Dormant Until Approached: Proximity-Graded Displays for Heritage Sites

Museums and historic sites increasingly install interactive screens, but they face two failures that pull in opposite directions. Passersby tend to treat any public screen as an advertisement and walk past it, a habit called display blindness. Heritage sites, meanwhile, ask that modern equipment be hidden or disguised so it does not clash with the historic setting. The trouble is that a screen disguised well enough to satisfy a conservator is also one that nobody recognizes as interactive. The two goals are usually traded off informally, through whatever a curator and a vendor agree on, and nobody measures what each choice costs.

We propose a display that stays visually dormant, matching the surrounding material palette, while people pass at a distance. It reveals that it can be used in stages, and only as a person actually approaches. A coarse distance sensor sealed inside the display's own housing drives this, so nothing is drilled or cabled into the building. The research contribution is a measurement as much as a device. We will run a multi-week, rotating-condition field study inside an operating heritage space and plot each design on two axes: how many passersby become users, and how intrusive passive visitors and conservators judge the installation to be. The core claim is that proximity-graded revelation sits strictly above the trade-off curve set by the obvious alternatives. One of those alternatives, a revelation that simply pulses on a timer with no sensing, is a real threat, and we treat it as the decisive ablation.

Core Claim

Two things have been measured separately. Display blindness research has asked how to make interactivity visible. Heritage work has asked how to make modernity invisible. Both can be satisfied at once if visibility depends on where the visitor is standing. People at a distance get camouflage. People who have slowed and turned toward the display get faint cues. People within reach get an explicit invitation to interact. If that is right, the gain over a uniformly camouflaged or uniformly bright display should show up specifically as higher conversion at equal intrusion. The gain should also disappear when the same staged reveal runs on a timer instead of on approach. The theory-practice direction gives the methodological stance. Following the closed-loop view of deployment in the IAAI-13 Preface and the AAAI/IAAI-96 Preface, we treat the site's constraints as inputs that shape the design, not as obstacles to engineer around afterwards. We report the deployment frictions themselves, such as sensing failures in crowds and conservator vetoes, as findings.

What Prior Work Gives Us and Leaves Open

Three papers motivate the gap. Memarović et al. deployed FunSquare on outdoor screens and described movement from a passive engagement zone to an active one. Their deployment lasted two days on commercial touchscreens in a modern plaza, so novelty effects and aesthetic constraints were never separated. Cheung et al. moved first contact onto visitors' own phones through QR codes, and they describe the honeypot effect, where one active user draws in onlookers. Their prototype was a Breakout game, they report no conversion numbers, and they name museums only as a motivating venue. Liu's account of the Palace Museum documents the preservation side: air-conditioning units hidden in mesh boxes painted to match the palace walls, and reversible installations in the old Icehouse. That account is descriptive and never asks what camouflage does to discoverability. The affordance study by Ye supports the general point that users fail to act when they cannot read what an interface allows.

Two papers supply the method. Proxemic Flow used an LED floor and an overhead depth camera to show passersby nested zones for observing, intending to interact, and touching. We borrow its zone structure but not its hardware: floor grids and ceiling-mounted sensors are exactly what heritage rules forbid. Its evaluation was qualitative only. Du's study of Digital Dunhuang argues that deliberately restrained interaction preserves a sense of reverence. Its example is a "Local Light Mode" that mimics a flashlight in a dark cave. We adopt that idea for the final revealed state, so the display can invite use without switching into billboard mode. Du ran no behavioral measurement, and her setting was solo VR, not shared on-site screens.

Two papers constrain the design. MuseumVisitors shows that pedestrian detectors in a real gallery degrade sharply when visitors cluster in groups. This matters because the honeypot effect the display tries to trigger produces exactly such clusters. We therefore ask the sensor only for coarse zone occupancy, never for individual identity tracking. Mirror Target YOLO shows that heritage rules against drilling and cabling can force new sensing designs: it uses mirrors and a masking module for fire detection. We take its lesson, that the sensor should fit the building rather than the reverse, but our sensor is self-contained in the display housing, so no mirrors are needed. No paper in the file connects all three directions.

Research Program

We will build a freestanding kiosk with a matte, low-reflectance screen behind a removable façade panel. The panel's color and texture are chosen with the host site's conservators, in the spirit of the Palace Museum's matched enclosures, and the unit needs only a floor footprint and an existing power outlet. A forward-facing depth sensor inside the housing classifies the floor area in front of the kiosk into three zones adapted from Proxemic Flow. In the far zone the screen shows a slow, desaturated still image drawn from the nearby collection, so it reads as an image, not as a screen. When someone enters the middle zone and slows or orients toward the kiosk, a single restrained cue appears. This is a narrow pool of light moving across the image, borrowed from Du's flashlight mode. When someone reaches the near zone, an explicit invitation appears: touch, or a QR code for private onboarding as in Cheung et al. All revelation is reversible and fades back within seconds of the zone emptying.

The sensing question is narrow on purpose. We need to know whether anyone occupies the middle and near zones, and whether that occupancy is approach or just passing. We do not need to know who they are. Before going into the field, we validate this reduced task offline on MuseumVisitors. We project its ground-plane annotations onto virtual zones in front of its artworks, then measure zone-occupancy accuracy separately for its individual and group scenarios. The analysis then follows the closed loop: field failures, such as zone flicker in crowds, conservator objections, or network dead spots in thick masonry, feed back into the next design iteration and are reported as results.

Baselines

The do-nothing reference from the preservation side is the same kiosk left permanently camouflaged, with the full interface available on touch but no reveal at any distance. It marks the low-intrusion end of the trade-off, and we expect very low conversion from it. The naive reference from the modernization side is the same kiosk with its façade panel removed, running a looping, full-brightness attractor video with a "touch to explore" banner. Memarović et al. and Cheung et al. both argue that this broadcast style triggers display blindness and social hesitation. Still, it is the default vendor choice and may win on raw conversion, so it marks the high-intrusion end.

The strongest existing method is Cheung et al.'s cross-device approach applied to a camouflaged display. In that condition the kiosk stays dormant and a small printed QR placard on the plinth invites private phone-based onboarding. It is the strongest competitor because it already addresses social inhibition and adds almost no visual intrusion. Our condition offers the same QR path in its near state, so the comparison isolates the effect of graded revelation. No reported result was found on Lacuna for conversion rates under Cheung et al.'s method, and FunSquare reports only that about 130 individuals interacted over more than 18 hours of observation, without a passerby denominator. We will therefore measure all baselines ourselves. For the offline sensing check, the reference numbers come from MuseumVisitors: an ACF-LDCF detector reached a 23% miss rate at $10^{-1}$ false positives per image on Camera 3 under reasonable conditions, and DPM reached 75% on Camera 1. Our zone-occupancy classifier should beat these error levels on the reduced task, especially in group scenarios.

The key ablation is a timer-driven reveal. It runs the same three visual stages on a fixed cycle regardless of who is present, at a frequency matched to the proposal's average daily reveal count, so total visual intrusion is equalized. A second ablation removes the restrained flashlight cue and jumps straight from dormant to the explicit invitation. This tests whether the intermediate stage matters.

All conditions use the identical kiosk, location, content set and interaction code. They differ only in the reveal policy and placard, and they rotate on a balanced daily schedule over at least six weeks. The proposal "beats" a baseline if it achieves higher conversion at no worse intrusion rating, or lower intrusion at no worse conversion, with the difference holding in the later weeks after novelty has faded.

Why not just…?

Why not just put a QR placard next to a disguised screen? This is the most credible competitor, because it costs nothing visually and Cheung et al. motivate it well. Its weakness is that a static placard is exactly the kind of signage that visitors already filter out. It also says nothing about the screen above it. The head-to-head in our study settles the question directly. If the placard condition matches the proposal's conversion, the sensing and staged reveal are not worth their cost, and we will report that.

Why not just use a bright attractor screen and accept the aesthetic cost? It may well convert more raw passersby. The research file's evidence that it fails comes from argument and observation in Memarović et al. and Cheung et al., not from controlled numbers, and Liu documents the aesthetic objection only descriptively. Our intrusion ratings put a measured price on the bright screen. If it dominates even after intrusion is accounted for, the premise of the proposal is wrong.

Why not just reveal on a timer, with no sensing at all? This is the threat we take most seriously. A slow periodic pulse is cheaper, has no tracking failures in crowds, and might break display blindness just as well. The matched-intrusion timer ablation exists to answer it. The proposal's mechanism predicts that a timed reveal mostly fires when nobody is near and is missed by those who are. If the timer condition converts as well as the proximity condition, we will drop the sensor and recommend the timer.

Why not just move everything online or into off-site VR? Du shows that reproductions can carry reverence. Liu, however, documents visitor volumes on the scale of 16 million a year at the Palace Museum, with visitors expecting enriched on-site experiences. Off-site VR does not address the people who are physically in the gallery, which is the population this study measures.

Proposed Experiments

The first stage is the offline zone-occupancy validation on MuseumVisitors. We report the accuracy of middle-zone and near-zone occupancy separately for individuals and for groups. The go/no-go criterion is whether group scenarios stay usable, since the honeypot effect makes groups the normal case in front of a successful display.

The second stage is the field deployment. It takes place in one gallery of a partner heritage institution, with the façade palette approved by its conservators beforehand. Six conditions rotate daily: dormant, bright attractor, placard, timer reveal, direct reveal, and graded reveal. The in-housing sensor logs anonymous zone transitions. Each condition also receives trained-observer counts on sampled hours to calibrate the logs, following FunSquare's observational method but over weeks, not two days.

The primary outcome is conversion, defined as the fraction of people entering the middle zone who then interact by touch or phone. Secondary outcomes are dwell time, the size of the follow-on crowd after a first user (the honeypot effect), and time spent looking at the nearby physical artwork. The last of these tests whether the display steals attention from the relics, a concern raised in Gap 4 of the file. Intrusion is measured two ways: short intercept surveys of far-zone visitors who did not interact, asking whether the object "belongs in this room," and blinded ratings by a conservator panel of photographs and video clips from each condition.

The decisive checks are three. Does graded reveal lie above the conversion–intrusion curve traced by the dormant, placard and attractor conditions? Does it beat the matched-intrusion timer? Does its advantage survive in weeks four to six? If all three hold, the claim stands that showing interactivity only as visitors approach dissolves much of the trade-off between preservation and engagement. If the timer or the placard ties it, we will have measured, for the first time, that the simpler option is enough.