STEMorph: A Set of Morphed Emotional Face Stimuli from Angry to Happy Derived from NimStim

This paper introduces STEMorph, a validated and ecologically valid stimulus set of morphed emotional faces derived from NimStim that utilizes neutral-anchored morphing and neural-network masks to overcome traditional methodological limitations and provide a standardized tool for investigating facial emotion recognition biases.

The Gist

Imagine you are trying to teach a computer, or even a group of people, how to tell the difference between a person who is furious and a person who is overjoyed.

To do this, you need a "ladder" of pictures. You start with a picture of pure anger at the bottom, a picture of pure happiness at the top, and you need a bunch of steps in between that slowly change from one to the other. This is what scientists call a "morphed" stimulus set.

However, there's a problem with how most scientists have built these ladders in the past.

The Problem: The "Uncanny Valley" Ladder

Think of traditional face-morphing like trying to blend two photos using a cheap photo editor. If you just smash an angry face and a happy face together, the middle steps often look weird. The hair might get stuck to the forehead, the ears might look like they're melting, or the face might look like a distorted alien.

It's like trying to blend a red smoothie and a blue smoothie, but in the middle, you get a purple sludge that tastes like plastic. If the picture looks fake, people's brains get confused. They might not be reacting to the emotion (anger vs. happiness); they might just be reacting to the weirdness of the picture. This makes scientific results messy and unreliable.

The Solution: STEMorph

The authors of this paper, led by researchers from Oxford and Iranian universities, built a new, better ladder called STEMorph. They wanted to create a set of faces that looked as real as possible, so people would react to the emotions naturally.

Here is how they did it, using some creative analogies:

1. The "Neutral" Anchor (The Safe Harbor)
Instead of just smashing "Angry" and "Happy" together, they used a Neutral face (a calm, blank expression) as a middleman.

• Analogy: Imagine you are driving from a freezing cold city (Anger) to a scorching hot city (Happiness). If you drive straight there, the temperature change is too sudden and jarring. Instead, you stop at a mild, comfortable town (Neutral) in the middle. This makes the transition smooth and natural. In STEMorph, the "Neutral" face acts as this safe harbor, ensuring the transition doesn't look like a glitch.

2. The "Neural Network" Mask (The Digital Hairdresser)
When you morph faces, the background (hair, ears, neck) often gets messed up. Traditional methods use a simple oval cutout, which is like cutting a cookie out of dough—it leaves jagged edges and cuts off important bits.

• Analogy: The authors used a super-smart AI (a neural network) to act like a digital hairdresser. Instead of just cutting a circle, the AI carefully snips around every single strand of hair and the curve of the ear, keeping the face perfectly intact while removing the messy background. This ensures the "weirdness" is gone, and the face looks like a real person.

The Experiment: Testing the Ladder

To see if their new ladder worked, they showed 50 volunteers (mostly medical students) 198 different faces.

• The faces ranged from 1 (Super Angry) to 9 (Super Happy).
• The volunteers had to guess where on the scale each face fell.

The Results:

• It Worked: The volunteers' guesses matched the scientists' plan almost perfectly. As the faces got "happier" in the computer code, the volunteers rated them as happier. The correlation was so strong it was practically a straight line.
• It's Reliable: They asked a group of people to do the test again two weeks later. The results were almost identical, proving the faces aren't confusing people; they are consistent.
• Gender Matters: They found that, just like in real life, people perceived emotions differently based on gender.
  • The Faces: People generally rated female faces as slightly more expressive or "happier" than male faces.
  • The Viewers: Female participants tended to be more sensitive to the subtle changes in the faces than male participants.

Why Does This Matter?

This paper is like giving scientists a high-quality ruler instead of a bent, broken one.

Before, researchers studying emotions (like how people with depression or anxiety see the world) had to use "bent rulers" (the old, weird-looking face sets). This made it hard to know if their results were real or just an artifact of bad pictures.

Now, with STEMorph, researchers have a tool that is:

1. Natural: It looks like a real human face.
2. Precise: It moves smoothly from anger to happiness.
3. Fair: It accounts for how men and women might see things differently.

This allows doctors and scientists to better understand how our brains process emotions, which could eventually help in treating mental health conditions where emotion recognition goes wrong. It's a small step in a lab, but a giant leap for making our understanding of human feelings more accurate.

Technical Summary

1. Problem Statement

Emotion recognition research relies heavily on facial stimuli, but existing methodologies face significant limitations:

• Inconsistency in Findings: Studies on emotional recognition biases (e.g., in depression or anxiety) often yield conflicting results, partly due to methodological variations in stimulus presentation.
• Artifacts in Traditional Morphing: Standard face morphing techniques often produce unnatural, "uncanny" artifacts, particularly in non-facial regions (hair, ears, background) and at transition points between opposing emotions (e.g., anger to happiness).
• Lack of Ecological Validity: Many existing stimulus sets (e.g., KDEF, POFA) use simple geometric masks (ovals/squares) or lack intermediate neutral anchors, leading to distorted facial features that confound the perception of genuine emotional gradients.
• Need for Standardization: There is a lack of publicly available, validated stimulus sets that combine naturalistic appearance with precise control over emotional intensity, specifically addressing gender-related perceptual biases.

2. Methodology

Stimulus Generation (STEMorph)

• Source Material: Derived from the NimStim set of facial expressions.
• Target Emotions: A continuum from Angry to Happy, anchored by a Neutral expression.
• Participants in Stimuli: 22 individuals (11 male, 11 female) were selected based on their ability to facilitate smooth morphing without extreme exaggeration.
• Morphing Process:
  • Neutral-Anchored Morphing: Instead of a direct angry-to-happy transition, the authors used a neutral face as a central anchor point. This prevents the "uncanny valley" effect caused by conflicting muscle activations (e.g., corrugator vs. zygomaticus) at the midpoint.
  • Steps: The transition involves 9 steps (1–9): 4 steps from Angry to Neutral, the Neutral point, and 4 steps from Neutral to Happy.
  • Software: Abrosoft Fantamorph was used for the morphing.
• Masking and Naturalism:
  • Neural Network Masking: To eliminate artifacts in hair and ears, the authors used FaceSwap software utilizing the BiSeNet-FP neural network. This generated distinct, naturalistic face masks for each morphing step, extracting only the face region.
  • Manual Correction: Manual editing was performed to fix residual artifacts (e.g., stray hair strands).
• Final Dataset: 198 distinct faces (22 identities × 9 morph steps).

Experimental Design & Validation

• Participants: 50 medical students (22 female, 28 male; mean age 21.5) from Iranian universities. All had normal/corrected vision and no psychiatric history.
• Task: Participants rated each face on a 9-point Likert scale (1 = Angriest, 9 = Happiest).
  • Procedure: Stimuli were presented for 4000ms in a dimly lit room. A noise mask (scrambled previous trial) followed each trial to prevent priming.
  • Counterbalancing: The scale direction was reversed for half the participants to control for motor/numerical bias.
• Reliability Check: A subgroup of 23 participants returned two weeks later to repeat the task.
• Statistical Analysis:
  • Validity: Ordinary Least Squares (OLS) linear regression modeling emotional ratings as a function of Morph Step, Face Gender, Participant Gender, and their interactions.
  • Reliability: Linear regression comparing initial ratings to retest ratings.
  • Subgroup Analysis: Partial regression analysis to isolate gender-specific effects.

3. Key Contributions

1. STEMorph Stimulus Set: A new, publicly available dataset of 198 morphed faces transitioning from anger to happiness via a neutral anchor.
2. Methodological Innovation:
   • Implementation of neutral-anchored morphing to ensure perceptual coherence at the emotional midpoint.
   • Use of neural-network-generated masks (BiSeNet-FP) to preserve anatomical fidelity and remove non-facial artifacts, superior to traditional geometric masking.
3. Comprehensive Validation: The set is psychophysically validated for both validity (correlation with intended emotion) and reliability (test-retest consistency).
4. Normative Data: Provides detailed normative data on how face gender and participant gender influence emotional perception, addressing a gap in gender-specific emotion research.

4. Results

Validity

• Strong Linear Correlation: There was a robust positive association between the morph step and subjective ratings ($\beta = 0.735$, $p < 0.001$).
• Model Fit: The regression model explained 74.1% of the variance in ratings ($R^2 = 0.741$).
• Bayesian Evidence: The Bayes Factor approached infinity, providing overwhelming evidence for the model over the null hypothesis.
• Gender Effects:
  • Face Gender: Female faces received higher emotional intensity ratings than male faces ($\beta = -0.248$).
  • Participant Gender: Male participants gave higher ratings than female participants ($\beta = 0.326$).
  • Interaction: Female participants showed steeper psychometric slopes (greater sensitivity to emotional gradations) compared to male participants.
  • Identity Variability: The gender effect was distributed across identities rather than driven by a few outliers, though individual identity characteristics still contributed to variability.

Reliability

• Test-Retest Consistency: The correlation between initial ratings and retest ratings (2 weeks later) was strong ($R^2 = 0.736$, $p < 0.001$).
• Stability: The high reliability indicates that the stimuli elicit consistent interpretations across multiple exposures, with minimal variability due to the stimulus design.

5. Significance and Implications

• Ecological Validity: By using neural-network masks and neutral anchoring, STEMorph overcomes the "uncanny" artifacts of previous morphed sets, making it more suitable for studying naturalistic emotion recognition.
• Clinical and Research Utility: The set is ideal for investigating emotion recognition biases in clinical populations (e.g., MDD, GAD, SAD) where subtle shifts in perception (e.g., interpreting neutral as angry) are critical.
• Standardization: STEMorph offers a standardized tool to reduce methodological heterogeneity across studies, potentially resolving inconsistencies in the literature regarding emotional biases.
• Open Science: The authors have made the stimulus set, data, and analysis code publicly available via GitHub and the NimStim website, fostering transparency and reproducibility in the field of social cognition and affective neuroscience.

In conclusion, STEMorph represents a significant methodological advancement in facial emotion research, providing a high-fidelity, validated, and gender-sensitive stimulus set that bridges the gap between experimental control and ecological realism.