{"id":"0199a111-019e-4091-8b94-d62bbf4418ee","arxiv_id":"2412.13168","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A learnable wavelet-lifting framework called IFDD separates emotion-related facial motion from static context and improves in-the-wild dynamic facial expression recognition.","lead":"This paper proposes a video model that separates emotion-related face movements from unchanging background context using a learnable version of the wavelet lifting scheme. The method improves accuracy on three real-world dynamic facial expression datasets, rivaling or beating prior supervised models.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The global context loss in Eq. 9 constrains only YS to match X's local average; it imposes no complementarity or independence penalty on YD, so the claimed implicit disentanglement is not enforced or measured, only inferred from downstream accuracy.","rationale":"The reader identified Eq. 9 as the load-bearing assumption, and my analysis agrees: the loss constrains YS to approximate local averages of X but never constrains YD to be free of global context. This is a precise gap between the paper's stated claim ('implicitly disentangles emotion-related dynamic information from emotion-irrelevant global context') and what the training objective provably enforces. The concern is not that the empirical accuracy results are wrong; the ablations in Tables 1-4 and the comparisons in Tables 5-7 are internally consistent and show real gains over baselines. The issue is that those gains are also compatible with a model that simply learns better features without achieving the claimed decomposition. Because the central claim is about mechanism, not only about accuracy, the paper should either (a) add a quantitative separation metric showing that YD is more invariant to identity/background than baseline features, or (b) ablate the subtraction step to show it is responsible for the gain. The reader's CONDITIONAL verdict already captures the need for such evidence, so my recommendation is to leave the verdict unchanged while sharpening the required condition: demonstrate that the disentanglement mechanism, not just the extra parameters or attention-based refinement, drives the improvement.","tokens_in":18318,"tokens_out":3903,"duration_ms":41182,"concrete_test":"Run a zero-predictor ablation: in Eq. 5 set P(YS|XD)=0 during both training and inference, so YD = XD, while keeping ISSM, the updater U, the LLift loss, and all other hyperparameters identical. Compare DFEW WAR against the reported 73.82. In parallel, run a zero-updater variant (YS = XS) to isolate U's contribution. If removing the subtraction leaves WAR essentially unchanged, the predictor/disentanglement step is not the source of the gains. As a complementary check, train a linear probe on YD and on baseline MViT features to predict a synthetic nuisance (e.g., injected background noise or pose perturbation) on held-out clips; if the nuisance remains as decodable from YD as from the baseline, then YD is not purified in the claimed sense.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that IFDD disentangles emotion-related dynamics YD from emotion-irrelevant global context YS. The mechanism rests on Eq. 5, YD = XD - P(YS|XD), and on the global context loss in Eq. 9. The objective does not actually enforce separation. LLift = Huber(Avg_{dim=T,C}(YS - PT_down2(X))) only asks YS to reproduce a temporally pooled local average of X; this is a low-frequency copy constraint on the static branch. It contains no term penalizing emotion-irrelevant, identity, or background content in YD and no orthogonality or independence term between YD and YS. The predictor P is trained only through LCLS(YD) plus this YS-side regularizer, so a degenerate solution is not excluded: P could output zero, leaving YD = XD, while U learns to make YS match X's local average. In that case the classification loss can still be minimized on XD and the accuracy gains would reflect feature refinement or added capacity rather than true context removal. The Grad-CAM and t-SNE evidence is qualitative and does not quantify whether YD retains identity or background informativeness. Thus the empirical margin over baselines in Table 5 does not by itself validate the specific disentanglement claim; an independent test of what YD actually removes is needed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes IFDD, a wavelet-lifting-inspired framework for dynamic facial expression recognition (DFER). IFDD consists of an Inter-frame Static-dynamic Splitting Module (ISSM), which adaptively splits latent frame features into static and dynamic groups, and a Lifting-based Aggregation-Disentanglement Module (LADM), which refines these groups with cross-attention updaters and predictors. The final dynamic features are used for classification, and a global context loss encourages the static branch to match a temporally pooled local average of the input. The authors evaluate IFDD with MobileNetV2 and MViT-S backbones on DFEW, FERV39k, and MAFW, reporting consistent gains over prior supervised methods, with per-emotion tables, ablations, and efficiency comparisons included.","tokens_in":18582,"tokens_out":3415,"duration_ms":33938,"significance":"If the central claim is correct, the paper offers a genuinely useful architectural pattern: a lightweight, backbone-agnostic module that removes emotion-irrelevant global context while keeping emotion-related dynamics, at modest computational overhead. The strengths are the breadth of experiments (three in-the-wild datasets, two backbones, many ablations, per-emotion accuracy, efficiency tables), the clear writing of the framework, the release of code, and the fair decision to compare only against supervised methods in the main tables. The main weakness is that the specific disentanglement claim is not directly tested; the evidence is indirect, consisting of downstream classification accuracy and qualitative visualizations. Because the accuracy gains could in principle arise from added capacity or feature refinement rather than from principled context removal, the central conceptual claim needs additional quantitative verification before the paper can be fully accepted.","major_comments":[{"comment":"The objective does not actually enforce disentanglement between YD and YS. LLift only requires that YS has the same local average as X along the temporal and channel dimensions; there is no term that penalizes emotion-irrelevant, identity, or background content in YD, and no orthogonality or independence penalty between YD and YS. A degenerate solution is not excluded: the predictor P could output zero, leaving YD = XD, while the updater U learns to make YS match the pooled average of X. In that case the same losses would be minimized as long as classification on XD performs well, and the reported accuracy gains would not validate the specific claim of implicit disentanglement. I ask for an independent, quantitative test of what YD actually removes, for example by training linear probes for identity or neutral/background content on YD, by measuring mutual information or distance correlation between YD and YS, or by using a synthetic setting where contamination is known and the model's ability to remove it can be directly verified. The Grad-CAM and t-SNE analyses in Section 'Visualization Analysis' are qualitative and do not resolve this issue.","section":"Decoupling Loss, Eq. (9)"},{"comment":"For the 3DViT backbone, the ISSM-only variant achieves UAR 55.53, which is lower than the 3DViT baseline's UAR of 57.00—a drop of 1.47 percentage points—even though WAR improves from 68.60 to 71.68. Since UAR is one of the two primary metrics and the text states that 'ISSM and LADM both show remarkable effectiveness in recognition accuracy,' the claim is only partially supported. The interaction between ISSM and the 3DViT backbone needs explanation, or the claim should be softened to reflect that ISSM alone is not consistently beneficial across backbones and metrics.","section":"Ablation Study, Table 1"},{"comment":"There is a direct contradiction about how the DFEW and MAFW numbers are obtained. The main text says 'As for DFER and MAFW with 5-fold cross-validation, we report metrics from a single run,' while Appendix F says that the final results are the average of 5-fold results and Table 13 lists per-fold numbers. If the main results are from a single run, then all comparisons on DFEW and MAFW have no error bars, which weakens the claimed margins over prior methods. If they are 5-fold averages, the text should say so and should provide fold-wise standard deviations or at least the variance across folds for the main tables. Please clarify and report run-to-run variability for the headline numbers.","section":"Implementation Details; Appendix F"}],"minor_comments":[{"comment":"The formula for XD[i] appears to use the index variable IF in some places and IS in others; this is likely a typo and should be made consistent with the definition of ID.","section":"Eq. (4)"},{"comment":"The phrase 'As for DFER and MAFW' should read 'As for DFEW and MAFW'; also the text near 'which all include neural emotion' should say 'neutral emotion.'","section":"Experiments, Datasets and Metrics"},{"comment":"The sentence 'We further visualize the gradient attention of YF and YS' uses YF where YD is intended; please correct the symbol.","section":"Visualization Analysis"},{"comment":"The Freq-HD row in Table 5 reports per-emotion numbers such as 6.9 and 64.61 that are inconsistent with the stated UAR/WAR and with the other rows; please verify the values and formatting.","section":"Table 5, Freq-HD row"},{"comment":"The phrase 'without exploit operations and external guidance' should likely be 'without explicit operations and external guidance'; the grammar should be corrected throughout.","section":"Abstract and Introduction"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical contribution to DFER and fits the journal's scope. The main risk is that the central 'implicit disentanglement' claim is not directly verified; the proposed addition of quantitative content-removal or independence metrics would substantially strengthen the paper. The author's comparison only against supervised methods is fair and appropriately defended. I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi,\n\nRead the IFDD paper. The bottom line: the empirical story is solid, and the accuracy improvements are the real thing. IFDD-3DViT beats the best supervised prior by about 4.6 WAR on DFEW, and the gains on FERV39k and MAFW are even larger. The architecture is a sensible extension of the wavelet lifting scheme: adaptive splitting via temporal correlation instead of fixed even-odd, and mutual cross-attention in the updater and predictor. That is a genuine step beyond DAWN/LGLFormer-style lifting, and the integration with video is not trivial. The paper also does a lot of things right: code is out, ablations are extensive (including per-fold results in the appendix), efficiency numbers are given, and the comparison with self-supervised methods is honest about being surpassed by MAE-DFER.\n\nThe soft spots are real but not fatal. The biggest one is the disentanglement claim itself. The global context loss in Eq. 9 only constrains YS to match the temporally pooled local average of X. There is no complementarity or independence penalty on YD, and nothing stops a degenerate solution where the predictor P outputs zero, YD = XD, and the classification loss is still minimized on XD. So the accuracy gains could in principle come from added refinement or capacity rather than from actual context removal. The paper infers disentanglement from downstream accuracy and from Grad-CAM/t-SNE, which are suggestive but not quantitative about what YD actually keeps or removes. I would want to see a direct test, like training an auxiliary classifier on YD for identity or background attributes, or a controlled experiment where YS is forced to contain the non-emotion content and YD is evaluated under perturbations of that content.\n\nTwo smaller things. The ablation in Table 1 shows ISSM-only for the 3DViT backbone drops UAR from 57.00 to 55.53 while improving WAR; the paper should explain why the coarse splitting hurts UAR in that setting. Also, the main text says results for DFEW and MAFW come from a single run, but the appendix says the final numbers are averages across five folds. That is presumably a per-fold single run averaged over folds, which is standard, but the wording should be cleaned up.\n\nWho is this for? The DFER community, and anyone doing dynamic video recognition with lifting-style decomposition. It deserves a serious referee: the empirical claim is strong, the code is available, and the architecture is novel enough to warrant discussion. My recommendation is to send it to review, with a request that the authors either soften the disentanglement wording or add an experiment that directly probes what YD retains. The accuracy claim stands regardless.","headline":"Real and sizable accuracy gains on three DFER benchmarks, but the 'implicit disentanglement' claim is more interpretation than demonstrated property.","tokens_in":19104,"tokens_out":3708,"would_cite":true,"duration_ms":35259,"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":"The paper proposes IFDD, a fully learnable wavelet-lifting framework that implicitly separates emotion-related facial dynamics from emotion-irrelevant global context in video, and reports that it outperforms prior supervised dynamic…","keywords":["dynamic facial expression recognition","wavelet lifting scheme","implicit disentanglement","spatiotemporal feature splitting","global context","in-the-wild video","cross-attention","vision transformer"],"falsifier":"Synthesize a benchmark in which the same identity, lighting, and background are held fixed while only the labeled expression changes, and train IFDD on it; then check whether the dynamic branch $Y_D$ is the only branch whose linear probe predicts the label. If the static branch $Y_S$ also recovers the emotion label, or if a label-shuffled version of the data still allows $Y_D$ to perform well, the disentanglement claim would be refuted.","tokens_in":18098,"feed_emoji":"😊","tokens_out":5963,"duration_ms":52018,"temperature":0.7,"pith_summary":"In-the-wild expression videos are mostly neutral or noisy frames, so emotion-relevant motion is diluted by global context and identity-bound appearance. This paper claims that the right way to handle that dilution is not to hand-design a motion detector—differencing frames, frequency analysis, optical flow—but to let a network learn to split video features into a static global-context group and a dynamic group, then refine the split with a learnable version of the wavelet lifting scheme. The result, IFDD, is claimed to outperform previous supervised methods on DFEW, FERV39k, and MAFW, with the largest reported gain on DFEW (73.82% WAR versus 69.25% for the best prior supervised model) and only modest added compute. If right, this gives a recipe, needing no external guidance or explicit motion priors, for getting cleaner expression representations out of raw video.","feed_headline":"New decoupling scheme beats prior supervised expression-video models","feed_subtitle":"Separating static context from emotion dynamics lifts weighted accuracy to 73.8% on DFEW and 51.1% on FERV39k.","key_machinery":"The load-bearing object is the wavelet lifting scheme (split, predict, update), expanded into a fully learnable, two-stage module. Standard lifting splits a signal into even and odd samples before predicting and updating; IFDD replaces the fixed split with ISSM, which uses softmax attention over temporal tokens to generate content-aware interpolated indices, and replaces the fixed predictor and updater with cross-attention MLP blocks in LADM, where the updater aggregates both groups and the predictor strips global context from the dynamic group. The resulting division is $Y_S$ for low-frequency global context and $Y_D$ for the residual high-frequency emotion dynamics, and the classification loss is applied only to $Y_D$.","core_discovery":"The central claim is that emotion-related dynamics can be disentangled from global context implicitly, by treating the video feature tensor as a signal and applying an expanded lifting scheme. ISSM learns content-aware temporal splitting indices from inter-frame correlation and interpolates the frame features into two halves: a relatively static group and a relatively dynamic group. LADM then runs an updater that aggregates both groups into a refined global-context estimate $Y_S$, and a predictor that subtracts global-context content from the dynamic group to yield purified emotion dynamics $Y_D$; a global-context loss forces $Y_S$ to keep the same local average as the input along temporal and channel dimensions, so the static branch is pinned to the global appearance. The classification head is applied only to $Y_D$. On DFEW, FERV39k, and MAFW the paper reports that IFDD variants surpass prior supervised DFER methods, and ablations tie the gains to both modules and to the global-context loss.","pith_inferences":["If the disentanglement generalizes, the same lifting recipe could apply to other dynamic tasks the authors mention—micro-expression recognition, optical flow estimation, and video compression—where separating static context from temporal residual is the core difficulty.","The implicit-disentanglement claim suggests a testable signature: the static branch $Y_S$ should carry identity and background information while being nearly label-blind, and the dynamic branch should be the opposite; a reader could verify this by training linear probes on each branch.","Because the splitting indices are themselves learned from temporal correlation, the method could in principle adapt to variable-length or irregularly sampled video, a setting the current fixed-length-clip experiments do not cover.","The comparison excludes self-supervised methods pretrained on large external data; the paper notes MAE-DFER still exceeds IFDD-3DViT on DFEW, so the practical claim is specifically about supervised training without external sources."],"forward_implications":["On the three in-the-wild benchmarks, IFDD-3DViT reports WAR of 73.82% on DFEW, 51.09% on FERV39k, and 53.92% on MAFW, beating the best prior supervised methods compared in the paper.","Ablations show the two stages are complementary: ISSM alone and LADM alone each improve over the backbone, and using both gives the largest gain.","The global-context loss is load-bearing: removing it drops UAR by 2.63 percentage points on DFEW, and constraining local averages along $\\{T,C\\}$ performs better than constraining all four dimensions.","IFDD is backbone-agnostic: the same modules attach to a 2D CNN (MobileNetV2) and a 3D ViT (MViT-S), with modest extra FLOPs of about 3.5% and 9.7%.","Because classification uses only $Y_D$, the framework predicts from residual emotion dynamics rather than from the full coupled spatiotemporal representation."],"supporting_citations":[{"why":"Supplies the lifting scheme (split-predict-update) that IFDD expands into a fully learnable framework.","marker":"(Sweldens 1998)"},{"why":"Introduces the deep adaptive wavelet network whose local-average idea motivates the global-context loss on $Y_S$.","marker":"(Bastidas Rodriguez et al. 2020)"},{"why":"Provides the DFEW benchmark and evaluation protocol on which the main ablations and comparisons are run.","marker":"(Jiang et al. 2020)"},{"why":"Provides the FERV39k dataset and the temporal 2D-CNN baselines used in comparison.","marker":"(Wang et al. 2022a)"},{"why":"Provides the MAFW dataset and the T-ESFL baseline IFDD claims to beat by the largest margin.","marker":"(Liu et al. 2022)"},{"why":"M3DFEL is the strongest prior supervised method on DFEW and FERV39k that IFDD is compared against.","marker":"(Wang et al. 2023)"},{"why":"NR-DFERNet is the explicit inter-frame-difference method that motivates the implicit alternative.","marker":"(Li et al. 2022)"},{"why":"Freq-HD is the explicit frequency-analysis method that IFDD contrasts with.","marker":"(Tao et al. 2023)"},{"why":"Supplies the self-attention and cross-attention machinery used in the updater and predictor.","marker":"(Vaswani et al. 2017)"}],"fun_headline_variants":["Implicit disentanglement lifts expression recognition in the wild","Learnable lifting scheme separates emotion dynamics from static context","IFDD: implicit disentanglement beats prior DFER methods","Lifting-scheme model untangles emotion dynamics in the wild","Emotion dynamics disentangled from context without explicit cues"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole argument hinges on the global-context loss: forcing the static branch $Y_S$ to match the local temporal-and-channel average of the input is what supposedly makes it absorb identity, background, and other emotion-irrelevant context, leaving $Y_D$ as purified emotion dynamics; if that constraint instead discards useful emotion cues or fails to absorb irrelevant context, the reported accuracy gains do not establish genuine disentanglement.","fun_headline_variants_meta":{"raw":{"variants":["Implicit disentanglement lifts expression recognition in the wild","Learnable lifting scheme separates emotion dynamics from static context","IFDD: implicit disentanglement beats prior DFER methods","Lifting-scheme model untangles emotion dynamics in the wild","Emotion dynamics disentangled from context without explicit cues"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000573,"raw_usage":{"total_tokens":2762,"prompt_tokens":1052,"completion_tokens":1710,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":1630}},"tokens_in":668,"tokens_out":1710,"duration_ms":12639,"temperature":1.0,"reasoning_tokens":1630,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:21:17.821480+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Synthesize a benchmark in which the same identity, lighting, and background are held fixed while only the labeled expression changes, and train IFDD on it; then check whether the dynamic branch $Y_D$ is the only branch whose linear probe predicts the label. If the static branch $Y_S$ also recovers the emotion label, or if a label-shuffled version of the data still allows $Y_D$ to perform well, the disentanglement claim would be refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the lifting scheme (split-predict-update) that IFDD expands into a fully learnable framework."},{"cited_title":"X.; Gruson, A.; Polania, L","cited_arxiv_id":null,"evidence_quote":"Introduces the deep adaptive wavelet network whose local-average idea motivates the global-context loss on $Y_S$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the DFEW benchmark and evaluation protocol on which the main ablations and comparisons are run."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MAFW dataset and the T-ESFL baseline IFDD claims to beat by the largest margin."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"M3DFEL is the strongest prior supervised method on DFEW and FERV39k that IFDD is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Freq-HD is the explicit frequency-analysis method that IFDD contrasts with."},{"cited_title":"N.; Kaiser, .; and Polosukhin, I","cited_arxiv_id":null,"evidence_quote":"Supplies the self-attention and cross-attention machinery used in the updater and predictor."}],"review_version":1}