acceptodds
Under review as a conference paper at ICLR 2027

M2E: Lightweight Motion-to-Emotion Inference for Edge-Oriented Micro-Expression Recognition

Abstract

Micro-expression recognition (MER) requires capturing subtle facial movements with low computational cost and providing interpretable decision evidence for edge deployment, yet existing deep models typically learn latent representations from facial images, incurring substantial resource overhead and offering limited process-level interpretability. In this paper, we propose M2E, a lightweight MER framework that encodes compensated facial dynamics as a compact, semantically explicit 12-dimensional regional representation. M2E suppresses head-motion interference, transforms compensated onset–apex optical flow into FACS-guided regional features, and employs an RBF-SVM with class weighting for efficient and interpretable classification. Under the MEGC2019 Composite Database Evaluation protocol, the lightweight anchor-based configuration achieves 76.94% UF1 and 79.99% UAR on the 3DB dataset, outperforming LAENet, the most lightweight existing baseline, by 1.26 and 5.94 percentage points, respectively, with lower computational overhead. When equipped with MEFlowNet, a more computationally intensive learning-based optical-flow compensation network, M2E further achieves 79.80% UF1 and 83.20% UAR on the 3DB dataset and matches or surpasses high-capacity expert models on several benchmark metrics. Raspberry Pi experiments, ablation studies, and SHAP analyses further demonstrate its edge applicability, component effectiveness, and process-level interpretability, while evaluations on MMEW and CK+ show its cross-dataset and cross-domain generalization. Code and data are available at https://anonymous.4open.science/r/M2E-F27.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
micro-expression recognition, optical flow, regional motion representation, edge deployment, interpretability

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_1BJEOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents M2E, which encodes compensated onset–apex optical flow into 12-dimensional regional motion features and uses a class-weighted RBF-SVM for micro-expression recognition. The authors evaluated benchmark performance, component ablations, cross-dataset transfer, SHAP explanations, and Raspberry Pi deployment.

Critical strengths
  1. The method addresses small-sample recognition by reducing dense facial motion to a compact set of regional measurements. Head-motion compensation and sample-wise normalization address nuisance variation before classification, reducing what the classifier needs to learn from limited data.
  2. The 12-dimensional encoding combines regional vertical motion with left–right horizontal differences and inner–outer eyebrow differences. This is a compact way to retain both local movement and relationships between regions without AU labels. It also gives the SHAP analysis concrete facial-motion variables to explain.
  3. The anchor-based version reports 76.94% UF1 and 79.99% UAR on Composite, exceeding the lightweight baselines listed in Table 1. The authors also test it on a Raspberry Pi and examine performance without CK+ augmentation. Together, these results support the practical value of a simple representation and classifier, even though the relative efficiency needs a fairer comparison.
Critical weaknesses
  1. M2E assumes the onset and apex frames are known and thus calculate the optical flow between the onset and apex. However, detecting the onset, apex, and offset is very challenging for MER/FER. This has been demonstrated from the experimental results on the SMIC, which is the only dataset without groundtruth apex labeling. The performance of the propose method is inferior to FRL-DGT.
  2. In Figure 1, TV-L1 and MEFlowNet can be mistaken for sequential steps, although they are alternative configurations. The regional abbreviations need a legend, and “Magnitude weighted” and “Direction mapping” do not explain how the 12 features are constructed. The symbols should also match the equations.
  3. In Section 3.4, the bilateral horizontal difference is described as capturing asymmetry. However, symmetric motion (−1, 1) and unilateral motion (−2, 0) both give −2. The feature cannot distinguish these patterns, so this interpretation should be revised.
  4. The claim that Equations (9)–(11) suppress isolated high-magnitude noise needs qualification. As a counterexample, 99 aligned flow values of 1 and one outlier of 1000 yield a response of approximately 10, compared with 1 without the outlier. This does not establish a failure on the benchmark, but shows that the aggregation is not generally robust to large outliers.
  5. Table 6 compares variants of the proposed features but does not include MDMO, a closely related regional flow descriptor. Table 3 tests arithmetic averaging, but this does not replace a comparison with an existing descriptor. A matched comparison using the same flow, compensation, normalization, and classifier is needed to establish the benefit of the proposed encoding.
  6. Tables 1 and 2 use different front ends, but Raspberry Pi measurements cover only the anchor-based version. The runtime comparison with LAENet in Appendix E.9 does not provide matched hardware measurements and includes training-data preparation. Table 12 reports memory increases rather than total peak memory. These measurements do not establish the full pipeline’s efficiency advantage.
Questions for authors
  1. How does the 12-dimensional representation compare with MDMO under identical preprocessing, training data, and classification settings?
  2. Can you clarify or revise the asymmetry and noise-suppression claims discussed above?
  3. Please report complete inference latency and total peak memory for both configurations and LAENet on the same device.
  4. Please specify the ROI expansion parameter, eye-closure threshold, regional landmark indices, and the conversion to the upward-positive coordinate convention.
8 comments
AuthorsOpenReview ↗

Reply to Reviewer 1BJE [Part 1/2]

We thank the reviewer for the detailed and technical questions, which helped us tighten several parts of the paper. Revised text is in orange, and the new measurements are in Appendices E.9–E.10.

W1 Dependence on onset and apex localization

M2E recognizes a micro-expression from an annotated or estimated onset–apex pair, which is also the setting of FRL-DGT (Zhai et al., CVPR 2023) and SelfME [R1]. We use official apex labels where they exist and estimate the SMIC apex with D&C-ROIs (Appendix D.3). The SMIC comparison does not isolate the effect of apex estimation, because the two methods also differ in motion representation, training objective, and model capacity.

W2 Figure 1

The two front ends are alternatives, and the revised Figure 1 draws them as parallel branches. MEFlowNet works on the onset and apex frames directly and never sees TV-L1 flow. The figure now outlines the ROIs on the face. The caption spells out the region abbreviations and defines μx1 / μx2 and μy1 / μy2 as the positive and negative parts of μx(R) and μy(R) in Eq. (11). Section 3.4 and Table 5 show how the regional responses become the 12 features.

W3 / Q2 Bilateral horizontal difference

Section 3.4 now calls this feature "the relative horizontal motion of the two sides" instead of "asymmetry". Keeping the left and right components separately, as a 15-dimensional variant, does not improve results consistently across the three datasets (revised Table 8, originally Table 6).

W4 / Q2 Outlier robustness of the aggregation

We removed the claim about isolated noise from Section 3.4. The text now only says that the weighting in Eqs. (9)–(11) gives more weight to strong motion vectors.

W5 / Q1 Comparison with MDMO

With the same flow, compensation, classifier, splits, and training pool, M2E outperforms MDMO (Liu et al., IEEE TAFFC 2016) on every dataset, with and without CK+. MDMO keeps its own 36 regions and its own magnitude normalization, which is part of how its features are built, and M2E keeps max-abs.

Table R1. Matched-input comparison (UF1 / UAR, %).

Training poolSubsetMDMOM2EΔ (pp)
MER onlyCASME II84.27 / 86.1792.48 / 93.94+8.21 / +7.77
MER onlySAMM64.72 / 71.6170.03 / 76.03+5.31 / +4.42
MER onlySMIC61.10 / 62.4969.16 / 70.30+8.06 / +7.81
MER only3DB70.57 / 73.1676.89 / 79.57+6.32 / +6.41
MER + CK+CASME II81.35 / 82.6691.16 / 93.18+9.81 / +10.52
MER + CK+SAMM62.19 / 66.5071.00 / 77.89+8.81 / +11.39
MER + CK+SMIC61.95 / 62.4369.19 / 70.71+7.24 / +8.28
MER + CK+3DB68.69 / 70.1176.94 / 79.99+8.25 / +9.88

M2E is also higher in the source-only setting, by 8.73 / 7.76 points on MMEW and 28.09 / 24.07 on CK+. A subject-level bootstrap on 3DB gives 95% intervals for the difference that all exclude zero. Details are in Appendix E.9 (Tables 14–16).

AuthorsOpenReview ↗

Reply to Reviewer 1BJE [Part 2/2]

W6 / Q3 Latency and memory

On the same desktop, the complete anchor-based M2E pipeline, from the given onset and apex frames to the predicted label, takes less time per sample than LAENet preprocessing alone on the three MER datasets. Feature construction, timed from reading the two frames and including face cropping, takes 10.00–35.00% less time than LAENet preprocessing (Table R2). The SVM stage adds almost nothing. Training and testing over all LOSOCV folds of 3DB take 2.05 s in total, which is under 5 ms per test sample even with training included (Table R4). M2E anchor and LAENet run on the CPU, and MEFlowNet uses the GPU.

Table R2. Per-sample time on the same desktop, including face cropping (s).

DatasetSamplesM2E anchor (CPU)LAENet (CPU)ReductionM2E MEFlowNet (GPU)
CASME II1450.520.8035.00%0.51
SAMM1330.861.3134.35%0.76
SMIC1640.450.5010.00%0.44
CK+2891.332.9655.07%0.51

Table R3 gives the mean of the per-sample peak host RSS, in addition to the memory increase in revised Table 17 (originally Table 12).

Table R3. Mean per-sample peak host RSS during feature construction or preprocessing (MB).

DatasetM2E anchorLAENetM2E MEFlowNet
CASME II140.86144.742223.17
SAMM152.82157.862225.34
SMIC134.64142.572223.33
CK+137.12140.912220.63

Table R4. Downstream learning and evaluation.

WorkloadTime (s)Host RSS (MB)Memory aggregation
M2E SVM LOSOCV (anchor)2.05124.69Whole-experiment peak
M2E SVM LOSOCV (MEFlowNet)1.89124.55Whole-experiment peak
LAENet CK+ pre-training10.812251.46Whole-pre-training peak
LAENet MER LOSOCV557.292255.28Mean of fold peaks

The anchor pipeline also runs on the Raspberry Pi (revised Table 17), and MEFlowNet was evaluated on the desktop GPU. All of these numbers are in Appendix E.10 (originally E.9).

Q4 Reproducibility details

Appendix A now specifies the ROI sizes, the eye-closure threshold, and the coordinate convention, with the FaceMesh indices in Table 6. Eyebrow boxes are expanded by δ = 4 pixels on each side. Cheeks, mouth corners, and chin use 8 × 8-pixel boxes around their landmarks. The weighting uses ε = 10⁻⁶, and the closed-eye threshold is EAR = 0.20 (Soukupova & Cech, CVWW 2016). We flip the sign of the vertical flow during aggregation so that positive means upward, and the feature names and SHAP labels follow this convention. These are the values in the released code.

We hope these additional results address your questions. We will respond promptly to any further questions during the discussion period.

References

[R1] X. Fan et al. SelfME: Self-Supervised Motion Learning for Micro-Expression Recognition. In CVPR, 2023.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer 1BJE,

Your questions led to several concrete changes, including the matched comparison with MDMO in Appendix E.9, the corrected wording in Section 3.4, the revised Figure 1, and the same-platform latency and memory measurements in Appendix E.10.

We would appreciate hearing whether these resolve the points you raised, and we can provide further details if needed.

If you find these changes satisfactory, we hope they merit a reconsideration of your assessment. Thank you very much.

Reviewer_6idFOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

Since micro-expression recognition (MER) requires capturing subtle facial movements with low computational cost and providing interpretable decision evidence for edge deployment, the paper propose a lightweight MER framework. On on hand, the proposed explicit 12-dimensional regional representation appears to be an effective design suitable for lightweight deployment at the edge. On the other hand, the experimental comparisons are relatively vague. The paper has not been compared with some of the latest work progress. In terms of pure accuracy comparison, there is a significant gap between the algorithm proposed in this paper and the SOTA model. If this paper emphasizes lightweight and high-performance computing, it does not provide a comparison of algorithm efficiency compared to the latest MER algorithm, nor does it provide actual performance verification on mainstream edge segment devices.

Critical strengths
  1. M2E replaces heavy end-to-end deep networks with 12 dimensional region semantic representation and RBF-SVM, reducing computing power and storage overhead.
  2. The regional feature semantics guided by FACS are clear, and SVM classification combined with SHAP analysis can trace the contribution of each region to decision-making. Compared with black box deep models, it is easier to provide decision evidence, which is beneficial for clinical assistance, human-computer interaction and other scenarios that require trust and debugging.
  3. The paper presents a lightweight, interpretable, edge usable, and performance balanced solution with high practical value.
Critical weaknesses
  1. The entire process of M2E is built on the basis of 'compensated onset apex optical flow'. However, optical flow estimation is already a fragile link in MER and even in a large number of computer vision tasks, so the practicality of such method design may still be limited.
  2. There are significant issues with the literature summary and experimental comparison in this article, which overlook a large amount of the latest published related work. Both research emphasizing performance and research emphasizing efficiency and lightweight are equally important, and it is also necessary to compare their latest progress and optimal performance with each other, in order to help relevant researchers better grasp the progress of the entire MER field.
  3. The gap between the accuracy of the current paper and the state-of-the-art algorithm that focuses on performance is too obvious. So, in the process of continuously optimizing the hardware performance of the edge segment, can this algorithm achieve an adaptive balance between lightweight and precision performance? 4.Some missing references, e.g., [1]DCE: Dynamic channel and spatial redundancy elimination for Micro-Expression recognition. Expert Syst. Appl. 314: 131675 (2026). [2]Temporal-spatial cross-fusion for dynamic micro expression recognition. Pattern Recognit. 179: 113715 (2026) [3]TGRF-Net: Text-Semantic Guided Region Focused Network for Micro-Expression Recognition. IEEE Trans. Affect. Comput. 17(3): 3434-3447 (2026) [4]Region-Aware Instance Consistency Learning for Micro-Expression Recognition. CVPR 2026: 34428-34438.
Questions for authors

See the weaknesses 1,2,3.

3 comments
AuthorsOpenReview ↗

Reply to Reviewer 6idF

Thank you for the constructive review. All changes are marked in orange in the revised PDF.

W1 Reliance on optical flow

Optical flow is the usual carrier of subtle facial motion in MER, and M2E builds on it. FRL-DGT (Zhai et al., CVPR 2023) and SelfME [R1] also start from onset–apex displacement or motion. What M2E adds is a compact, human-readable summary of that motion. It works with both a classical front end (TV-L1) and a learned one (MEFlowNet), and with either front end it transfers from 3DB to MMEW and CK+ without target-domain training or adaptation (revised Table 13). Appendix E.12 now discusses this dependence on the upstream steps.

W2 / W4 Recent literature

Section 2 of the revision now cites and discusses DCE (Yu & Chen, ESWA 2026), TGRF-Net (Hu et al., IEEE TAFFC 2026), Ra-ICL (Cai et al., CVPR 2026), and TSFmicro (Liu et al., Pattern Recognition 2026). All four are learned expert networks. DCE, TGRF-Net, and Ra-ICL aim at accuracy, and TSFmicro reports 9.5M parameters (Table 4 of that paper).

Our two configurations sit at the two ends of the accuracy–efficiency range. The anchor-based one is compared in Table 1 with lightweight models, LAENet (Gan et al., The Visual Computer 2024) with 843 parameters and STSTNet (Liong et al., FG 2019) with 1,670. With the CK+-augmented training pool, it reaches 76.94 / 79.99 UF1 / UAR on 3DB on the desktop and 76.01 / 78.67 on a Raspberry Pi 5, against 75.68 / 74.05 for LAENet. The MEFlowNet configuration runs on a desktop GPU, reaches 79.80 / 83.20, and matches or surpasses the high-capacity models of Table 2 on several metrics.

W3 Accuracy–efficiency balance

M2E offers two fixed configurations that share the same 12 features and the same SVM. Only the motion front end changes. The anchor-based one runs on a Raspberry Pi 5. On the desktop and without CK+, it reaches 76.89 / 79.57 on 3DB. The MEFlowNet one runs on a desktop GPU and reaches 79.80 / 83.20. The SVM cost is almost the same in both cases (2.05 s and 1.89 s for the full LOSOCV, revised Table 20).

We hope these additional results address your questions. We will respond promptly to any further questions during the discussion period.

References

[R1] X. Fan et al. SelfME: Self-Supervised Motion Learning for Micro-Expression Recognition. In CVPR, 2023.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer 6idF,

A short follow-up on our response. Section 2 of the revised paper now covers the four works you listed (DCE, TGRF-Net, Ra-ICL, and TSFmicro). Our reply also explains where the two configurations stand: the anchor-based one runs on a Raspberry Pi 5 and stays ahead of LAENet, and the MEFlowNet one is compared with the high-capacity models in Table 2. On the reliance on optical flow, the same representation transfers across datasets with both the TV-L1 and MEFlowNet front ends (Table 13).

Please let us know if any point needs further clarification.

If our responses have addressed your concerns, we would be grateful if you could consider raising your score. Thank you very much.

AuthorsOpenReview ↗

Follow-up before the discussion closes

Dear Reviewer 6idF,

As the discussion period closes in a few hours, we would be grateful for any feedback on our response and the revised paper.

Thank you very much.

Reviewer_A51dOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents M2E as an alternative to computationally expensive micro-expression recognition models. Instead of learning high-dimensional representations through deep neural networks, it uses optical flow to capture facial changes between onset and apex frames. To minimize interference from head movement, the authors investigate two compensation strategies: facial landmark-based motion correction and MEFlowNet-based estimation. The resulting motion information is summarized across predefined facial regions into 12 numerical features, which encode regional displacement and asymmetry. These features are classified using a class-weighted RBF-SVM. Experiments on SMIC, CASME II, SAMM, and the combined 3DB benchmark show that the approach is competitive with existing lightweight methods. The authors also examine feature importance using SHAP and evaluate deployment on a Raspberry Pi 5. While the explicit motion representation improves interpretability and reduces computational requirements, the approach remains dependent on reliable facial landmark localization and optical-flow estimation, which may limit its performance under unconstrained recording conditions

Critical strengths
  1. The method addresses unwanted head movement through two compensation strategies, offering a choice between a simpler approach and a more sophisticated optical flow estimation method.
  2. The authors evaluate the method on few established micro expression benchmarks and compare it with handcrafted, lightweight, and deep-learning-based approaches using metrics that account for class imbalance.
  3. The ablation studies examine how motion compensation, regional motion weighting, and feature normalization affect recognition performance, helping clarify the contribution of individual design choices.
  4. The combination of SHAP based feature analysis and experiments on Raspberry Pi 5 provides useful evidence of the model’s interpretability and its potential for deployment on resource-constrained devices
Critical weaknesses
  1. The methodological novelty mainly lies in the regional feature design and motion compensation, while several underlying components are existing techniques.
  2. The ablation results indicate that the effectiveness of certain design choices varies across datasets. Arithmetic averaging outperforms magnitude-weighted aggregation on CASME II, while L2 normalization has better results than max absolute normalization on SMIC and CASME II. Although the authors justify their choices in terms of balanced performance across datasets, these results suggest that the proposed design choices are not universally optimal.
  3. The experiments do not include a direct comparison with MDMO which is a closely related regional optical flow descriptor discussed in the paper.
  4. Using only onset and apex frames discards the full motion trajectory. This can be a significant limitation for the task defined by brief, dynamic facial movements.
  5. A small feature representation may miss useful information. 12 dimensions are efficient and interpretable but the method may miss local deformation patterns, complex interactions and subtle motion that is not captured by the selected regional statistics.
  6. The paper reports feature-construction time of approximately 1.13-5.11 seconds per sample on a Raspberry Pi 5. That indicates that the SVM may be lightweight, but it does not by itself establish real time inference.
  7. The method estimates a single global displacement from facial landmarks and subtracts it from the entire optical flow field. This approach may not adequately correct more complex head movements, such as rotation or spatially varying motion, and the paper does not report a dedicated evaluation under these conditions
Questions for authors
  1. Could the authors clarify whether the baseline methods in Tables 1 and 2 were evaluated under the same experimental protocol, including CK+ training-data augmentation, or whether their results were taken from the respective original publications?
  2. Could the authors compare M2E with MDMO under the same evaluation protocol to clarify the benefit of the proposed feature representation?
  3. Does M2E automatically detect onset and apex frames, or does it rely on pre-annotated frames?
  4. Given the relatively small size of micro-expression datasets, have the author evaluated the statistical significance and variability of their result across different subject splits or repeated experiments?
  5. Can M2E capture important micro-expression movements occurring outside its predefined facial regions? Can M2E capture important micro-expression movements occurring outside its predefined facial regions?
5 comments
AuthorsOpenReview ↗

Reply to Reviewer A51d [Part 1/2]

We thank the reviewer for the careful reading and the concrete suggestions. Changes in the revised PDF are marked in orange. The new analyses are in Section 4.3 and Appendices E.9–E.10.

W1 Novelty

Our starting point is that the evidence for a micro-expression lies in the direction of motion within a few facial regions. We remove the dominant head translation, pool the compensated onset–apex flow over FACS-guided regions, and keep 12 named variables, such as the vertical motion of the left mouth corner. The classifier can then stay small, and each decision can be read as where the face moved and in which direction. Under the matched settings in W3 / Q2, this representation outperforms MDMO on every dataset, with and without CK+.

W2 Aggregation and normalization choices

Arithmetic averaging, L2 normalization, and absolute-sum normalization each stay within two points of our settings on every dataset and metric (revised Table 4, originally Table 3). Arithmetic averaging is 0.65 / 0.38 points higher in UF1 / UAR on CASME II and lower on SMIC and SAMM. L2 and absolute-sum are slightly higher on SMIC and CASME II, and max-abs is best on SAMM. All three normalizations divide the vector by one positive number, so the signs and the ratios between dimensions do not change. We keep max-abs because it scales each nonzero sample so that its largest absolute component is 1, which makes the feature values in the SHAP plots (Figure 2) comparable across samples. The same rule is used for all datasets, and Section 4.5 now explains the choice.

W3 / Q2 Comparison with MDMO

With identical inputs, M2E outperforms MDMO on every dataset, with and without CK+ (Table R1). MDMO (Liu et al., IEEE TAFFC 2016) keeps its own 36 regions and its own magnitude normalization. Everything else is shared: onset–apex pairs, face crops, anchor compensation, RBF-SVM, LOSOCV splits, and training pool.

Table R1. Matched-input comparison (UF1 / UAR, %).

Training poolSubsetMDMOM2EΔ (pp)
MER onlyCASME II84.27 / 86.1792.48 / 93.94+8.21 / +7.77
MER onlySAMM64.72 / 71.6170.03 / 76.03+5.31 / +4.42
MER onlySMIC61.10 / 62.4969.16 / 70.30+8.06 / +7.81
MER only3DB70.57 / 73.1676.89 / 79.57+6.32 / +6.41
MER + CK+CASME II81.35 / 82.6691.16 / 93.18+9.81 / +10.52
MER + CK+SAMM62.19 / 66.5071.00 / 77.89+8.81 / +11.39
MER + CK+SMIC61.95 / 62.4369.19 / 70.71+7.24 / +8.28
MER + CK+3DB68.69 / 70.1176.94 / 79.99+8.25 / +9.88

On 3DB without CK+, the margin is 6.32 / 6.41 points in UF1 / UAR.

M2E also transfers better. Trained only on 3DB, MDMO reaches 66.90 / 70.62 on MMEW and 49.41 / 53.36 on CK+. M2E is higher by 8.73 / 7.76 and 28.09 / 24.07 points.

On 3DB, the gain over MDMO is significant under a subject-level bootstrap. We resampled the 67 subjects 10,000 times, and none of the 95% intervals for M2E − MDMO contains zero.

Table R2. Bootstrap 95% CI of M2E − MDMO on 3DB (pp).

Training poolUF1UAR
MER only[+3.04, +9.99][+2.94, +10.24]
MER + CK+[+4.94, +12.10][+6.48, +13.85]

Appendix E.9 has the full comparison (Tables 14–16).

W4 Onset and apex frames

Using only the onset and apex frames is a deliberate cost choice, and it is common in MER. FRL-DGT (Zhai et al., CVPR 2023) takes an onset–apex image pair as input, and SelfME [R1] uses the motion between the same two frames. With two frames, the edge pipeline needs a single flow estimate per sample.

W5 / Q5 Motion outside the predefined regions

M2E encodes motion inside the FACS-guided regions of the eyebrows, cheeks, mouth corners, and chin. The eyebrow boxes are expanded by 4 pixels to catch nearby displacement, and the other regions use 8 × 8-pixel boxes around their landmarks (Appendix A). Table 5 lists what each of the 12 dimensions measures. A 15-dimensional variant that keeps the left and right horizontal components separately does not improve consistently across the three datasets (revised Table 8).

W6 Raspberry Pi runtime

We do not claim real-time processing. The Raspberry Pi experiment shows that the anchor-based pipeline runs entirely on a 2 GB edge device with modest memory. The MEFlowNet configuration is a desktop GPU study of what the same 12 features achieve with a learned flow front end.

AuthorsOpenReview ↗

Reply to Reviewer A51d [Part 2/2]

W7 Rotation and spatially varying head motion

Anchor compensation removes translation, and Section 3.3 now says so explicitly. It subtracts one reference vector from the flow of the cropped face. Even this simple step matters. Without it, UF1 / UAR on SAMM drop by 8.94 / 10.90 points (revised Table 4).

Q1 Evaluation protocol of the baselines

The baseline numbers in Tables 1 and 2 come from the original papers, except for MDMO in Table 1, which is our matched-protocol implementation. For M2E, label-compatible CK+ samples are added only to the training side of each LOSOCV fold, and the held-out MER subject never enters training. The SMIC, CASME II, and SAMM scores are read off the same 3DB run. Without CK+, anchor-based M2E still reaches 76.89 / 79.57 on 3DB (revised Table 14), ahead of LAENet (Gan et al., The Visual Computer 2024) at 75.68 / 74.05 and STSTNet (Liong et al., FG 2019) at 73.53 / 76.05. LAENet also uses macro-expression data, in its case for pre-training. The captions of Tables 1 and 2 now state the training pool, and Appendix E.5 gives the per-dataset MER-only results.

Q3 Onset and apex detection

No. M2E uses the official onset and apex annotations, and for SMIC, which has no apex labels, the apex is estimated with D&C-ROIs (Appendix D.3).

Q4 Statistical significance and variability

Across the subject-deletion runs, the mean UF1 and UAR stay within 0.44 points of the full-data scores, and the standard deviations are at most 1.05 points (Table R3, Section 4.3). For significance, see the bootstrap intervals in Table R2.

Table R3. Subject-deletion analysis (UF1 / UAR, %).

DatasetConfig.Full dataDeletion meanDeletion std.
CASME IIAnchor91.16 / 93.1890.88 / 93.020.66 / 0.46
CASME IIMEFlowNet95.12 / 95.7494.80 / 95.380.68 / 0.74
SAMMAnchor71.00 / 77.8970.72 / 77.450.76 / 0.91
SAMMMEFlowNet73.90 / 81.5473.81 / 81.410.88 / 1.05
SMICAnchor69.19 / 70.7169.26 / 70.770.98 / 0.90
SMICMEFlowNet70.79 / 72.9270.68 / 72.800.91 / 0.76
3DBAnchor76.94 / 79.9976.83 / 79.880.58 / 0.53
3DBMEFlowNet79.80 / 83.2079.64 / 83.010.48 / 0.45

We hope these additional results address your questions. We will respond promptly to any further questions during the discussion period.

References

[R1] X. Fan et al. SelfME: Self-Supervised Motion Learning for Micro-Expression Recognition. In CVPR, 2023.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer A51d,

We hope you have had a chance to look at our response. Two additions may be most relevant to your review: the matched-input comparison with MDMO in Appendix E.9, where M2E is higher on every dataset with and without CK+, and the 3DB results without CK+ that you asked about in Q1.

If any part of the response is unclear, we are happy to follow up. Thank you very much.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the revision and discussion

Dear Area Chair and Reviewers,

Thank you for handling and reviewing our submission. To help with the discussion, we summarize the main additions in the revised PDF, where all changes are marked in orange.

  1. Comparison with MDMO (A51d, 1BJE). Under the same inputs and classifier, M2E outperforms MDMO on every dataset, with and without CK+. Subject-level bootstrap intervals on 3DB exclude zero (Appendix E.9).
  2. Results without CK+ (A51d). M2E stays above all handcrafted and lightweight baselines on 3DB (Table 14).
  3. Onset and apex frames (1BJE, A51d). In the discussion we showed that moving the onset or apex of CASME II by up to four frames costs less than 0.5 points in UF1 and UAR. All SMIC results already rely on the estimated apex, and Appendix E.12 states this assumption.
  4. Efficiency (1BJE, A51d). On the same desktop, anchor-based feature construction is faster than LAENet preprocessing and uses less memory (Appendix E.10).
  5. Related work and clarity (6idF, 1BJE, A51d). Section 2 now covers the four 2026 works raised by Reviewer 6idF. Sections 3.3 and 3.4, Figure 1, and Appendix A were revised for clarity.

Reviewers A51d and 1BJE have raised their scores after the discussion. We remain happy to answer any further questions.

Thank you for your time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.