acceptodds
Withdrawn from ICLR 2027

RethinkFuse: Cross-Window Evidence for Micro-Expression Event Decoding

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Long-video micro-expression analysis requires an event decoder to determine whether a local response is genuine and which temporal boundary best represents it. A single-window decoder couples candidate survival and boundary construction to one temporal scale, leaving no independent evidence for either decision. We introduce RethinkFuse, an event decoder for frozen joint spotting-recognition backbones. Its central observation is that cross-window agreement provides evidence of event support, while boundary disagreement exposes complementary localization hypotheses. RethinkFuse decodes the same score sequence with micro and macro windows, associates compatible candidates into event clusters, constructs a stronger single-window candidate and a fused candidate , and uses an Adaptive Gated Fusion router to reject the cluster, keep , or select . The router is fitted using only training subjects in each outer fold and leaves backbone parameters and frame-level predictions unchanged. Under fully nested subject-disjoint evaluation, RethinkFuse improves the spotting-recognition synergy score (STRS) by approximately 9.6% on SAMMLV and 30.1% on CAS(ME) relative to native decoding. Controlled comparisons attribute a positive STRS gain to adaptive routing on SAMMLV. On CAS(ME), the corresponding interval reaches zero at the reported precision.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
CV: Motion, Tracking & Activity Analysis, CMS: Affective Computing & Social Cognition, CV: Biometrics, Face, Gesture & Pose

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Xo9kOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

RethinkFuse improves event decoding for frozen micro-expression models through dual temporal windows, candidate association, and a logistic router choosing rejection, single boundaries, or fusion. Some experiments are conducted to prove the effectiveness.

Critical strengths

The paper addresses a practical issue in micro-expression spotting and recognition: converting frame-level predictions into reliable event detections.

Critical weaknesses
  1. Despite my research experience in micro-expression spotting and recognition, I found the paper very difficult to follow and understand.
  2. The proposed decoder is employed on the outputs of ME-TST+, but the final results are worse than ME-TST+ (authors didn't show the results of ME-TST+ in the manuscript).
Questions for authors

Please provide the explanation about the weaknesses mentioned above.

Reviewer_NW41OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes RethinkFuse, an event decoder for micro-expression spotting and recognition in long videos using predictions from a frozen backbone. It associates event candidates from two temporal windows and learns to choose among rejecting an event, retaining a single-window hypothesis, and fusing boundaries. Experiments on SAMMLV and CAS(ME)³ show improved STRS over native decoding, with stronger evidence for adaptive routing on SAMMLV.

Critical strengths
  1. Separating whether to retain an event from how to determine its boundaries addresses a practical limitation of single-window decoding and provides a reasonable motivation.
Critical weaknesses
  1. The paper does not establish whether cross-window evidence adds value beyond confidence. It reports no feature-group ablation or confidence-only router, and the authors acknowledge that Figure 4 does not establish superiority over a tuned threshold.
  2. The results offer limited evidence that the full design outperforms simpler alternatives. Its STRS advantage over Independent Always-Fuse on SAMMLV remains uncertain, and the gain over Nested Single-Window on CAS(ME)³ is small. The second-backbone results use the complete target cache for selection and fitting, so they do not establish independent generalization.
Questions for authors
  1. Section 3.3 keeps only the higher-confidence member as the single-window option, and the output in Figure 4(b) appears to nearly coincide with the macro-window candidate. What justifies excluding the other member as a separate action? Could you report both candidates’ intervals and confidences to clarify what fusion adds in this example?
  2. Appendix B.2 reports an 18.7% rescue rate on CAS(ME)³, but the rescue-on/off comparison in B.4.2 uses an exploratory full-cache configuration. Under the main nested setting, how much of the improvement comes from AGF itself and how much from confidence-based rescue?
  3. Although the gate uses emotion agreement, its action labels depend only on IoU, while recognition F1 is measured on retained matched events. Relative to Native, how does the number of correctly localized and classified events change on SAMMLV, where TP falls from 52 to 46?
  4. Figure 4 shows successful rejection and boundary refinement but no failures. Can you contrast these with cases where AGF rejects a true event or fusion worsens localization, and explain which candidate characteristics make the decisions succeed or fail?
  5. Appendix B.4 reports full-cache exploration, while the nested evaluation uses a fixed candidate-window pool and several fixed heuristics. How were this pool and these settings chosen, and did their development use annotations or results from subjects later treated as held-out test subjects?
Reviewer_9X5xOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method named RethinkFuse for micro-expression spotting. The authors argue that traditional single-window methods couple the existence of a micro-expression with its temporal boundaries. To solve this, this paper introduces a decoupling framework. It uses both macro and micro windows to generate initial proposals. Then, it uses an Adaptive Gated Fusion (AGF) routing module that takes scalar differences between the two windows as input and outputs routing actions, including reject, single, fuse, to decouple the alleged "candidate survival" from "boundary construction". The method is evaluated using the Leave-One-Subject-Out cross-validation protocol on standard datasets.

Critical strengths
  • The core idea to separate the binary decision of event existence from the regression of temporal localization is reasonable. It provides a clear perspective to handle the boundary blurring problem which is often seen in single-window methods.
  • The paper follows the LOSO protocol. The emphasis on "subject-disjoint AGF routing" ensures that the post-processing fusion weights are learned strictly on the training set. This effectively prevents subject-level data leakage during the evaluation phase.
Critical weaknesses
  • The overall pipeline leans heavily towards manual feature engineering. The input to the AGF routing module consists of hand-crafted scalar differences (such as the peak difference Δp, confidence difference Δc, duration difference Δd, and emotion difference Δe). Furthermore, the AGF module itself is essentially a simple logistic regression decision maker (a switch module). As a result, the technical novelty is moderate for a top-tier machine learning venue.

  • The paper frequently uses the term "candidate survival". I don't think it is a standard term in the micro-expression spotting domain. The domain usually uses terms like false positive suppression or proposal filtering. This unconventional usage creates an unnecessary reading barrier.

  • Figure 1 lacks the necessary explanations to clearly convey the initial motivation of the paper.

Questions for authors
  • Figure 1 is difficult to understand without detailed captions. The authors may provide a clear and step-by-step explanation of this figure to better illustrate the motivation?

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

Related papers

Discussion (0)

Sign in to comment.