Distributional Pixel Agreement (DPA)

• Distributional Pixel Agreement (DPA) is a metric that quantifies the agreement between predicted and ground-truth pixel proportions for each semantic class, emphasizing distributional accuracy over precise spatial alignment.
• It computes per-class scores by normalizing predicted and actual pixel counts to proportions, then evaluates the absolute difference, with weighted averaging used to aggregate scores across classes.
• DPA is particularly applied in plate-level food waste estimation, where accurate area proportions of food classes are more relevant than the spatial configuration of each class.

Distributional Pixel Agreement (DPA) is a distribution-based semantic segmentation metric that evaluates whether a predicted mask assigns approximately the correct proportion of pixels to each class, rather than whether it localizes those pixels in the correct spatial positions. In the food-waste case study that explicitly introduces the metric, the paper names it Distributional Pixel Accuracy (DPA) while also characterizing it as emphasizing distributional agreement rather than spatial alignment; in that usage, DPA is a custom metric for plate-level food waste estimation from semantic segmentation masks (Rokhva et al., 19 Jul 2025). The central idea is that, for applications driven by class-wise surface coverage rather than precise geometry, agreement in pixel-frequency histograms can be more task-aligned than overlap-based measures such as IoU or Dice.

1. Definition and conceptual scope

DPA is defined on predicted and ground-truth segmentation masks with the same label set. For a class $c$, it compares the predicted proportion of pixels assigned to $c$ with the ground-truth proportion of pixels assigned to $c$, ignoring spatial arrangement. In the formulation given in the food-waste study, the per-class score is

$$\mathrm{DPA}_c = 1 - \left| \frac{p^{\text{pred}}_c}{T^{\text{pred}}} - \frac{p^{\text{gt}}_c}{T^{\text{gt}}} \right|,$$

where $p^{\text{pred}}_c$ is the number of pixels predicted as class $c$, $p^{\text{gt}}_c$ is the number of pixels labeled as class $c$ in the ground truth, and $T^{\text{pred}}, T^{\text{gt}}$ are the total numbers of pixels in the predicted and ground-truth masks, respectively (Rokhva et al., 19 Jul 2025).

This definition makes DPA a class-proportion agreement measure. If the predicted and true class proportions coincide exactly, then $\mathrm{DPA}_c = 1$. If they differ by 0.2, then $c$0. The metric therefore evaluates whether the segmentation preserves the correct area fraction of each class within an image.

The paper’s motivation is explicit: the goal is not “structurally perfect segmentations” but estimation of “the proportions (proportional distribution) of different classes within each segmentation mask.” DPA is therefore designed for settings in which the downstream variable of interest is the fraction of image area occupied by each semantic component, not the exact location or boundary of that component (Rokhva et al., 19 Jul 2025).

2. Mathematical formulation and aggregation

For a single image, DPA is computed by first extracting per-class pixel counts from the predicted and ground-truth masks, then normalizing those counts into class proportions. The paper describes the pipeline in the following sequence: compute $c$1 and $c$2, compute totals $c$3 and $c$4, form the proportions $c$5 and $c$6, and finally evaluate $c$7 (Rokhva et al., 19 Jul 2025).

The reported scores are not limited to per-class values. The same source states that experimental results use weighted-averaged DPA over classes, and then over the dataset, in the same weighted sense used for IoU and Dice. For an image, if $c$8 is a class weight proportional to the ground-truth frequency of class $c$9, normalized so that $c$0, the aggregate image-level score is

$c$1

The evaluation set $c$2 excludes background in reported results, although the paper notes that background exists in practical deployment. Reported dataset-level values are then averages of image-level weighted DPA values, or equivalently weighted aggregates over the dataset (Rokhva et al., 19 Jul 2025).

Two implementation choices are especially important. First, the authors explicitly prioritize weighted rather than macro averaging, arguing that in food-waste estimation “weighted metrics are more appropriate” because macro averages can over-penalize trivial errors in underrepresented classes. Second, reported metrics exclude the background class for clarity, so DPA primarily reflects agreement on food classes rather than on plate or scene background (Rokhva et al., 19 Jul 2025).

3. Relation to Pixel Accuracy, IoU, and Dice

DPA differs fundamentally from conventional segmentation metrics because it does not require spatial coincidence between predicted and ground-truth regions. Pixel Accuracy counts correctly classified pixels and is heavily influenced by majority classes. IoU and Dice both depend on true positives and therefore require spatial overlap. DPA instead compares class-frequency histograms, or more precisely, class-wise normalized pixel counts (Rokhva et al., 19 Jul 2025).

This divergence is most apparent in the paper’s three qualitative regimes. In the first regime, when prediction and ground truth are identical, Pixel Accuracy, IoU, Dice, and DPA all equal 1. In the second regime, with partial alignment, all metrics lie strictly between 0 and 1. In the third regime, which the paper calls an “extreme and rare” case, there is no spatial overlap but the class proportions are identical. In that case, IoU and Dice are 0, Pixel Accuracy is close to 0, yet DPA remains close to 1 because the predicted and true class proportions match (Rokhva et al., 19 Jul 2025).

This behavior is not an accident or a weakness relative to the stated objective; it is the metric’s defining property. The paper explicitly states that DPA “does not care where class $c$3 is located in the image” and instead asks whether the model assigns about the right fraction of the image area to that class. A plausible implication is that DPA is best understood not as a replacement for overlap metrics in general segmentation, but as a task-specific complement whose utility depends on whether downstream inference is driven by area proportions rather than by spatial configuration.

Empirically, the study reports that DPA values are consistently higher than IoU and Dice for the same model-category pair. The authors interpret this as expected because DPA is less stringent: small boundary errors or local misalignments can degrade IoU or Dice while barely affecting class histograms. For challenging categories such as visually ambiguous stews, DPA remains relatively high even when IoU and Dice are lower, provided that total class areas are still estimated reasonably well (Rokhva et al., 19 Jul 2025).

4. Role in plate-level food waste estimation

The clearest explicit use of DPA is in plate-level food waste estimation from RGB images taken before and after consumption. The application models food waste through surface-area proportions of semantic classes such as rice, stew, meat, or fries. In that setting, the decisive quantity is not where a component appears on the plate, but how much of the plate area it occupies before and after eating (Rokhva et al., 19 Jul 2025).

The study defines class proportion for class $c$4 in a ground-truth image as

$c$5

then computes average pre-consumption and post-consumption proportions $c$6 and $c$7, an eating rate

$c$8

an average eating rate over dishes,

$c$9

and finally a remaining rate

$$\mathrm{DPA}_c = 1 - \left| \frac{p^{\text{pred}}_c}{T^{\text{pred}}} - \frac{p^{\text{gt}}_c}{T^{\text{gt}}} \right|,$$0

These formulas depend directly on class-wise pixel proportions, which is precisely what DPA measures (Rokhva et al., 19 Jul 2025).

Within the evaluation protocol, the models are trained with a capped dynamic inverse-frequency cross-entropy loss and selected by highest validation weighted IoU, but DPA is treated as the primary task-relevant indicator at test time. The paper states that DPA is “the most important” metric for the downstream objective and that it “better reflect[s] the nature of this task” because it prioritizes pixel class frequencies and proportions over spatial alignment (Rokhva et al., 19 Jul 2025).

The reported practical benchmark is also explicit: DPA $$\mathrm{DPA}_c = 1 - \left| \frac{p^{\text{pred}}_c}{T^{\text{pred}}} - \frac{p^{\text{gt}}_c}{T^{\text{gt}}} \right|,$$1 is interpreted as strong alignment in pixel-wise proportion estimates. The abstract and discussion state that for each food type, at least one model approached or surpassed 90% DPA, and that high DPA values, mostly exceeding 0.9 in some models, support the accuracy of downstream estimates such as consumption and waste rates (Rokhva et al., 19 Jul 2025).

5. Interpretation, limitations, and appropriate use

DPA should be interpreted narrowly and task-dependently. A value close to 1 indicates that the model’s predicted pixel-wise class proportions closely match the ground truth; in the food-waste setting, this suggests that downstream percentage estimates of remaining food should be reliable, subject to the independent limitation that the measurements are derived from 2D surface coverage rather than 3D volume (Rokhva et al., 19 Jul 2025).

The paper is equally explicit about the metric’s limits. It states that DPA “is not suitable for applications requiring high spatial precision, such as medical imaging, autonomous driving, or any other safety-critical or sensitive systems.” The reason is direct: DPA ignores shape, boundary quality, and absolute location. A segmentation can therefore score highly under DPA while being spatially poor, so long as class areas are approximately correct (Rokhva et al., 19 Jul 2025).

This has several methodological consequences. DPA is insensitive to object geometry and can understate the severity of misplaced or structurally implausible predictions. Weighted aggregation also means that majority classes dominate the overall score, although the paper mitigates one obvious failure mode by excluding background from reported metrics. A plausible implication is that DPA should be paired with overlap- or boundary-sensitive metrics whenever spatial correctness matters, even if area estimation is the main operational target.

The same source consistently positions DPA in exactly this way: not as a universal segmentation metric, but as a metric whose adequacy depends on whether “class-wise pixel distributions before and after food consumption accurately mirror real-world conditions.” In such applications, DPA is deliberately robust to scattered small mislabels, minor boundary jitter, and post-consumption visual ambiguity, all of which can distort IoU and Dice more than they distort area totals (Rokhva et al., 19 Jul 2025).

6. Broader research context and acronym ambiguity

The acronym DPA is not unique across arXiv literature, and this creates a recurring source of confusion. In the food-waste segmentation study, DPA denotes the custom metric discussed above (Rokhva et al., 19 Jul 2025). By contrast, in object detection, DPA denotes Dual Probabilistic Alignment, a framework for Universal Domain Adaptive Object Detection that models domain probability as Gaussian distribution and performs Global-level Domain Private Alignment, Instance-level Domain Shared Alignment, and a Private Class Constraint (Zheng et al., 2024). That work is described as a form of “distributional/probabilistic agreement,” but it is not a segmentation metric.

Related segmentation literature nevertheless contains methods that can plausibly be interpreted as distributional pixel agreement mechanisms even when they do not define a metric called DPA. DRSL introduces a “pixel-level distribution regularization scheme” for self-supervised domain adaptation of semantic segmentation, with class-aware multi-modal distribution learning and stochastic mode alignment that enforce agreement between source and target pixel embeddings conditioned on class and mode (Iqbal et al., 2022). DACL similarly defines a dual-agreement consistency loss combining pixel-wise KL divergence with entropy-guided confidence alignment, thereby enforcing agreement between full per-pixel predictive distributions across a CNN and a Transformer under semi-supervised learning (Wang et al., 24 Jun 2026).

A separate and unrelated use appears in unsupervised representation learning, where DPA stands for Distributional Principal Autoencoder rather than Distributional Pixel Agreement. That work concerns distributionally correct reconstruction and score-oriented encoder level sets, not pixel-wise segmentation evaluation (Leban, 17 Feb 2025).

Taken together, these usages suggest that “Distributional Pixel Agreement” occupies two distinct but adjacent conceptual roles in current literature. In the strictest sense, it denotes a simple class-proportion metric for segmentation masks (Rokhva et al., 19 Jul 2025). More broadly, the phrase also resonates with a family of methods that align or regularize per-pixel distributions, pseudo-labels, or domain-conditioned pixel embeddings in segmentation and detection pipelines (Iqbal et al., 2022, Wang et al., 24 Jun 2026, Zheng et al., 2024). A common misconception is to treat these as interchangeable. They are not: one is an evaluation metric based on class-frequency agreement, while the others are training-time alignment mechanisms or entirely different models sharing the same acronym.