City in Sight
城市在眼前
Research on urban visibility, perception, and visual experience of the city.
研究城市可见性、感知与城市视觉体验。
2023 - present
Most urban analytics still look at the city from the outside—satellite tiles, street networks, land-use parcels. City in Sight asks a complementary question: what does the city look and feel like from where people actually live, walk, and dwell? The programme treats urban visual experience as something that can be measured, mapped, and linked back to planning and design—not as a decorative afterthought to morphology or property markets.
From an information-theoretic view, the city is a complex system continuously sampled by countless partial observers. No single vantage captures the whole. A resident at a window, a pedestrian on a street—each obtains a local, channel-specific readout of urban form and atmosphere. Everyday urban life is therefore a vast, distributed observation process: the built environment as latent state, visual channels as measurement operators, human perception as structured but noisy sensing.
A compact formalisation makes this precise. Let $X(\mathbf{s})$ denote the urban environment at location $\mathbf{s}$. Observer $i$, anchored at $a_i$, samples through a visual channel $c_i \in \lbrace\mathrm{window}, \mathrm{street}, \ldots\rbrace$ and produces a perceptual response $Y_i = h(g(X, a_i, c_i)) + \varepsilon_i$. Each observation is necessarily partial—it informs a neighbourhood of the city, not the system as a whole. Over time, urban life generates a stream $\mathcal{O}={(a_i,c_i,I_i,Y_i)}$. City in Sight asks how to recover a city-wide perceptual field $\hat{P}(\mathbf{s},d)$ from $\mathcal{O}$: the expected human response at $\mathbf{s}$ along dimension $d$ (preference, vividness, thermal affordance, and others), inferred by propagating sparse judgments through imagery and spatial structure.
To do this, we combine three ingredients: image sources that reflect lived experience (real-estate window views, street view imagery along pedestrian networks); rigorous human-perception data (crowdsourced pairwise comparisons, survey-based visual indicators); and interpretable models that scale local labels to urban maps while preserving compositional meaning.
Two strands anchor the work so far. From the window, we study how city landscapes read through domestic views—how floor level, sky, greenery, and built form shape preference and vividness across a metropolis. Along the street, we study how the visible streetscape signals thermal comfort potential (thermal affordance) and related visual-perceptual cues. Together, vertical and horizontal sight lines offer dual portals into urban experience—a case for treating what is in sight alongside geometry, accessibility, and environmental simulation.
Related empirical work is listed below.
大多数城市分析仍从外部俯视城市——卫星图、路网、用地斑块。城市在眼前提出一个互补问题：从人们真实居住、行走与停留的位置看，城市是什么样、感受如何？该研究将城市视觉体验视为可测量、可制图、可回馈规划与设计的对象，而非形态或房价分析的装饰性附庸。
从信息论视角看，城市是一个被无数局部观察者持续采样的复杂系统。没有任何单一视点能捕捉全貌。窗边的居民、街道上的行人——各自通过特定的视觉通道，获得关于城市形态与氛围的局部读数。日常城市生活因此构成一个庞大的分布式观测过程：建成环境是 latent state，视觉通道是测量算子，人类感知是结构化但含噪的 sensing。
一个简洁的形式化表述如下。设 $X(\mathbf{s})$ 表示位置 $\mathbf{s}$ 处的城市环境。观察者 $i$ 锚定于 $a_i$，通过视觉通道 $c_i \in \lbrace\mathrm{window}, \mathrm{street}, \ldots\rbrace$ 采样，产生感知响应 $Y_i = h(g(X, a_i, c_i)) + \varepsilon_i$。每次观测必然是局部的——它告知城市的某个邻域，而非整个系统。随时间推移，城市生活生成数据流 $\mathcal{O}={(a_i,c_i,I_i,Y_i)}$。城市在眼前要解决的问题是：如何从 $\mathcal{O}$ 恢复城市尺度的感知场 $\hat{P}(\mathbf{s},d)$——即位置 $\mathbf{s}$ 沿维度 $d$（偏好、生动性、热舒适可供性等）的期望人类响应，通过稀疏判断在影像与空间结构上的传播进行推断。
为此，我们结合三类要素：反映生活体验的影像来源（房产窗景、沿步行网络的街景）；严格的人类感知数据（众包成对比较、调查型视觉指标）；以及可扩展至城市地图、同时保留组合语义的 interpretable 模型。
目前已有两条主线。从窗户，研究城市景观如何通过居家视窗被阅读——楼层、天空、绿化与建成形态如何塑造整座城市中的偏好与生动性。沿街道，研究可见街景如何信号化热舒适潜力（thermal affordance）及相关视觉—感知线索。垂直与水平的视线共同构成进入城市体验的双重入口——主张将眼前所见与几何、可达性与环境模拟并列对待。
相关实证工作见下文。
