Models trained on crowdsourced street-view ratings can predict fairly well whether people will judge a scene safe, lively, or depressing. For a city planner, though, a score without a reason is close to useless. The planner needs to know which part of the scene, and in particular which instance of an ordinary object such as a wall, a sidewalk or a row of trees, is pushing the judgment up or down. Saliency maps are heatmaps that assign each pixel a share of the model's output. They look like the right tool for this. Yet the one serious attempt to use them on urban disorder found that its object rankings were driven mostly by the semantic segmentation, the per-pixel object labels, and only weakly by the heatmap itself (UPDExplainer). If that holds in general, saliency adds little beyond counting how much of the image is "wall" or "vegetation."
This note proposes to settle that question directly. We will attribute a pairwise-ranking perception model's continuous score to image regions using gradient path attribution, with a reference set drawn from the same city. The claim is that the resulting per-object contributions carry information that object area cannot carry, namely which particular wall or tree people respond to. We test this with new human judgments on image pairs that have been matched to have the same object composition, so that area-based explanations are blind by construction.
Core Claim
Gradient-based attribution of a ranking model's perception score, pooled inside segmentation masks, identifies within-category appearance cues that drive human pairwise judgments. These are cues such as a graffitied versus clean wall, or ragged versus maintained vegetation. Segmentation area and Score-CAM-style object rankings do not capture them. The claim is falsifiable in two ways. First, on composition-matched pairs, the model's preferred image should agree with the human majority well above chance. Second, if we blur the single object class the attribution names as responsible, human preference should collapse toward 50/50 far more than when we blur an equal-area class chosen at random. If either test fails, the skeptics are right and segmentation is the whole story.
Research Program
Perception model. We train a ranking network in the style of the Ranking Siamese CNN from Deep Learning the City. That network jointly learns from pairwise human choices and a scalar ranking score $f(x)$ on Place Pulse 2.0, which contains 110,988 images and 1.17M comparisons across six attributes. We use a ResNet50 backbone so that we can reuse multi-layer attribution analysis validated on that architecture (There and Back Again). Explanations from a weak model would be uninformative, so the model must at least match the reported 73.5% pairwise accuracy on "safe" before we attribute anything.
Attribution. Most saliency methods are defined for a softmax class score. Ours explains the scalar ranking score $f$. We use Expected Gradients (Axiomatic Attribution Priors and Expected Gradients), the reference-distribution form of Integrated Gradients (Axiomatic Attribution for Deep Networks). The reference images $x'$ are drawn from $D_c$, the set of images from the same city $c$:
$$\phi_i(x) = \mathbb{E}_{x' \sim D_c,\ \alpha \sim U(0,1)}\left[(x_i - x'_i)\,\frac{\partial f(x' + \alpha(x - x'))}{\partial x_i}\right], \qquad \sum_i \phi_i(x) = f(x) - \mathbb{E}_{x' \sim D_c} f(x').$$
Here $\phi_i$ is pixel $i$'s contribution, and the second identity is the completeness property inherited from Integrated Gradients. Because of completeness, summing $\phi$ inside each segmentation mask $k$ gives object-class contributions $\Phi_k$. These add up exactly to "how much safer this scene looks than a typical scene in its city," and that is the quantity a local planner cares about. We also chose this attribution method for a second reason. Saturating gradients are a plausible issue for perception scores that plateau, since saturation effects are pervasive in well-being determinants (systematic review). Path integration was designed for exactly that case.
We acknowledge one cost of the city-conditioned reference. Interpolating between two unaligned photographs produces off-manifold blends. For that reason, blurred and black references are run as variants, so that we can see whether the choice of reference changes the conclusions.
For small cues such as litter or cracks, we add NormGrad multi-layer fusion as a second gradient method (There and Back Again). That work showed single-layer Grad-CAM washes out early-layer detail. Masks come from a Cityscapes-trained segmenter (Cityscapes, 67.1% mIoU for Dilated10). Cityscapes images come from 50 cities, mostly European, so we will audit mask quality on a hand-labelled sample of 300 Place Pulse images before relying on the masks globally.
Composition-matched pairs. For each attribute, we sample about 400 same-city image pairs whose segmentation class fractions match within a few percentage points per class. Within each pair, the model must predict a clear score gap. We then collect about 25 fresh pairwise judgments per pair. The large number is deliberate. Inter-rater reliability on place appearance averaged only 0.35, and roughly 22–32 raters per item were estimated as necessary for stable ground truth (Using human raters…).
Baselines
The naive reference point is a ridge or gradient-boosted regression from segmentation class fractions to each image's ranking score. It is the tabular-plus-segmentation approach common in the well-being literature, as surveyed in the systematic review, and it is the approach the UPDExplainer ablation implies may be sufficient. It is trained on the same Place Pulse splits and given the same hyperparameter search budget as the deep model. No reported result found on Lacuna for segmentation-fraction regression on Place Pulse 2.0 pairwise accuracy, so we will report it ourselves. On matched pairs it should sit near 50% by construction. Its value lies in the unmatched evaluation, where it shows how much of the deep model's accuracy area alone already explains.
The strongest existing explanation pipeline is UPDExplainer's design: a heatmap from Score-CAM, a forward-pass masking variant of class activation mapping, merged with segmentation into per-object "activation density" (UPDExplainer). On its own disorder benchmark it reports 79.89% detection accuracy and 87.76% Top@1 ranking precision. We re-run it on our backbone and our continuous ranking score, so that only the attribution step differs. Single-layer Grad-CAM (Grad-CAM) is run alongside it, again on the same backbone and masks.
The most direct competitor is segment occlusion: replace each segment with a blurred version and record the score drop. It is also given the same number of forward passes as our attribution budget.
Two ablations isolate our key ingredients. The first is identical attribution with a black reference instead of the city reference. The second is attribution without multi-layer fusion.
Beating these baselines means two things. Our top-named segment must produce a significantly larger collapse in human preference when blurred than Score-CAM's or Grad-CAM's top segment does, assessed with a paired bootstrap over pairs. And it must at least tie occlusion.
Why not just…?
Why not just regress perception on segmentation area? This is the real threat. UPDExplainer's own ablation found that segmentation quality, not the heatmap, dominated its ranking accuracy. Our matched-pair design is built so that this alternative cannot explain any human preference that remains. If humans show no consistent preference on matched pairs, or the model cannot predict that preference, we will report that area suffices for these attributes. That would be a useful negative result for the field.
Why not just use Score-CAM, as UPDExplainer did? It works as a post-hoc map. However, it has no completeness property, so its per-object scores do not add up to a score difference a planner can budget against. It also cannot later serve as a differentiable training constraint of the kind used in Where Not to Learn. Whether its rankings are less faithful is an empirical question, and the human blur-collapse test answers it head to head.
Why not just occlude segments directly? Occlusion is simple and directly operates at the object level, and it may do as well on the model-side test. Its weaknesses are that a blurred or gray patch is itself an unusual visual stimulus, and that occluding one segment at a time ignores how segments interact. If occlusion matches gradient attribution on the human test, we will recommend occlusion and keep gradients only for the multi-layer small-cue analysis. That is an outcome we are prepared to report.
Proposed Experiments
The first check is predictive parity: our ranking model's pairwise accuracy on all six Place Pulse attributes, compared against 73.5% on "safe."
The second is model-side faithfulness. For each method, we blur the segments it ranks highest and measure the deletion curve of $f$, using all methods on identical masks.
The third and decisive check is the human study on composition-matched pairs. We measure two things there. One is agreement between the model's choice and the human majority. The other is the drop in human preference margin when the top-attributed segment class is blurred in both images, compared with blurring a random equal-area class. We hypothesize that the gap will be largest for "depressing" and "safe," where disorder cues lie inside walls and sidewalks, and smallest for "wealthy," where composition may dominate.
The fourth is a robustness check. We repeat the analysis with black and blurred references and across held-out cities. If the identity of the top-named segment changes with the reference choice, the explanations are not stable enough to act on, and we will say so.
References
- UPDExplainer: an Interpretable Transformer-based Framework for Urban Physical Disorder Detection Using Street View Imagery
- Deep Learning the City : Quantifying Urban Perception At A Global Scale
- There and Back Again: Revisiting Backpropagation Saliency Methods
- Improving performance of deep learning models with axiomatic attribution priors and expected gradients
- Axiomatic Attribution for Deep Networks
- Assessing environmental determinants of subjective well-being via machine learning approaches: a systematic review
- The Cityscapes Dataset for Semantic Urban Scene Understanding
- Using human raters to characterize the psychological characteristics of GPS-based places
- Grad-CAM: Why did you say that?
- Where Not to Learn: Prior-Aligned Training with Subset-based Attribution Constraints for Reliable Decision-Making