},{"why":"MPIIFaceGaze provides the second benchmark dataset and its baseline angular errors.","marker":"[17]"},{"why":"RT-Gene provides the third benchmark dataset and its baseline angular errors.","marker":"[18]"},{"why":"Supplies the shifted-window mechanism that SCA uses for global spatial information exchange.","marker":"[14]"},{"why":"Supplies the visualization method used to compare UC receptive fields with standard convolutions.","marker":"[27]"}],"fun_headline_variants":["Unidirectional convolution slashes gaze model size 75% and beats SOTA","Gaze model with 75% fewer params outperforms heavier nets","Lightweight gaze net beats state-of-the-art with 75% fewer params","Efficient gaze model: 75% less params, better accuracy than SUGE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that SUGE and the other baselines in Table I were evaluated under the same protocol (same data splits, input resolution, preprocessing, and pretraining schedule); Section III-A gives only the pretraining dataset, with no code or hyperparameters, and no repeated-run variance is reported.","fun_headline_variants_meta":{"raw":{"variants":["Unidirectional convolution slashes gaze model size 75% and beats SOTA","Gaze model with 75% fewer params outperforms heavier nets","Lightweight gaze net beats state-of-the-art with 75% fewer params","Efficient gaze model: 75% less params, better accuracy than SUGE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000746,"raw_usage":{"total_tokens":3372,"prompt_tokens":1041,"completion_tokens":2331,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":657,"completion_tokens_details":{"reasoning_tokens":2249}},"tokens_in":657,"tokens_out":2331,"duration_ms":16070,"temperature":1.0,"reasoning_tokens":2249,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:32:23.943159+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce both models under one fixed protocol with identical ETH-XGaze pretraining, input resolution, face crop, training schedule, and test splits, and check whether Multitask-Gaze still shows lower mean angular error than SUGE on Gaze360 and MPIIFaceGaze; if the gap vanishes, the central accuracy claim collapses.","supporting_citations":[{"cited_title":"Suppressing uncertainty in gaze estimation,","cited_arxiv_id":null,"evidence_quote":"Defines the SUGE baseline whose angular error, parameters, and FLOPs set the comparison."},{"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 the pretraining dataset used before evaluation on the three benchmarks."},{"cited_title":"It’s written all over your face: Full-face appearance-based gaze estimation,","cited_arxiv_id":null,"evidence_quote":"MPIIFaceGaze provides the second benchmark dataset and its baseline angular errors."},{"cited_title":"Rt-gene: Real-time eye gaze estimation in natural environments,","cited_arxiv_id":null,"evidence_quote":"RT-Gene provides the third benchmark dataset and its baseline angular errors."}],"review_version":1}