When shoppers move online, merchants in historic towns lose foot traffic. A plausible response is to compete harder for the passer-by who is left: bigger signboards, deeper awnings, booths that spill into the street. Each of these is a small, cheap act of physical modernization, and together they erode the historic streetscape that heritage law exists to protect. The opposite story is just as plausible. A merchant who sells through a phone app may stop depending on being seen from the street, and may stop building outward. Two literatures sit on either side of this question without meeting. Work on digital economic displacement documents offline retail losing ground to platforms. Work on heritage documents merchants physically encroaching on historic fabric. Nobody has measured whether the first causes the second, or relieves it.
We propose to build a measuring instrument and then use it. The instrument turns ordinary repeated street-level photographs into a per-facade estimate of the "second contour," the layer of signage, canopies and extensions that merchants add to a historic facade. It is calibrated against terrestrial laser scans of a heritage town. With it, we can track encroachment across many streets and years, and test a pre-registered contrast. The test compares commercial frontage with residential frontage on the same street, before and after local platform commerce arrives. The deliverable is a validated measurement and a causal estimate whose sign settles which story holds. A new segmentation architecture is not the goal. We will adopt whatever vision method passes calibration.
Core Claim
The empirical anchor is Computing Urban Fabric of a Historical Town, which laser-scanned 2,500 m of streets in Lukang, Taiwan. It defined a modification ratio: the projected area of added elements divided by total facade area. The ratio reaches nearly 20% in commercial districts near markets and stays below 5% in residential and preservation districts. The authors attribute the difference to commercial competition and corroborate growth with satellite imagery from 2002 to 2018. That study plays the motivating role here: it establishes that encroachment tracks commercial pressure. It cannot tell us what happens when the nature of that pressure changes, because it never measures digital exposure. Laser scanning is also too expensive to repeat across many towns and years.
The displacement literature supplies the competing mechanisms. How the e-commerce environment affects the economic development of traditional enterprises describes e-commerce removing spatial entry barriers for shoppers. Traditional firms are pushed into marketing innovation and drift toward homogenized, price-driven competition, which is a route to louder storefronts. Reconceptualizing Entrepreneurial Marketing in the Age of Digital Commerce argues that success becomes "platform-embedded." Some firms bypass physical premises entirely, which is a route to decoupling. Both papers are conceptual, and the file notes that the latter lacks large-scale quantitative validation. Does the Internet deserve everybody? warns that platforms can collapse local retail before alternatives take root. That predicts a third outcome: shuttered frontage with a falling ratio for reasons nobody would celebrate.
Our claim is that these three outcomes leave distinguishable signatures in the second contour, and those signatures can be measured from street imagery. The three signatures are intensified encroachment, relief, and abandonment. To make the claim concrete, let $m_{s,t}$ be the modification ratio of street segment $s$ in year $t$. Let $C_s$ indicate commercial frontage, and let $P_{s,t}$ indicate that platform commerce has arrived in the segment's town. The quantity of interest is
$$ \Delta = \big(\bar m_{C=1,\,post} - \bar m_{C=1,\,pre}\big) - \big(\bar m_{C=0,\,post} - \bar m_{C=0,\,pre}\big), $$
the change in commercial encroachment beyond the change on residential frontage of the same towns. Residential frontage serves as the control because the Lukang data show it is largely insensitive to retail pressure. A positive $\Delta$ supports the inverted-pressure story and a negative $\Delta$ supports relief. Abandonment is separated from relief by also recording vacancy, meaning shuttered shopfronts, from the same images.
Research Program
The first component is the instrument. The difficulty is not detecting a sign. It is estimating how much of the facade is added rather than original, on vernacular buildings that break the assumptions most facade parsers rely on. Specialized parsers such as DeepFacade and Facade Segmentation in the Wild build in rectilinear, "Manhattan-world" geometry, meaning walls and openings on a horizontal-vertical grid, along with repetition priors. They report high pixel accuracy on Haussmannian Paris, but the file notes they struggle with irregular, weathered vernacular buildings. These papers constrain the design: grid priors are a liability on our targets.
Two techniques supply the method. Deep Facade Parsing with Occlusions infers structure hidden behind occluders. FacadeNet separates view-dependent elements from invariant wall structure using self-supervised image features. We reuse the idea, not the grid priors. The pipeline treats added elements as a removable layer, estimates the underlying facade behind them, and reports the added area as a fraction of the facade.
The ingredient we expect to matter most is facade-matched temporal differencing. The same frontage is geolocated across years of street imagery, and the instrument reports change in added area relative to that facade's own earlier appearance. Absolute ratios are not compared across architecturally different buildings. Per-building biases from unusual styles, such as a curved eave the model mistakes for an awning, largely cancel in a difference. The causal estimand needs only change, so this ingredient gives up nothing we need. The ablation below tests whether it actually helps.
The second component is calibration. We will render or capture street-level views of the Lukang scanned segments. The point cloud gives per-facade ground-truth modification ratios, and the instrument is scored against them. Access to the Lukang data, or a rescan of comparable streets, is a dependency we would secure through collaboration with the original authors. A small second calibration set will come from the Wuyi facade images in CGAN-Assisted Renovation of the Styles and Features of Street Facades. These 164 images of traditional regional facades are clean, so pasting known signage and canopies onto them gives exact ground truth on a non-Manhattan vernacular style.
The third component is the panel and the causal analysis. We will assemble repeated street imagery for heritage commercial streets with residential stretches. The file names candidates in Lukang, Zhaoqing Fucheng and Haikou Qilou. We will pair this imagery with a town-level arrival date for local platform commerce. That exposure series does not exist in the file and must be built from platform coverage announcements and merchant listings, so its quality is a stated risk. Staggered arrival across towns gives the before/after variation. The commercial/residential split within towns gives the control.
A merchant-level sub-study adds a sharper contrast within commercial frontage. Shops observed to list on platforms, for example through visible delivery-platform signage or a matched online storefront, are compared with neighbouring shops that do not list.
One further finding shapes how we interpret results. Research on the Integration of Cultural Heritage Scenes and Urban "Night Economy" Development found in a survey of 310 visitors that authentic aesthetic and educational experience drives spending, while pure entertainment did not. If so, encroachment is a commons problem. Each merchant's sign helps that merchant and degrades the scene that brings everyone revenue. A positive $\Delta$ would therefore mean digital displacement is quietly accelerating the destruction of the asset that offline merchants most depend on.
Baselines
The naive reference for the instrument predicts every facade's modification ratio as the mean for its district type, roughly 20% commercial and under 5% residential as reported for Lukang. For the causal question, the naive reference assumes no change. Any instrument must beat the district mean in per-facade error to be worth deploying, and must resolve within-type variation, which the district mean cannot.
The first general-purpose competitor is an off-the-shelf semantic segmenter in the DeepLabV3 family, the baseline that Building Facade Parsing R-CNN compares against. It is fine-tuned on the same training images to label signage, awnings and extensions, with added area taken as the labelled pixel fraction. The second is Facade R-CNN itself, the strongest reported street-level facade parser in the file. It reports 53.8% mean intersection-over-union on the Oxford RobotCar facade benchmark, which includes a "shop" class, and 74.34% on RueMonge2014. The third is the occlusion-completion network RPCNet from Deep Facade Parsing with Occlusions, which reports 83.8% mean intersection-over-union on ENPC Art-deco under 50–60% occlusion. RPCNet is the closest existing method to our "remove the added layer" step. No reported result was found on Lukang for estimating the modification ratio from images by any method.
All of these baselines receive the same training images, the same backbone size where architectures allow, and the same hyperparameter search budget. All are scored on the same metric: absolute error in per-facade modification ratio against laser-scan ground truth, plus rank correlation across facades.
The ablation that isolates our key ingredient runs the identical model per-image versus facade-matched across years, and scores error on change in ratio. Beating the baselines means lower change-error than the best of them on Lukang and on the pasted-signage Wuyi set. The bar is set in advance: error small enough to detect a shift of a few percentage points, well inside the 5-to-20 range. If a baseline meets that bar, we use it and drop our pipeline.
For the causal stage, the competitor is a cross-sectional regression of modification ratio on town-level platform adoption. Our difference-in-differences should give a different answer from it, and should pass the pre-trend checks that the cross-section cannot.
Why not just…?
Why not just fine-tune an off-the-shelf segmenter for signs and awnings? This is the most serious threat, and we treat it as a likely winner rather than a straw man. Facade R-CNN's comparison suggests generic segmenters produce structurally implausible masks under street-level perspective. However, that evidence comes from mean intersection-over-union on Western facades, not from area fractions on vernacular frontage. An area estimate can be accurate even when mask shapes are ugly. The calibration experiment settles this directly. If the fine-tuned segmenter's change-error matches ours, the instrument becomes that segmenter plus temporal differencing, and the contribution moves wholly to the causal study.
Why not just use satellite imagery, as the Lukang study already did? Overhead imagery sees roofs and the plan footprint. The second contour is mostly vertical: signboards, hanging canopies and projecting booths seen in elevation. Shih et al. used satellite series only to corroborate trends found by laser scanning, not to measure the ratio. We will include a satellite-derived footprint-change proxy as a further baseline in the calibration. We expect it to correlate weakly with laser-measured facade additions, and that is testable.
Why not just survey merchants or use permit records? Many second-contour elements are temporary, retractable or illegal by construction, which is exactly why the Lukang study measured them physically. Permit records undercount them by design. Surveys about illegal extensions invite misreporting, and cannot be run retroactively for the pre-platform years the comparison needs. Surveys remain useful for mechanism, asking why a merchant added a sign, and we would run a small one alongside the panel. They cannot replace the measurement.
Why not just correlate encroachment with e-commerce adoption across towns? Towns that adopt platforms early also differ in tourism volume, wealth and enforcement, and any of these could move encroachment on its own. The residential control within the same town absorbs town-wide shocks such as an enforcement campaign or a tourism boom. The staggered timing of arrival guards against fixed differences between towns. We will report the cross-section next to the difference-in-differences so readers can see how much confounding the design removes.
Decisive Checks
The project stops or pivots on two results. First, if no method, including ours, estimates change in modification ratio on Lukang with error below the pre-registered detectable effect, the imagery route is not viable. The fallback is repeated laser scanning of a few streets, which supports a much smaller study.
Second, if the instrument works, we check pre-trends: commercial and residential frontage must move in parallel before platform arrival. Then we report $\Delta$ together with the vacancy series. A positive $\Delta$ with stable vacancy supports inverted pressure. A negative $\Delta$ with stable vacancy supports relief. A negative $\Delta$ with rising vacancy indicates abandonment, the displacement outcome the colonialism critique predicts. The merchant-level contrast between listing and non-listing shops on the same street serves as a replication of whichever sign appears. Any of these results is informative for heritage policy, because each calls for a different response: enforcement plus collective signage design, platform-integration support, or protection of the retail base itself.
References
- Computing Urban Fabric of a Historical Town
- How the e-commerce environment affects the economic development of traditional enterprises: Taking business innovation as an example
- Reconceptualizing Entrepreneurial Marketing in the Age of Digital Commerce: The Case of China
- Does the Internet Deserve Everybody?
- DeepFacade: A Deep Learning Approach to Facade Parsing
- Facade Segmentation in the Wild
- Deep Facade Parsing with Occlusions
- FacadeNet: Conditional Facade Synthesis via Selective Editing
- CGAN-Assisted Renovation of the Styles and Features of Street Facades—A Case Study of the Wuyi Area in Fujian, China
- Research on the Integration of Cultural Heritage Scenes and Urban "Night Economy" Development
- Building Facade Parsing R-CNN