ACETADA: Nutrition Analysis Benchmark

• ACETADA is a curated dataset pairing realistic meal images with dietitian-verified nutrient labels and rich contextual metadata for precise nutritional estimation.
• It supports single-meal nutrient regression through multimodal reasoning experiments that integrate image data with GPS coordinates, timestamps, and food-presence cues.
• The controlled-feeding trial and detailed annotation protocol enable rigorous evaluation of large multimodal models, significantly reducing estimation errors when using contextual data.

Searching arXiv for the specified paper and closely related nutrition-analysis benchmark context. ACETADA, short for A Controlled-Feeding Trial Dataset for Automated Dietary Assessment, is a food-image dataset created to fill a gap in publicly available nutrition-analysis benchmarks by pairing dietitian-verified nutrient labels with realistic smartphone meal images, precise timestamps, GPS coordinates, and post-meal weights (Coburn et al., 9 Jul 2025). It serves as the empirical foundation for evaluating how Large Multimodal Models (LMMs) estimate calories, macronutrients—protein, carbohydrates, and fat—and portion sizes, especially when image inputs are augmented with contextual metadata such as location, meal timing, and visible food items. The dataset is positioned as a benchmark for single-meal nutrient regression, multimodal reasoning with image + contextual metadata, and prompt-engineering studies involving Chain-of-Thought, Multimodal Chain-of-Thought, Scale Hint, Few-Shot, and Expert Persona prompting (Coburn et al., 9 Jul 2025).

1. Scope, motivation, and benchmark role

ACETADA was introduced in the context of a broader study on nutrition analysis with LMMs. That study notes that existing work primarily evaluates proprietary models, such as GPT-4, and that the broad range of LLMs remains underexplored. It further identifies the influence of integrating contextual metadata—and its interaction with reasoning modifiers—as largely uncharted (Coburn et al., 9 Jul 2025).

Within that framing, ACETADA functions as a benchmark designed for context-aware nutrition analysis. Its defining feature is the combination of meal imagery with metadata derived from multiple channels: GPS coordinates converted to location/venue type, timestamps transformed into meal/day type, and food items present in the image. This design directly supports evaluation of whether contextual information improves nutritional estimation relative to straightforward image-only prompting.

The benchmark targets five prediction variables: Calories (Energy), Protein, Carbohydrates, Fat, and Portion size. The intended tasks are explicitly formulated as single-meal nutrient regression, multimodal reasoning with image + contextual metadata, prompt-engineering studies, and evaluation of closed-weight vs. open-weight Large Multimodal Models (Coburn et al., 9 Jul 2025). A plausible implication is that ACETADA is not merely a corpus for supervised estimation, but also an experimental substrate for studying inference-time prompting behavior under realistic metadata conditions.

2. Data acquisition and experimental setting

The dataset comprises 806 “before-meal” RGB images (pre-consumption) distributed across breakfast (36 %), lunch (32 %), and dinner (32 %). Collection was conducted through a controlled-feeding, randomized crossover trial with 152 adult participants in Perth, Australia (Coburn et al., 9 Jul 2025). Meals were laboratory-prepared and served over three nonconsecutive days, and both foods and beverages were weighed to the nearest 0.1 g before serving and after consumption.

Image acquisition was performed via the mFR24 smartphone app immediately before and after intake. Each image included three instrumentally important components:

• A fiducial marker for visual scale calibration
• Device-local timestamp
• Latitude–longitude coordinates (when available) from the onboard GNSS chip

This acquisition protocol links computer-vision cues with precise measurement infrastructure. The fiducial marker provides a scale reference for portion assessment; the before/after pairing supports consumed-mass estimation rather than served-mass estimation; and the timestamp and geolocation enable downstream construction of contextual variables. The dataset also covers dozens of distinct food and beverage types spanning 11 cuisine categories, with a median of five distinct food items per meal (Coburn et al., 9 Jul 2025).

The report describes ACETADA as enabling research under realistic, free-living conditions with dietitian-verified ground truths (Coburn et al., 9 Jul 2025). Since meal preparation and weighing occurred in a controlled-feeding trial, this wording suggests a hybrid methodological position: realistic smartphone capture and contextual metadata are combined with tightly controlled nutritional measurement.

3. Annotation protocol and nutritional ground truth

Ground-truth construction relied on accredited practicing dietitians. They reviewed every image pair, enumerated all visible foods and beverages in each “before” image, assigned consumed-mass portion weights (served minus leftovers), and calculated nutrient profiles using AUSNUT 2011–2013 food-composition factors (Coburn et al., 9 Jul 2025).

The annotation workflow is specified as a five-step protocol:

1. Pairwise image review (before/after consumption).
2. Food-item identification and normalization of dish names (lowercasing, punctuation removal).
3. Food-item quantity estimation via the fiducial marker and visual cues.
4. Assignment of gram-level consumed weights and lookup of nutrient values in AUSNUT.
5. Verification of total calories, protein, carbohydrate, and fat by a second dietitian as needed.

This protocol ties label generation to both image interpretation and post-consumption evidence. The use of served minus leftovers makes the reference variable consumed mass, not simply plated mass. That distinction matters for nutrition-analysis benchmarks because it couples recognition and portion estimation to actual intake rather than presentation alone.

A fully annotated example in the report consists of a pre-meal image of scrambled eggs and toast, with fiducial marker visible, and metadata specifying Location: “Café, Perth, Australia”, Timestamp: “2024-06-15 at 07:43 AM (Breakfast, Weekday)”, and visible foods scrambled egg, whole-meal toast, butter. The corresponding nutritional labels are 345 kcal, 14 g protein, 28 g carbs, 18 g fat, and 165 g portion (Coburn et al., 9 Jul 2025).

A notable limitation is that inter-annotator agreement is not explicitly reported in the paper. Instead, quality control relied on accredited dietitians and established protocols rather than formal agreement metrics (Coburn et al., 9 Jul 2025). This is an important qualification for methodological interpretation: the benchmark emphasizes expert curation and verification, but does not provide a formal reproducibility statistic for annotation consensus.

4. Metadata design and representation

A central feature of ACETADA is its explicit treatment of contextual metadata. The report distinguishes three metadata sources.

GPS-derived metadata include reverse-geocoded venue type and name (e.g., café, home, restaurant) as well as city and country. Timestamp-derived metadata include exact date and time (human-readable), meal type label (Breakfast, Lunch, Dinner), and day type (e.g., Weekday vs. Weekend). Food-presence metadata consist of a dietitian-verified comma-separated list of visible items per meal (Coburn et al., 9 Jul 2025).

In prompting experiments, metadata are concatenated as natural-language fragments. In downstream modeling, metadata can be encoded as categorical one-hot vectors (meal type, venue type) or normalized embeddings (e.g., time-of-day sine/cosine, location embeddings) (Coburn et al., 9 Jul 2025). This bifurcation is methodologically important because it separates prompt-level contextualization from architecture-level feature engineering.

The paper’s broader argument is that the interpretation of contextual metadata can enhance LMM performance in estimating nutritional values. Within ACETADA, metadata are not auxiliary bookkeeping fields; they are benchmark variables explicitly intended to alter inference quality. A plausible implication is that the dataset operationalizes dietary assessment as a multimodal reasoning problem in which semantic scene context and behavioral regularities—such as meal timing and venue priors—can constrain plausible nutritional outputs.

5. Statistical characteristics and label space

The dataset’s target labels are reported in fixed units: Energy in kilocalories (kcal), Macronutrients in grams (g), and Portion size in grams (g) (Coburn et al., 9 Jul 2025). The per-meal distributions are summarized as follows:

Variable	Distribution summary	Unit
Energy	mean ≈ 600, SD ≈ 240; range ≈ 100 – 1 200	kcal
Carbohydrates	mean ≈ 60, SD ≈ 30; range ≈ 10 – 150	g
Fat	mean ≈ 25, SD ≈ 15; range ≈ 2 – 65	g
Protein	mean ≈ 20, SD ≈ 12; range ≈ 3 – 75	g
Portion	mean ≈ 550, SD ≈ 200; range ≈ 100 – 1 200	g

The metadata distribution is also characterized. Meal type counts are Breakfast 290, Lunch 258, Dinner 258. Venue types (approx.) are Home 45 %, Café 25 %, Restaurant 15 %, and Others 15 %. The corpus spans 11 cuisine categories, with no one class exceeding 25 % of meals (Coburn et al., 9 Jul 2025).

These statistics indicate moderate heterogeneity in both nutritional and contextual dimensions. The range of Energy from approximately 100 to 1 200 kcal and the portion range from approximately 100 to 1 200 g imply a broad but bounded regression space. The venue and cuisine distributions suggest some diversification of eating contexts without a single cuisine dominating the sample. This suggests that ACETADA is structured to evaluate contextual robustness rather than only canonical meal recognition.

The report also provides the macronutrient-based calorie approximation:

$$\mathrm{Calories} \approx 4 \times \mathrm{Protein\ (g)} + 4 \times \mathrm{Carbohydrates\ (g)} + 9 \times \mathrm{Fat\ (g)}$$

This formula functions as a nutritional consistency relation rather than a benchmark metric. It provides a useful interpretive link between the four nutrient labels and the energy label.

6. Evaluation protocol, metrics, and reported benchmark behavior

ACETADA underpins an evaluation across eight LMMs (four open-weight and four closed-weight) (Coburn et al., 9 Jul 2025). The evaluation first establishes the benefit of contextual metadata integration over straightforward prompting with images alone, and then examines how contextual information interacts with reasoning modifiers such as Chain-of-Thought, Multimodal Chain-of-Thought, Scale Hint, Few-Shot, and Expert Persona.

Two error metrics are defined:

$$\mathrm{MAE}(y) = \frac{1}{n} \sum_{i=1}^{n} | y_i - \hat{y}_i |$$

$$\mathrm{MAPE}(y) = \frac{100\%}{n} \sum_{i=1}^{n} \left| \frac{y_i - \hat{y}_i}{y_i} \right|$$

The reported baseline and metadata-enriched results focus especially on caloric and portion estimation. Baseline image-only MAE (calories) across eight LMMs is reported as ≈ 270 kcal. The best metadata-enriched MAE (calories) across models is ≈ 70 kcal, corresponding to an average reduction ≈ 200 kcal. Portion-size MAE is reported to be reduced by ≈ 50 g on average when including GPS and timestamp metadata. For reasoning modifiers, the report states that Expert Persona can further benefit from ≈ 285 kcal MAE down to ≈ 75 kcal when combined with full metadata (Coburn et al., 9 Jul 2025).

The abstract also states that integrating metadata intelligently, when applied through straightforward prompting strategies, can significantly reduce both MAE and MAPE in predicted nutritional values (Coburn et al., 9 Jul 2025). In benchmark terms, ACETADA is therefore designed not only to test raw visual recognition but also to quantify the value of contextual conditioning. A common misconception would be that meal-image nutrition analysis is primarily an image-understanding problem; the reported results instead support a context-aware formulation in which location, time, and food-presence cues materially affect error rates.

7. Access, release configuration, and research significance

The initial release is not pre-split. Researchers are encouraged to apply standard splits (e.g., 70/15/15) or cross-validation (Coburn et al., 9 Jul 2025). This choice preserves flexibility for benchmarking, though it places responsibility for split design and evaluation comparability on downstream users.

The planned release is under an open license (e.g., CC BY 4.0) via GitHub and institutional data archive, and is described as including raw images, annotation files (CSV/JSON), and preprocessing code. The repository URL is given as https://github.com/ACETADA-dietary-benchmark, to be activated upon publication (Coburn et al., 9 Jul 2025).

In research terms, ACETADA occupies a specific niche: it joins dietitian-verified annotations, before/after meal imaging, fiducial-marker scale cues, and contextual metadata in a single benchmark for LMM evaluation. Its significance lies less in raw scale than in the alignment of measurement protocol with multimodal prompting research. The dataset supports comparisons between closed-weight and open-weight systems, and between image-only and metadata-enriched inference strategies. This suggests that ACETADA is particularly suitable for studying whether nutritional estimation performance depends on explicit contextual priors rather than visual evidence alone.

At the same time, several constraints remain explicit in the report: inter-annotator agreement is not explicitly reported, and the dataset is not pre-split in the initial release (Coburn et al., 9 Jul 2025). Those properties do not diminish its utility, but they shape how results derived from ACETADA should be interpreted and compared. Within those bounds, the benchmark provides a structured basis for developing and evaluating context-aware nutrition-analysis systems grounded in dietitian-verified ground truths (Coburn et al., 9 Jul 2025).