Evaluating a Multitask AI Model versus Humans for Portion Size Estimation

This study demonstrates that while AI models show promise for estimating Central Asian portion sizes, culturally adapted visual food atlases significantly outperform both unassisted human judgment and AI in accuracy, particularly for small and complex food items.

The Gist

Imagine you are trying to guess how much food is on a plate at a big family dinner in Central Asia. The food is served on huge, shared platters, and there are no measuring cups or scales. You just have to guess: "Is this a small bowl of rice, a medium one, or a giant mountain?"

This is exactly the problem doctors and researchers face when trying to understand what people eat in Central Asia. If they get the portion sizes wrong, they can't figure out if people are getting enough nutrients or if they are eating too much, which leads to health problems like diabetes and heart disease.

This paper is a "taste test" of three different ways to solve this guessing game. The researchers wanted to see which method is the best at guessing the weight of food:

1. The "Gut Feeling" Method: Just guessing with your eyes (Unassisted).
2. The "Picture Book" Method: Using a special digital photo album that shows you what small, medium, and large portions look like (Visual Atlas).
3. The "Robot Brain" Method: Using a smart computer program (AI) that has studied thousands of food photos to guess the weight (AI Model).

Here is how the race went down, explained with some simple analogies:

🏁 The Race: Who Won?

1. The "Gut Feeling" (Unassisted Humans)

• The Analogy: This is like trying to guess the weight of a watermelon while blindfolded. You have no reference point.
• The Result: This was the worst method. People guessed wildly. They often thought a small portion was huge, or a huge portion was tiny. The errors were massive, like guessing a 500g steak was actually 1kg.

2. The "Robot Brain" (The AI Model)

• The Analogy: Imagine a super-smart robot chef who has memorized a library of food photos. It's great at recognizing things that look the same every time, like a perfectly round cookie or a glass of soda.
• The Result: The AI was a mixed bag.
  • When it shined: It was excellent at guessing big, clear portions (like a full bowl of soup) and drinks. It was almost as good as the picture book for these.
  • When it stumbled: It got very confused by small portions and messy foods (like a pile of meat or a salad). It's like the robot tried to count the grains of sand in a small pile and got the math wrong. It struggled with "amorphous" foods—things that don't have a fixed shape.

3. The "Picture Book" (The Visual Atlas)

• The Analogy: This is like giving someone a ruler and a set of reference photos. Instead of guessing, they can say, "Oh, this plate looks exactly like the 'Medium' photo in my book."
• The Result: This was the clear winner. The people who used the digital photo book made the fewest mistakes. Because the photos showed real Central Asian dishes served in real local bowls, it bridged the gap between "what I see" and "how much it weighs."

🥘 The Special Challenges: Meat and Dairy

Central Asian diets rely heavily on meat and dairy. The researchers found that:

• Meat: The AI got really confused with small pieces of meat (like a single meatball). It thought they were huge! The picture book, however, helped people see exactly how big a "normal" serving of meat is.
• Dairy: The AI was actually quite good at guessing big bowls of yogurt or milk, but it failed miserably at small amounts. The picture book was consistent and reliable for everything.

🧠 The Big Takeaway

The study teaches us two main lessons:

1. Technology isn't a magic wand yet. While AI is amazing and getting better, it still struggles with the messy, complex reality of real-world food, especially small portions and traditional communal meals. It needs more training on "small stuff."
2. Simple tools work best. A well-designed, culturally specific picture book (a digital atlas) is currently the most reliable tool we have. It helps people who aren't nutrition experts to finally understand what a "healthy portion" actually looks like in their own culture.

In a nutshell: If you want to know exactly how much food someone in Central Asia is eating, don't just ask them to guess, and don't just rely on a robot yet. Give them a picture book that shows them what a real portion looks like. It's the most accurate way to get the numbers right so we can build better health policies for the future.

Technical Summary

1. Problem Statement

Accurate dietary assessment is critical for precision nutrition and public health surveillance, yet portion size estimation remains a significant bottleneck, particularly in Central Asia.

• Cultural Context: Traditional Western assessment tools fail in Central Asian contexts due to communal eating practices (sharing from large platters), non-standardized portioning, and a lack of familiarity with metric units (grams/cups).
• Data Gap: Existing AI models and visual atlases are predominantly validated in Western settings. There is a lack of robust, region-specific data to determine whether AI or human-assisted visual tools perform better in low-resource, communal-eating environments.
• Consequence: Inaccurate estimates lead to flawed nutrient intake data, undermining evidence-based policy and interventions for Non-Communicable Diseases (NCDs) like obesity and diabetes.

2. Methodology

The study employed a cross-sectional comparative validation design involving 128 adult participants recruited from Astana, Kazakhstan.

Experimental Groups

Participants were randomized into two human estimation groups, while a third group utilized an AI model:

1. Unassisted Human Judgment: Participants estimated portion sizes based solely on standardized food photographs without aids.
2. Atlas-Assisted Human Judgment: Participants used a culturally adapted digital visual food atlas containing three portion sizes (small, medium, large) for 60 food types and 8 beverages, derived from the Central Asian Food Dataset (CAFD) and Central Asian Food Scenes Dataset (CAFSD).
3. AI Model: A YOLOv12-based end-to-end multitask framework trained on over 21,000 regionally sourced images. The model utilized pre-trained weights from the Food Portion Benchmark (FPB) dataset and was fine-tuned on Central Asian food categories to perform simultaneous food detection and volume estimation.

Data Collection & Analysis

• Stimuli: 51 food types and 8 beverages presented in three portion sizes (small, medium, large).
• Reference Standard: Actual laboratory-measured weights/volumes served as the ground truth.
• Metrics: Accuracy was evaluated using Mean Absolute Error (MAE) (magnitude of error in grams) and Mean Absolute Percentage Error (MAPE) (relative error).
• Stratification: Results were analyzed by food category (Meat, Dairy, Flour, Salads, etc.) and portion size.

3. Key Contributions

• Validation of Region-Specific Tools: The study validated a Central Asia-specific digital visual food atlas, confirming its efficacy in improving estimation accuracy for populations with communal eating habits.
• AI Benchmarking: It provides the first direct comparison of a specialized AI model against human judgment and visual atlases specifically within a Central Asian cultural context.
• Dataset Integration: The research leveraged and fine-tuned models using the CAFD and CAFSD, addressing the scarcity of high-quality, culturally diverse food datasets for AI training.
• Granular Performance Analysis: The study dissects performance not just by method, but by food texture/structure (e.g., amorphous vs. dense) and portion size, offering specific insights into where AI succeeds or fails.

4. Key Results

Overall Accuracy

• Atlas-Assisted Group: Achieved the highest accuracy overall.
  • MAE: 80.81g
  • MAPE: 44.76%
• AI Model: Showed promising results for average/large portions but struggled significantly with small portions.
  • MAE: 79.07g (comparable to atlas for large items)
  • MAPE: 67.91% (higher error rate due to small portion sensitivity)
• Unassisted Group: Performed the worst.
  • MAE: 133.86g
  • MAPE: 79.40%

Performance by Food Category & Portion Size

• Small Portions: The AI model's performance degraded sharply for small portions, particularly for meat-based items (MAPE reached ~144.55% for small meat portions). The atlas-assisted method reduced MAPE by nearly 50% in this category compared to unassisted estimates.
• Food Texture:
  • AI Strengths: Performed well with dense, single-component items (e.g., qurt, cheburek) and beverages (e.g., coke, kymyz).
  • AI Weaknesses: Struggled with amorphous foods, salads, and liquid-based dishes (soups) where 2D images fail to capture volume and texture variations.
• Dairy & Meat: The atlas consistently outperformed both AI and unassisted methods across these central dietary categories. However, the AI model achieved the lowest MAPE (27.57%) for large dairy portions, suggesting it is viable for standardized, large-volume servings.

5. Significance and Implications

• Public Health Strategy: The findings support the integration of culturally adapted visual atlases as the primary tool for dietary monitoring in Central Asia, especially for small and complex communal meals.
• AI Deployment Roadmap: While AI holds promise for scalability and real-time assessment, it currently requires further refinement (specifically training on small portions and complex textures) before it can replace human-assisted methods in diverse settings.
• Future Directions: The study suggests a hybrid approach: using visual atlases for baseline accuracy and AI for scalable, automated tracking of standardized or large-portion foods. It highlights the need for expanded datasets including more liquid-based and small-portion traditional foods to improve AI robustness.
• Policy Impact: Accurate tools are essential for addressing the dual burden of malnutrition and NCDs in Central Asia, enabling better-targeted nutritional interventions.

Conclusion: While AI models are advancing rapidly, culturally tailored visual aids currently remain the superior method for accurate portion size estimation in Central Asian communal eating contexts, particularly for small and structurally complex foods.