BIRD: A Museum Open Dataset Combining Behavior Patterns and Identity Types to Better Model Visitors' Experience

Lack of data is a recurring problem in Artificial Intelligence, as it is essential for training and validating models. This is particularly true in the field of cultural heritage, where the number of open datasets is relatively limited and where the data collected does not always allow for holistic modeling of visitors’ experience due to the fact that data are ad hoc (i.e. restricted to the sole characteristics required for the evaluation of a specific model). To overcome this lack, we conducted a study between February and March 2019 aimed at obtaining comprehensive and detailed information about visitors, their visit experience and their feedback. We equipped 51 participants with eye-tracking glasses, leaving them free to explore the 3 floors of the museum for an average of 57 minutes, and to discover an exhibition of more than 400 artworks. On this basis, we built an open dataset combining contextual data (demographic data, preferences, visiting habits, motivations, social context. . . ), behavioral data (spatiotemporal trajectories, gaze data) and feedback (satisfaction, fatigue, liked artworks, verbatim. . . ). Our analysis made it possible to re-enact visitor identities combining the majority of characteristics found in the literature and to reproduce the Veron and Levasseur profiles. This dataset will ultimately make it possible to improve the quality of recommended paths in museums by personalizing the number of points of interest (POIs), the time spent at these different POIs, and the amount of information to be provided to each visitor based on their level of interest.

💡 Research Summary

The paper addresses a critical gap in artificial intelligence research for cultural heritage: the scarcity of comprehensive, publicly available datasets that capture the full spectrum of museum visitor experience. To fill this void, the authors present BIRD (Behavioral and Identity‑Related Dataset), a multimodal open dataset collected from 51 participants who freely explored the three‑floor Nancy Museum of Fine Arts in early 2019. Each participant wore Tobii Glasses 2 eye‑tracking devices (100 Hz) that recorded both gaze (only for paintings) and a scene video, while a custom web platform was used to manually annotate the visitor’s position on a pixel‑based floor plan every few seconds. The raw trajectories were later normalized using MovingPandas with a 2‑second sampling interval, yielding consistent spatiotemporal streams (position, floor, timestamp) and derived metrics such as average speed, number of stops, total distance, and number of artworks observed.

In addition to behavioral data, the study gathered extensive identity information through pre‑ and post‑visit questionnaires (23 and 29 items respectively). These surveys covered demographics (age, gender, education), visit motivations, frequency, psychological factors (fatigue, crowd tolerance), and learning objectives based on Falk’s visitor identity framework. After the tour, participants used a mosaic interface to select artworks they liked, producing explicit feedback (visitor_id, item_id, selection_time) that reflects both perceived preferences and coverage gaps (artworks liked but not seen). All data are provided in anonymized JSON and CSV files, together with museum floor plans (PDF/JSON) and a catalog of 400+ artworks (metadata on period, theme, description).

Statistical analysis of the 51 trajectories shows a diverse visitor population (27 male, 23 female, 1 non‑binary; mean age 33 years, range 12‑75). The average visit lasted 57.6 minutes, covering roughly 838 meters at a mean speed of 0.26 m/s, with about 54 stops per visit and 144 artworks observed on average. Each artwork was viewed for roughly 29 seconds. These figures indicate a low‑speed, stop‑heavy navigation pattern typical of museum contexts.

To demonstrate the dataset’s suitability for identity modeling, the authors performed K‑means clustering on the global trajectory features (duration, speed, stops, length, items seen). Using elbow, silhouette, Davies‑Bouldin, and Calinski‑Harabasz indices, they identified four optimal clusters, which they labeled “Grasshopper”, “Ant”, “Fish”, and “Butterfly”. The clusters correspond to distinct visitor archetypes (high mobility vs. low mobility, varied stop patterns) and align with the Veron and Levasseur profiles cited in the literature, confirming that BIRD can support fine‑grained visitor segmentation.

The paper situates BIRD within related work, noting that existing museum datasets often focus on a single modality (e.g., Bluetooth beacons for trajectories, questionnaires for preferences) and are rarely released publicly. By integrating gaze, precise spatiotemporal movement, rich questionnaire data, and explicit artwork feedback, BIRD offers a uniquely holistic view of visitor experience. This multimodality enables a range of research applications: personalized path recommendation systems that adapt POI density, dwell time, and information depth to individual profiles; simulation environments (the authors mention a Unity‑based virtual replica of the museum) for testing crowd dynamics and virtual guide interactions; and HCI studies on how technology (e.g., AR guides) influences learning and satisfaction.

Limitations are acknowledged. The sample size (51) is modest, restricting statistical power and generalizability. Data collection was confined to a single European art museum with a specific three‑floor layout, so transferability to other museum types (science, history, open‑air sites) remains to be validated. Gaze data were limited to paintings; sculptures and other media were not captured, leaving a partial picture of visual attention. Manual annotation of trajectories, while necessary for high spatial precision, introduces potential human error despite subsequent smoothing.

Future work outlined includes scaling the participant pool, extending data collection to multiple museums and diverse exhibition formats, incorporating 3‑D gaze tracking for sculptures and installations, and enriching the dataset with physiological signals (e.g., heart rate) for deeper affective modeling. The authors also plan to continuously update the dataset (the “progressively increase the number of participants” statement) and to provide tools for researchers to query and visualize the multimodal data.

In summary, BIRD represents a significant contribution to the museum informatics community by delivering an openly accessible, richly annotated dataset that bridges the gap between behavioral tracking and visitor identity modeling. Its comprehensive structure, combined with thorough documentation and preprocessing pipelines, makes it immediately valuable for developing and benchmarking AI‑driven personalization, recommendation, and simulation systems in cultural heritage contexts.