Micro-Expression Grand Challenge (MEGC) 2026

Facial micro-expressions (MEs) are spontaneous, fleeting facial movements (less than 500ms) that reveal suppressed emotions, making them critical for high-stakes analysis but notoriously difficult to detect due to data scarcity and annotation challenges. MEGC 2026 aims to bridge this gap by leveraging the emerging capabilities of Vision-Language Models (VLMs) and Multimodal Large Language Models (LLMs). While these models excel in general reasoning, their ability to interpret the subtle, fine-grained cues of micro-expressions remains an open research question that this competition seeks to answer through novel benchmarking tasks.

By introducing Micro-Expression Short-Video (ME-VQA) and the pioneering Long-Video Question Answering (ME-LVQA) tracks, the competition pushes boundaries from simple recognition to interpretable natural language understanding. Advancing this technology will have transformative impacts across healthcare (for mental state assessment), human–machine interaction (creating empathetic systems), and security (enhancing risk assessment), ultimately deepening our scientific understanding of nonverbal human communication.

In detail, MEGC 2026 introduces:

  • Task 1 - ME-VQA: a continuation of the ME-VQA task (MEGC 2025) for benchmarking short-clip subtle-emotion reasoning.
  • Task 2 - ME-LVQA: a new ME-LVQA task Micro-Expression Long-Video Question Answering, requiring reasoning and description across long videos.

Please contact Xinqi (x.fan@mmu.ac.uk) with an email subject starting with MEGC2026 if you have any questions about the competition.

News

  • 22/12/2025: Website made live
  • 29/01/2026: Task 1: ME-VQA is updated with an improved ME-VQA-v2
  • 01/02/2026: Task 2: ME-LVQA test sets released
  • 09/02/2026: Task 1: ME-VQA testing platform is available at Codabench ME-VQA
  • 19/02/2026: Task 2: curated ME-LVQA dataset released

Important Dates

The competition schedule is aligned with the FG 2026 timeline:

  • Competition Period: 23rd December 2025 – 9th March 2026
  • Training Set Release: 23rd December 2025
  • Test Set Release: 1st February 2026
  • Competition End Date: 30th March 2026
  • Winner Paper Submissions and Code Verification: 6th April 2026
  • Notification of Acceptance: 13th April 2026
  • Camera-Ready Deadline: 21st April 2026 (FG 2026 Camera-Ready)
  • Challenge Date: 25th or 29th May 2026 (during FG 2026)

Tasks and Application Scenarios

Task 1: ME-VQA (Micro-Expression Visual Question Answering)

This task extends MEGC 2025. Participants receive short video clips containing a single ME and must answer natural-language questions such as:

  • "What fine-grained emotion is expressed?"
  • "What action units are shown on the face?"
  • "Please provide an analysis and the reasoning process of the expression."
  • ...

Task 2: ME-LVQA (Micro-Expression Long-Video Question Answering)

This is a new FG 2026 task. Participants are given long videos. Each video may contain multiple or zero MEs. Participants must describe and explain all ME-related information contained in these long videos. Questions may include:

  • "How many micro-expressions occur in this video?"
  • "What emotions are expressed?"
  • "Explain which facial actions occur in this video and why they indicate these emotions."
  • ...

Data

The training data for the competition is recommended from existing ME datasets that are publicly available. These recommendations are split based on the tasks previously described. In addition, we include unseen test data that is used for the final results calculation for both tasks.

Task 1 - ME-VQA - Recommended Training Data

  • Curated ME VQA dataset: Improved from the MEGC2019 composite dataset with clips from SAMM, CASME II, and SMIC by adding QA pairs.
    • The ME-VQA dataset used for MEGC 2025 (last year) is here: ME-VQA (ME-VQA-v1).
    • We have improved the comprehensive QA pairs by including some variations, and made the new ME-VQA-v2.
    • In ME-VQA-v2, the comprehensive QA questions become more general, and the answers have a few variations.
    • Reference (MEGC 2025 with ME-VQA): X, Fan, J. Li, J. See, M. H. Yap, W.-H. Cheng, X. Li, X. Hong, S.-J. Wang and A. K. Davison., "MEGC2025: Micro-expression grand challenge on spot then recognize and visual question answering" Proceedings of the ACM International Conference on Multimedia (ACM MM)., 2025.
  • SAMM (159 ME clips at 100 fps):
    • To download the dataset, please visit: http://www2.docm.mmu.ac.uk/STAFF/M.Yap/dataset.php. Download and fill in the license agreement form, email to M.Yap@mmu.ac.uk with email subject: SAMM videos.
    • Reference: Dvison, A. K., Lansley, C., Costen, N., Tan, K., & Yap, M. H. (2016). SAMM: A spontaneous micro facial movement dataset. IEEE Transactions on Affective Computing, 9(1), 116-129.
  • CASME II (247 ME clips at 200 fps):
    • To download the dataset, please visit: https://melabipcas.github.io/melab/en/db/casme2.html. Download and fill in the license agreement form, submit throuth the website. >.
    • Reference: Yan, W. J., Li, X., Wang, S. J., Zhao, G., Liu, Y. J., Chen, Y. H., & Fu, X. (2014). CASME II: An improved spontaneous micro-expression database and the baseline evaluation. PloS one, 9(1), e86041.
  • CAS(ME)3 (1300 long videos at 30 fps):
    • To download the dataset, please visit: https://melabipcas.github.io/melab/en/db/casme3.html. Download and fill in the license agreement form, submit throuth the website.
    • Reference: Li, J., Dong, Z., Lu, S., Wang, S. J., Yan, W. J., Ma, Y., ... & Fu, X. (2022). CAS (ME)3: A third generation facial spontaneous micro-expression database with depth information and high ecological validity. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 3, pp. 2782-2800, doi: 10.1109/TPAMI.2022.3174895.
  • SMIC-E-long (162 long videos at 100 fps):
    • To download the dataset, please visit: https://www.oulu.fi/cmvs/node/41319. Download and fill in the license agreement form (please indicate which version/subset you need), email to Xiaobai.Li@oulu.fi.
    • Reference: Tran, T. K., Vo, Q. N., Hong, X., Li, X., & Zhao, G. (2021). Micro-expression spotting: A new benchmark. Neurocomputing, 443, 356-368.
  • 4DME (1068 ME clips at 60 fps):

Task 2 - ME-LVQA - Recommended Training Data

  • Curated ME LVQA dataset: Improved from the SAMM-LV and CAS(ME)3 (part A) by adding QA pairs.
    • The ME-LVQA dataset can be found here: ME-LVQA.
    • We apologize that the ME-LVQA dataset may be updated in the future, but we will keep you informed of any changes.
  • SAMM Long Videos (147 long videos at 200 fps):
    • To download the dataset, please visit: http://www2.docm.mmu.ac.uk/STAFF/M.Yap/dataset.php. Download and fill in the license agreement form, email to M.Yap@mmu.ac.uk with email subject: SAMM long videos.
    • Reference: Yap, C. H., Kendrick, C., & Yap, M. H. (2020, November). SAMM long videos: A spontaneous facial micro-and macro-expressions dataset. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020) (pp. 771-776). IEEE.
  • CAS(ME)2 (97 long videos at 30 fps):
    • To download the dataset, please visit: https://melabipcas.github.io/melab/en/db/casme_sq.html. Download and fill in the license agreement form, submit throuth the website. .
    • Reference: Qu, F., Wang, S. J., Yan, W. J., Li, H., Wu, S., & Fu, X. (2017). CAS (ME) $^ 2$: a database for spontaneous macro-expression and micro-expression spotting and recognition. IEEE Transactions on Affective Computing, 9(4), 424-436.
  • CAS(ME)3 (1300 long videos at 30 fps):
    • To download the dataset, please visit: https://melabipcas.github.io/melab/en/db/casme3.html. Download and fill in the license agreement form, submit throuth the website.
    • Reference: Li, J., Dong, Z., Lu, S., Wang, S. J., Yan, W. J., Ma, Y., ... & Fu, X. (2022). CAS (ME)3: A third generation facial spontaneous micro-expression database with depth information and high ecological validity. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 3, pp. 2782-2800, doi: 10.1109/TPAMI.2022.3174895..
  • SMIC-E-long (162 long videos at 100 fps):
    • To download the dataset, please visit: https://www.oulu.fi/cmvs/node/41319. Download and fill in the license agreement form (please indicate which version/subset you need), email to Xiaobai.Li@oulu.fi.
    • Reference: Tran, T. K., Vo, Q. N., Hong, X., Li, X., & Zhao, G. (2021). Micro-expression spotting: A new benchmark. Neurocomputing, 443, 356-368.
  • 4DME (270 long videos at 60 fps):
  • ME-LVQA: Will be prepared and released when we release the competition baseline.

Unseen Test Data

The unseen test set for the VQA task contains 24 ME clips, including 7 clips from SAMM (SAMM Challenge dataset) and 17 clips from different videos in CAS(ME)3.
The unseen test set for the LVQA task will contain 30 long videos, including 10 long videos from SAMM and 20 clips cropped from different videos in CAS(ME)3.

To obtain the test sets, participants will be required to download and fill in the license agreement forms for the SAMM Challenge dataset and CAS(ME)3, and upload them to the organisers through a query link.

Evaluation Metrics

To evaluate the performance of both our ME-VQA and ME-LVQA, we report metrics for both emotion classification and the overall language answer quality. The final ranking of the challenge participants will be based on the average (Avg) of all the overall metrics.

Emotion Classification

For both coarse- and fine-grained emotion classification, we use Unweighted F1 Score (UF1) and Unweighted Average Recall (UAR) to ensure balanced evaluation across classes.

  • Coarse classes: positive, negative, surprise.
  • Fine-grained classes: happiness, surprise, fear, disgust, anger, sadness.

Generated Text Quality

For all VQA answers, we report BLEU and ROUGE-1 to assess the quality of generated text.

BLEU evaluates n-gram precision between predicted and reference answers as:

$$ \mathrm{BLEU} = \exp \left( \min\left(1 - \frac{r}{c}, 0\right) + \sum_{n=1}^{N} w_n \log p_n \right) $$

ROUGE-1 score is defined as the recall of unigram overlaps between the candidate answer C and the reference answer R:

$$ \mathrm{ROUGE\text{-}1} = \frac{\sum_{w \in V} \min\big(\mathrm{N}_p(w),\ \mathrm{N}_r(w)\big)}{\sum_{w \in V} \mathrm{N}_r(w)} $$

Submission

Competition Submission

The MEGC 2026 competition will use a dedicated Codabench-based competition website.

  • Challenge submission platform for ME-VQA task: TBC
  • Challenge submission platform for ME-LVQA task: TBC
  • Evaluation result file (.json) : This file must include, for every tested video, the dataset name, video ID, and the corresponding predicted outputs. The Codabench platform will parse and evaluate these results, displaying metric-specific scores and leaderboard rankings in real time. A daily submission limit will be imposed to prevent leaderboard overfitting.

Task 1: ME-VQA Submission details

  • Participants can fill in the test VQA to-answer josonl files and renamed them as xxx_pred.jsonl. These files are available at Google Drive and Baidu Drive.
    • me_vqa_casme3_v2_test_to_answer.jsonl
    • me_vqa_samm_v2_test_to_answer.jsonl
  • Submissions to the Leaderboard must be made in the form of a zip file containining the predicted jsonl files with the following filenames:
    • me_vqa_casme3_v2_test_pred.jsonl (for the unseen CAS(ME)3 ME clips)
    • me_vqa_samm_v2_test_pred.jsonl (for the unseen SAMM ME clips)

Task 2: ME-LVQA Submission details

  • Participants can fill in the test LVQA to-answer josonl files and renamed them as xxx_pred.jsonl. These files are available at Google Drive and Baidu Drive.
    • me_lvqa_casme3_test_to_answer.jsonl
    • me_lvqa_samm_test_to_answer.jsonl
  • Submissions to the Leaderboard must be made in the form of a zip file containining the predicted jsonl files with the following filenames:
    • me_lvqa_casme3_test_pred.jsonl (for the unseen CAS(ME)3 ME videos)
    • me_lvqa_samm_test_pred.jsonl (for the unseen SAMM ME videos)

Rules

  • Use of external data: Permitted if clearly declared.
  • Use of pretrained models: Permitted (e.g., VLMs, LLMs) if publicly available and disclosed.
  • Proprietary models: Proprietary models (such as, GPT API) NOT allowed.
  • Prohibition of test-set leakage: Teams must not attempt to access or reverse engineer test labels.
  • Submission limits: Maximum of 3 submissions per day to Codabench.
  • Mandatory technical report: Required for eligibility.
  • Mandatory Github open access codebase: Required for eligibility.

Cheating Prevention

Hidden test sets, rate-limited submissions, metadata monitoring, cross-checks for label leakage, and manual verification for winning teams will be employed.

Paper Submission

  • Paper submissions will follow the FG 2026 formatting (long paper format, 8 pages + refs) and use the CMT system. Accepted papers will be included in the FG 2026 challenge proceedings.
  • For all other required files besides the paper, please submit in a single zip file and upload to the submission system as supplementary material. It is compulsory to include:
    • GitHub repository URL containing codes of your implemented method, and all other relevant files such as feature/parameter data.
    • CSV files reporting the results.
    The organizers have the right to reject any submissions that: 1) are not accompanied by a paper, 2) did not share the code repository and reported results for verification purposes.
  • Submission Link:TBC

For inquiries, please get in touch via the email addresses listed below.