---
title: 'Mean Object Hue (MoH): Safety Perception Metric'
url: https://www.emergentmind.com/topics/mean-object-hue-moh
type: topic
---

# Mean Object Hue (MoH): Safety Perception Metric

Mean Object Hue (MoH) is a pixel-level attention score for each semantic object, normalized and inverted so that higher values directly indicate stronger human focus. In the study of human safety perception in urban street-view imagery, MoH was introduced to move beyond object presence and quantify how much attention each object actually received, using hue values derived from human gaze heatmaps and semantic labels from ADE20K-trained segmentation. Within the broader framework combining eye-tracking systems, street view images, and deep learning, MoH was used alongside Mean Object Ratio in Highlighted Regions (MoRH) to identify which visual environmental factors attract attention when participants judge whether one street scene appears safer than another [2509.25457].

## 1. Conceptual role in safety-perception analysis

MoH was formulated in a setting where volunteers in Helsingborg, Sweden, viewed pairs of street view images and indicated which image appeared safer. Eye-tracking systems were used to quantify not only what users were looking at but also how long they engaged with specific environmental elements. In that context, MoH addressed a distinct analytical question: not merely whether an object was present in a scene, but how intensely it was attended at the pixel level [2509.25457].

The metric therefore complements object-ratio approaches. MoRH tells us “how often an object appears within the top \(t\) percentile of gaze,” whereas MoH goes a step further and measures “how intense was the attention to each object?” This distinction is central to the interpretation of the results. An object can occupy a large image area or appear frequently in highlighted regions without receiving strong per-pixel focus. MoH was designed precisely to capture that difference [2509.25457].

## 2. Formal definition and mathematical structure

For a given object class \(o\), let \(H(i,j)\) be the hue value at pixel \((i,j)\) of the human-attention heatmap after rescaling into the \(0\ldots150\) range, where \(0\) denotes strongest attention and \(150\) denotes weakest attention. Let \(p(i,j)\) denote the semantic category at that pixel as determined by the ADE20K-trained segmentation. The raw mean hue is defined as

$$
\mathrm{MoH}_{\mathrm{raw}}(o)=
\frac
{\sum_{i=1}^{h}\sum_{j=1}^{w}\mathds{1}\bigl(p(i,j)=o\bigr)\times H(i,j)}
{\sum_{i=1}^{h}\sum_{j=1}^{w}\mathds{1}\bigl(p(i,j)=o\bigr)}.
$$

Because lower hue means higher attention, the raw value is inverted into a more interpretable attention score by subtracting from the maximum hue value:

$$
\mathrm{MoH}(o)=150-\mathrm{MoH}_{\mathrm{raw}}(o).
$$

This formulation makes the direction of the metric explicit. Higher MoH indicates stronger attention, not weaker attention. The normalization divides by the pixel count of object \(o\), so MoH measures per-pixel average attention rather than total occupied area. A common source of confusion is to treat MoH as an object-prevalence statistic; the normalization shows that it is instead an attention-intensity statistic [2509.25457].

## 3. Computational workflow

The computation of MoH proceeds through a fixed sequence. Gaze points are first recorded from all participants viewing a street-view image pair. The resulting gaze fixations are aggregated into a continuous heatmap, and raw density is converted into an HSV hue channel in the \(0\ldots150\) range. A dense-prediction transformer (DPT) is then run to label each pixel with one of 150 ADE20K object categories. For each object \(o\) in an image, all pixels satisfying \(p(i,j)=o\) are collected, the average hue \(H(i,j)\) over those pixels is computed to obtain \(\mathrm{MoH}_{\mathrm{raw}}\), and the value is inverted via \(150-\mathrm{MoH}_{\mathrm{raw}}\) to yield MoH. The procedure is repeated over all images in the “safe” and “unsafe” groups, and the resulting values are averaged across images to obtain \(\overline{\mathrm{MoH}(o)}\) [2509.25457].

The hue channel itself is central to interpretation. It is derived from human-attention heatmaps and linearized to \(0\ldots150\) for visualization. Red and yellow at the low end denote high attention, whereas green and blue at the high end denote low attention. Because the final MoH score inverts this scale, the reported object-level quantity aligns directly with attention strength: larger values denote stronger human focus [2509.25457].

## 4. Semantic categories and measurement scope

MoH is defined over the 150 urban-scene categories used in ADE20K segmentation. The examples explicitly reported include “stairway,” “car,” “flag,” “sky,” “chairs,” “railings,” “bus,” “truck,” “plant,” and “sidewalk.” The metric is therefore object-specific at pixel resolution rather than scene-global. Each semantic class is evaluated by aggregating the heatmap hue over all pixels assigned to that class [2509.25457].

This design gives MoH a different scope from coarse scene descriptors. Because the score is computed after dense semantic labeling, it can separate attention to distinct environmental elements even when they are spatially proximate. The result is a human-centered characterization of which urban features attract gaze during safety judgments. A plausible implication is that MoH is especially useful when large objects with low attentional salience coexist with smaller objects that receive concentrated fixation, since the per-pixel averaging and inversion preserve that contrast.

## 5. Reported empirical patterns in safe and unsafe image groups

Figure 8, as summarized in the reported account of the metric, gives approximate rankings of MoH in “safe” and “unsafe” image groups. In safe images, objects with the highest MoH included stairways at approximately 140, floors at approximately 135, railings at approximately 130, booths at approximately 125, and flags at approximately 123. In unsafe images, the highest MoH values were reported for buses at approximately 145, vans at approximately 142, cars at approximately 140, fences at approximately 128, and hovels at approximately 125. By contrast, “sky” and “ground/land” had very low MoH values, below 20 [2509.25457].

| Image group | High-MoH objects | Low-MoH objects |
|---|---|---|
| Safe | stairways, floors, railings, booths, flags | sky, ground/land |
| Unsafe | buses, vans, cars, fences, hovels | sky, ground/land |

These rankings were tied directly to safety perception. Objects that attained high MoH in the safe image set, including stairways, floors, railings, public-space furniture, and flags, positively correlated with participants’ choices for safety. In the unsafe set, transportation vehicles such as buses, vans, and cars dominated MoH, reflecting participants’ tendency to associate heavy traffic with unsafety. Objects with systematically low MoH, including sky and expanses of ground, appeared to play little direct role in shaping safety judgments even when they occupied large image areas [2509.25457].

These results also clarify what MoH is not measuring. Low values for sky and ground/land do not imply visual absence; instead, they indicate weak per-pixel attentional engagement. Conversely, high values for infrastructure, furniture, or vehicles indicate concentrated gaze rather than merely frequent occurrence.

## 6. Relation to MoRH and comparison with XAI saliency maps

MoH was analyzed jointly with MoRH and with CAM-based saliency maps. The two object-level metrics are complementary. MoRH measures how often an object appears within highly attended regions, while MoH measures the intensity of attention directed to that object. The reported comparison shows that some objects, such as ceilings or sky, may have high MoRH but low MoH, or vice versa. This means that frequent inclusion in highlighted regions does not guarantee strong per-pixel focus [2509.25457].

The study also compared real human attention from eye-tracking systems with attention maps obtained from explainable artificial intelligence. Seven CAM variants were examined. XGradCAM and EigenCAM correlated best with human MoH, both at the scene level, where they achieved the lowest \(L2\)/LPIPS, and at the object level, where they achieved the highest cosine similarity of MoH vectors. Even so, detailed examination showed a residual gap: the best XAI maps tended to highlight broader regions, whereas human MoH pinpointed a few critical object patches [2509.25457].

This residual gap is significant for the interpretation of saliency-based explanations. It suggests that alignment at coarse scene level or by aggregate similarity metrics does not eliminate finer discrepancies in what humans actually inspect. In the safety-perception setting, MoH thus functions not only as an attention metric but also as a benchmark for evaluating whether XAI saliency maps reproduce human perceptual patterns.

## 7. Interpretive significance and limitations of scope

MoH supports a more human-centered understanding of which environmental elements influence human safety perceptions. By combining semantic segmentation with gaze-derived hue encoding, it identifies object classes that attract strong human attention and distinguishes them from classes that are visually present but weakly implicated in the perceptual decision. In the reported application, this yielded the conclusion that certain urban infrastructure and public-space features draw more human attention, while the sky is less relevant in influencing safety perceptions [2509.25457].

At the same time, the metric should be interpreted within its defined scope. It is computed from gaze heatmaps derived from participants viewing street-view image pairs, and its meaning depends on the hue remapping, per-pixel normalization, and inversion convention. It is therefore not a generic measure of urban quality, object frequency, or semantic importance in isolation. Rather, it is a scene-conditioned measure of human attentional intensity over segmented objects during safety judgments. This suggests that its principal contribution lies in linking human gaze behavior, semantic object structure, and comparative image-level safety perception within a single quantitative framework [2509.25457].

Source: https://www.emergentmind.com/topics/mean-object-hue-moh