1 Introduction
The past decade has seen an increase in the use of photogrammetry for digitally recording and sharing cultural heritage. The trend is widespread not only in institutions and groups that are specialised in working with cultural heritage [
2,
57,
70,
71]; groups and individuals with particular interests in heritage are also using photogrammetric models in a myriad of academic research and creative projects, re-purposing such models for use in designs and entertainment within the creative sector [
1,
12,
20,
21]. While such endeavours are worthwhile, it may be unsustainable in terms of resources to digitally capture, process, and store artefacts at all levels of priorities due to the inestimable amount of cultural heritage objects scattered over large geographical regions. The British Museum, for example, has been collaborating with Sketchfab – one of the largest online 3D model platforms since 2014. At the time of writing, the British Museum has digitised and published 273 high-quality 3D models from the museum’s collection, which has gained 1.6M views and 13.6K likes [
63]. Despite the quality, the publication of the models has been relatively slow, implying that closed digitisation work may require greater financial budgeting, technical support and human resources that are part of the museum’s digital team. That is why mass photogrammetry [
13] that uses crowdsourcing means to achieve its goals may become necessary.
Our main goal is to obtain 3D reconstructions of cultural heritage objects that can faithfully capture the form and colour of the actual artefacts. The 3D models are neither digital modelling that requires designers’ interpretations nor technology-savvy documentation that involves additional geometric information or interior measurement [
44,
45]. Mass photogrammetry rarely achieves professional, archivable copies due to the high variability of imaging devices, lens settings and lighting conditions. We aim to capture digital surrogates with adequate visual resemblances to the original physical objects for most consumption scenarios, including communication, sharing, cultural exchanges, creative products, and even research [
13]. Since offline close-range photogrammetry procedures can be separated into on-site image acquisition and off-site image processing [
40], the time-intensive image acquisition may benefit from engaging the public. Our research, therefore, proposes crowd-based image acquisition as it can use ubiquitous smartphones and accessible DSLR cameras, together with the crowds’ voluntary attitudes, to reconstruct photogrammetry-based 3D artefacts at a massive scale and possibly beyond the reach of institutions.
High-quality image acquisition is not easy as far as mass photogrammetry is concerned. Following the offline close-range photogrammetry with multi-image acquisition approaches, the object should be, in principle, acquired by multiple overlapping images from locations chosen to enable sufficient intersecting angles of bundles of rays in object space [
40]. The number of photographs is unspecified, all based on the target’s shape, form, and details. When extended to mass photogrammetry, the simple solution would be to collect as many high-resolution images as possible and hope that the images have at least 60% overlaps and cover every possible angle at different distances [
13]. To reconstruct high-quality 3D models, we focus primarily on capturing cultural heritage objects that possess the following characteristics:
•
Display sufficient space for manoeuvring one’s camera all around the display
•
Material avoid materials with low contrast to the background, e.g., transparent, translucent, glow, or reflective
•
Illumination suitable lighting conditions to avoid too much exposure and shadows
Therefore, the challenge is to outsource image acquisition tasks to a large group of amateurs so that their contributions can be used to reconstruct high-quality photogrammetric 3D models. Prevalent crowdsourcing trends we observed focus on stimulating as many independent contributions as possible to ensure the diversity of the crowdsourced data [
73]. We think collecting isolated contributions may not be the best way to crowdsource 3D models because this may yield inadequate and fragmented data (please see our definition of what constitutes high-quality data in
Section 2.3 and why conventional crowdsourcing is not suitable in
Section 2.4). To create crowdsourcing activities that can be mapped to collective tasks to ensure the quality of photogrammetric 3D models, more effective mechanisms should be explored.
It is a known fact that collaborative teamwork can yield synergetic effects [
27] so as to solve more complex problems [
6,
54]. Some researchers have proposed the implementation of team structures in typical crowdsourcing mechanisms [
4,
43,
59,
60,
67] to reduce individual workload, shorten the time, boost individual performance, and so on. However, the majority of existing collaborative crowdsourcing approaches fall short of the ability to facilitate productive collaboration due to the inflexible and inactive team mechanisms [
66]. Therefore, it is vital to understand the crowd behaviours and collaboration dynamics for it can help reveal the characteristic behaviours embedded in the crowdsourcing activity.
The unique nature of crowdsourcing 3D cultural heritage objects introduces a variability of elements in terms of task complexity, professionalisation, incentives of the crowd, and so on. With the introduction of team structures, there is also the added layer of disorganised and self-organised groups that may come together asynchronously and synchronously to complete a set of tasks. These unknowns have never been studied, and we believe that understanding these phenomena may provide us with the knowledge for facilitating a more optimised crowdsourcing workflow to obtain large numbers of high-quality 3D models. To fill this research gap, we aim to investigate the effectiveness and differences of crowdsourcing behaviours that are carried out following two different crowdsourcing mechanisms:
•
Group A a group of disorganised individuals working independently and asynchronously
•
Group B a crowd-based self-organised team working together synchronously
Group A follows the bottom-up procedure as most conventional crowdsourcing approaches, which implicitly harnesses the collective efforts since individual participants are unaware that their contributions will be later aggregated. Group A allows for isolated, asynchronous contributions. On the other hand, Group B can be seen as a top-down approach because participants will be assigned to a team and are fully aware they will be synchronously working together towards a common goal. We investigate the viability of incorporating team structures in terms of 3D model completeness by testing three hypotheses:
•
H1 The disorganised group of individuals working asynchronously will produce an incomplete 3D model.
•
H2 The crowd-based self-organised team working synchronously will produce a complete 3D model.
•
H3 The combination of imageries generated by the disorganised group and the crowd-based self-organised team will produce a complete 3D model that is better than each group separately.
The paper is organised as follows. We first reviewed existing practices on crowdsourcing 3D models. We then proposed a plausible solution to mitigate the challenge of obtaining high-quality data via implementing team structures in the collaborative crowdsourcing mechanism, assisted with proper task assignment and coordination. A Web App was developed to facilitate our crowdsourcing activity in terms of image acquisition and uploads. In the next section, our exploratory study investigated two scenarios constructed for two different crowdsourcing mechanisms. We used multiple metrics to measure the crowdsourcing performances at both group level and individual level. These include Productivity (quality of the 3D reconstruction); Dedication (time and efforts spent); Experience (image acquisition and application interaction); Psychological Ownership (to what extent does participants think they own the images and reconstructed 3D models); and Future Intention (willingness to take more pictures and recommend this activity to friends). This is followed by data analysis and findings. The paper ends with a discussion on the restrictions and limitations of this work and how it can be generalised to establish a more optimised workflow for future crowdsourcing activities.
6 Discussion
The present research successfully demonstrated that self-organised team structures can leverage collective behaviours for collaborative crowdsourcing and facilitate mass photogrammetry of 3D objects. Through our experimental study, we have validated the hypotheses raised at the beginning of the article.
The disorganised crowdsourcing Group A conducted is a pervasive approach in most crowdsourcing scenarios. It implicitly aggregates isolated contributions by encouraging individuals to explore the tasks freely and contribute asynchronously. The results (Figure
4) validated that H1 - The disorganised group of individuals working asynchronously will produce an incomplete 3D model. Group B follows the self-organised synchronous collaborative crowdsourcing mechanism that integrates crowd-based team structures for effective collaboration. The results (Figure
5) verified H2 - The crowd-based self-organised team working synchronously will produce a complete 3D model. The results (Figure
7) obtained by combining data from both groups validated H3 - The combination of imageries generated by the disorganised group and the crowd-based self-organised team will produce a complete 3D model that is better than each group separately.
Our experiments revealed that a crowd-based, self-organised team working in synchrony could effectively facilitate crowdsourcing 3D models regarding model completeness. It also alleviated, to some degree, participant fatigue in the repeated photo-taking practices. Conventional crowdsourcing, which engages a group of disorganised participants working independently, was sub-optimal because the aggregation of isolated contributions can only produce a partially reconstructed 3D model. The model generated using the combination of contributions from both groups obtained the best details and completeness. But most of the credits should be attributed to the contributions from Group B - the crowd-based self-organised team. The crowd-based self-organised team performs better and more consistently than the traditional crowdsourcing approach throughout our fieldwork.
We investigated further the reasons why crowd-based, self-organised team have positively affected crowdsourced models. The nature of our crowdsourcing was voluntary and location-dependent. It provided no tangible rewards and required spontaneous participants conduct on-site crowdsourced tasks. As compared to the disorganised and independent crowdsourcing context (Group A), forming an ad-hoc team required relatively high coordination costs. Although scheduling could be flexible, the lighting and weather conditions were limitations that could greatly affect the quality of the final reconstruction. Additionally, non-experts were recruited to compose a group of amateurs who would be similar to future volunteers from any community. No administrator or team leader was assigned, and no formal rules or established norms were set; thus, the team was self-organised with fluid team boundaries, which has less pressure and more freedom and flexibility.
After the team was formed, synchronous collaboration was scheduled at the designated location. The emergent collaboration created a social context that was different from traditionally individual-based crowdsourcing scenarios (Group A). Participants’ physical presence allowed them to observe peer activities and adapt their own via vicarious learning. Through implicit communication, exchanged information reinforced a shared mental model pertaining to their responsibility and collective goal. Evidences can be found when members of Group B constantly gave credits to their teammates. They had higher familiarity with the photogrammetric process, and became more concerned about the 3D reconstruction. As more time and effort were invested, participants experience psychological ownership towards the contents they have contributed, further stimulating their prosocial motivations. Participants’ prosocial behaviours led to higher dedication and more contributions. Such a positive cycle is beneficial for mass photogrammetry.
Although crowd-based team structures can promote individual behaviour and performance, the inherent attitudes would be difficult to change. Participants with little or no intrinsic motivations are unlikely to participate or contribute in future activities. This indicates that, in the long run, there will be challenges in maintaining such activities in terms of coordination and motivation. This also implies that the right kinds of volunteering will need to be targeted as this can be a critical step towards success in collective performance. We also learn that participants can be motivated by simple acknowledgements, such as giving credit where credit is due. These can be a viable approach for maintaining participant interest.
Despite the positive findings, there are limitations and issues that can be further investigated and improved. Future research can explore influential factors such as task complexity, team composition and personality traits, technology-mediation, remote work, and so on. Our participants were limited to the university demographics, who may have similar experiences and interests according to their age groups. We felt that further work could be conducted within a larger context with more diversified demographics so we can generalise our findings to other scenarios, especially when demographic factors may affect team collaboration. In the present research, the crowd-based team were randomly assigned by the initiator. Results may improve if the process of team formation is from the bottom-up, i.e., self-selection where participants can decide whether they would want to work alone or collaboratively.
For our data collection and analysis, we collected the image contributions via the Web App, tracked users’ behavioural data from the metadata files stored in the images; and learnt self-reported feedback using questionnaires and semi-structured interviews. Although we re-generated the footprints of participants after the campaigns, it would have been better if there were observers during the fieldwork so that behavioural data can be documented. This may help researchers gain a deeper understanding of correlations between specific behaviours and cognition that might have shaped these behaviours. We feel that future research can include timestamps during collaboration. While we have recorded the total time participants have spent in the tasks, a more accurate record of time may help uncover collaborative dynamics from both the individual and the team level. This awareness can assist initiators in identifying potential inhibiting factors in organisational structures within specific contexts. Furthermore, more appropriate and accurate prescriptive measurements (such as bipartite network analysis) can be used, with particular criteria revealing structural features embedded within the collaboration network and temporal dynamics.
In this experiment, we demonstrated through data that the disorganised group of individuals working asynchronously could improve the overall model completeness by voluntarily adding more pictures with detailed features. Since the members of this group expressed a higher interest in taking more pictures and a considerable higher willingness to recommend this activity to friends, we can explore the use of this factor in organising asynchronous collaborative crowdsourcing in the future. For instance, a future crowdsourcing campaign could visualise reconstruction process in real-time, transforming an initially incomplete model and refining it as more participants contributed photos. We could also explore approaches to facilitate online asynchronous collaboration via technology-mediated communications strategies. Since participation is often spontaneous, it is of importance to investigate sustained participation in ongoing crowdsourcing campaigns.