MEGC2026: Micro-Expression Grand Challenge on Visual Question Answering

The MEGC 2026 challenge introduces two new tasks, Micro-Expression Video Question Answering (ME-VQA) and Micro-Expression Long-Video Question Answering (ME-LVQA), to advance the analysis of facial micro-expressions by leveraging the multimodal reasoning capabilities of large vision-language models on both short and long-duration video sequences.

The Gist

Imagine you are at a high-stakes poker game. Everyone is trying to look calm, but one player just got a terrible hand. For a split second—faster than a camera shutter—his face twitches. His lip curls slightly, or his eyebrow raises. Then, just as quickly, he masks it with a poker face. That tiny, involuntary flicker is a Micro-Expression (ME). It's the truth trying to peek out from behind a lie.

For years, scientists have been trying to build computers that can spot these fleeting "truth flashes." But now, the game has changed. The MEGC 2026 (Micro-Expression Grand Challenge 2026) is a new competition asking a different, more human question: "Can you not just spot the twitch, but actually talk about it?"

Here is a simple breakdown of what this paper is about, using some everyday analogies.

The Big Idea: From "Spotting" to "Chatting"

In the past, computers were like security guards with a checklist. They would scan a face and say, "Yes, I see a micro-expression," or "No, I don't."

MEGC 2026 is upgrading the computer to be more like a detective with a notebook. Instead of just checking a box, the computer is given a video and asked natural questions like:

• "What emotion was the person feeling right before they smiled?"
• "How many times did they try to hide their anger?"
• "Describe the specific muscle movement around the eyes."

This is powered by Multimodal Large Language Models (MLLMs). Think of these models as super-smart students who have read every book on human emotion and watched millions of videos. They can look at a video and "chat" about what they see, combining visual clues with language skills.

The Two Main Challenges (The Tasks)

The competition has two levels, like a video game with a "Tutorial" and a "Boss Level."

Level 1: The Short-Clip Detective (ME-VQA)

• The Scenario: You are shown a very short video clip (a few seconds long) of someone's face.
• The Task: You are asked a specific question about that clip.
  • Example: "Did the person show a 'lip corner depressor' (a sign of sadness)?" or "Is this a happy or angry micro-expression?"
• The Goal: The computer must answer in full sentences, explaining why it thinks that. It's like a teacher asking a student to explain their answer, not just give a number.

Level 2: The Long-Haul Detective (ME-LVQA)

• The Scenario: This is the hard mode. You are given a long video (like a whole conversation or a tense meeting) that might last several minutes.
• The Task: The video is full of normal talking, laughing, and frowning. Hidden inside are tiny, fleeting micro-expressions. You have to find them and answer complex questions.
  • Example: "How many times did the person try to suppress a smile during the meeting?" or "List all the different facial movements that happened."
• The Challenge: This is like finding a needle in a haystack, but the haystack is moving, and the needle is invisible to the naked eye. The computer has to remember what happened 30 seconds ago to answer a question about what happened 5 minutes ago.

The "Test Drive" (Baseline Results)

The authors of the paper didn't just propose the idea; they tried it out using two powerful AI models (called Qwen). Think of these models as two different cars the researchers rented to see if they could win the race.

• The Zero-Shot Test (Driving without a map): They asked the AI to do the job without any special training on micro-expressions.
  • Result: The AI was okay at spotting big, obvious emotions (like "Is he happy?"), but it was terrible at spotting the tiny, subtle ones. It was like a driver who can see a stop sign but misses a small pothole.
• The Fine-Tuned Test (Driving with a map): They gave the AI a crash course using specific micro-expression videos.
  • Result: The AI got better at writing good sentences and describing what it saw. However, it still struggled with the hardest part: counting exactly how many micro-expressions happened and pinpointing exactly when they occurred in long videos.

The Takeaway

The paper concludes that while AI is getting smarter at "talking" about emotions, it still has a long way to go to truly "understand" the subtle, split-second lies and truths of the human face.

• The Good News: AI can now describe micro-expressions in natural language, making the technology more useful for real-world applications (like lie detection or mental health analysis).
• The Bad News: Long videos are still too confusing for current AI. The models get lost in the noise of normal facial movements and miss the tiny details.

In short: MEGC 2026 is inviting researchers to teach computers how to be better observers of human emotion, moving from simple "spotting" to complex "storytelling" about what people are really feeling.

Technical Summary

Here is a detailed technical summary of the paper "MEGC2026: Micro-Expression Grand Challenge on Visual Question Answering."

1. Problem Definition

The paper addresses the evolving challenges in Facial Micro-Expression (ME) analysis. MEs are involuntary, fleeting facial movements (typically <500ms) that occur when a person attempts to suppress an emotion, often in high-stakes environments. While previous research focused on recognition, spotting, and generation, the field is now transitioning toward Multimodal Large Language Models (MLLMs) and Large Vision-Language Models (LVLMs).

The core problem is to leverage these advanced models to perform Visual Question Answering (VQA) on ME videos. The challenge is divided into two distinct tasks:

• ME-VQA (Short Video): Analyzing short, pre-segmented clips to answer questions about specific ME attributes (e.g., emotion class, action units).
• ME-LVQA (Long Video): Analyzing long-duration, naturalistic video sequences. This requires models to handle temporal reasoning, distinguish between macro-expressions (MaE) and micro-expressions (ME) over time, and count events or identify specific attributes within a continuous stream of facial activity.

2. Methodology

The authors propose a framework utilizing state-of-the-art LVLMs to tackle these tasks, establishing baselines to guide future research.

• Baseline Models: The challenge employs Qwen2.5VL-3B and Qwen3VL-4B as baseline models. These models integrate a vision encoder, a language backbone, and a cross-modal fusion module.
• Training Strategies:
  • Zero-Shot (ZS): Evaluating the models' inherent capabilities without task-specific training.
  • Fine-Tuning (FT): Using QLoRA (Quantized Low-Rank Adaptation) to efficiently update parameters. The adapters are applied to the vision encoder, projection layers, and the query/key components of the language model.
• Datasets:
  • Training: Participants are encouraged to use standard datasets like SAMM, CASME II, SMIC, CAS(ME)3, and 4DME.
  • ME-VQA: Utilizes the ME-VQA-v2 dataset (curated from SAMM and CAS(ME)3), containing 24 test clips.
  • ME-LVQA: Utilizes a curated training set and a test set of 30 long videos (10 from SAMM, 20 from CAS(ME)3).
• Task Formulation:
  • Inputs are video sequences + natural language questions.
  • Outputs are natural language answers describing MEs, counting events, or classifying expression types.

3. Key Contributions

• Introduction of ME-LVQA: The primary novelty is the ME Long-Video Question Answering task, which moves beyond short clips to realistic, long-duration scenarios requiring temporal localization and event aggregation.
• Multimodal Paradigm Shift: The challenge reformulates traditional ME annotations (Action Units, emotion labels) into Question-Answer (QA) pairs, promoting interpretability and human-aligned interaction.
• Comprehensive Benchmarking: The paper provides a rigorous evaluation of current LVLMs (Qwen series) on ME tasks, establishing the first public leaderboards for both short and long-video ME-VQA.
• Dataset Curation: Release of refined datasets (ME-VQA-v2 and ME-LVQA) specifically formatted for training and testing VQA models on micro-expressions.

4. Experimental Results

The baseline results highlight significant gaps between current LVLM capabilities and the requirements of ME analysis.

Task 1: ME-VQA (Short Video)

• Performance: Both Qwen2.5VL-3B and Qwen3VL-4B showed moderate performance in coarse-grained emotion classification (Positive/Negative/Surprise) but extremely poor performance in fine-grained classification (specific emotions like fear, disgust, etc.).
• Fine-Tuning Impact: Fine-tuning improved language generation metrics (BLEU, ROUGE) significantly, indicating models learned to speak "correctly." However, fine-grained emotion recognition (UF1) remained near zero or very low, suggesting models struggle to capture subtle visual cues.
• Key Finding: Models can distinguish broad emotional valence but fail at the subtle distinctions required for MEs.

Task 2: ME-LVQA (Long Video)

• Performance Degradation: As expected, performance dropped significantly compared to short clips.
• Regression Tasks (Counting): Models exhibited high Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) in counting ME and MaE events.
• Classification Tasks:
  • AU Recognition: Set-based F1 scores were low (~0.16–0.24), indicating difficulty in identifying specific Action Units.
  • ME vs. MaE Classification: Unweighted F1 (UF1) scores ranged from 0.36 to 0.53, showing limited ability to distinguish micro from macro expressions in long contexts.
• Fine-Tuning Limitations: While fine-tuning helped, the gains were modest. The authors attribute this to the limited diversity of the fine-tuning subset (only 10 subjects), suggesting models are overfitting to specific identities rather than learning robust temporal dynamics.

5. Significance and Future Directions

• Real-World Applicability: By introducing long-video analysis, the challenge bridges the gap between controlled lab datasets and real-world scenarios where MEs occur amidst natural, continuous facial behavior.
• Limitations Identified: The results underscore that current LVLMs are not yet robust enough for fine-grained ME analysis. The "subtle" nature of MEs requires models to be highly sensitive to temporal dynamics and identity-invariant features, which current baselines lack.
• Call to Action: The paper emphasizes the need for:
  1. Larger and more diverse training data to prevent identity overfitting.
  2. Advanced temporal reasoning architectures capable of handling long-range dependencies.
  3. Better integration of visual cues to improve fine-grained emotion and Action Unit detection.

In conclusion, MEGC 2026 sets a new standard for ME research by shifting focus from pure classification to multimodal reasoning and long-term temporal understanding, revealing that while LVLMs offer a promising direction, significant architectural and data-driven advancements are required to achieve human-level performance in micro-expression analysis.