“objects produce knowledge and meaning both in and of the world” Geismar, 2018).
Museums are perfect sites for the production and display of collections. The objects in a given collection offer not only a visual way to learn about the world but are also, through their curatorial framing, a subjective representation of identity, culture and history.
The relationship between museums and collections should be carefully explored. Items in museum collections are usually ones of high value. While most of us get a chance to observe these items, only few can own them — an elite and exclusive class that gets to define cultural value. The museum setting, and the imagined setting of a private collection, create an aura around these material objects, seeding and stoking our desire for them. This desire is commonly expressed through the act of collecting souvenirs — personal collections of mass-produced items — from places and moments that we were a part of.
Can a collection of mass-produced items be considered part of a prestigious collection, in the traditional sense? Must a standing collection be a relic of the museum or private aristocrat home? Collections of mass-produced items link between public ownership and curatorial practice. Fairs, events and exhibitions produce cheap symbolic items for people to purchase, as a token of their experience and their shared memory. Once they are sold, they belong to different people in disparate cultural environments, and the context in which they exist is subject to different curatorial logic, if any. The items are owned by anonymous individuals, who each have the agency to resell, set their own value, and present them in whatever way they see fit. These scattered “collections” are maintained by a general public that sets the rules and generates its own historic and cultural knowledge and value. There is little knowledge and cultural attention dedicated to these types of collections, yet they can tell us a lot about the mechanisms and infrastructure that define how cultural and economic value is determined.
This case study thinks about these types of collections — both those found in a museum and in anonymous homes across the world — in a historical and in a material context. We zoomed in on souvenirs from the New York World’ Fair, which took place in 1939 and 1964. Using a museum collection of these souvenirs as a point of departure, we create a web-based portal that maps the same mass-produced objects’ present-day location and curates them into one accessible, digitized display. Using a digital platform and state-of-the-art technology, we aim to highlight alternative forms of collections and speculate on the future of how they get displayed. For this experiment, we invite viewers to meditate upon this type of collection, through one specific collectible item — souvenir plates. This case study aims to explore the gap between the virtual and physical aspects of memorabilia, as well as the meaning of collecting for individuals and the authority of museums to define cultural value.
A prior conversation with assistant curator Sophia Marisa Lucas from the Queens Museum led us to look into the collection from the New York World’s Fairs of 1939 and 1964. The fairs took place where the Queens Museum stands today, and in many ways they seemed to have shaped the museum’s identity — from the Unisphere statue that has become the iconic symbol of the area through the panorama to the vitrine at the back of the museum showcasing a collection of items from the events.
As two foreign artists, we were fascinated by these events that holds the premise of bringing the entire world to one physical place. As digital artists, we wondered if the internet is the true fulfillment of the fair’s intention, to bring people together to produce cultural exchange. Does the lack of physical space of the internet affect the way we collect, share, and remember?
The World’s Fair, held every few years since the mid-nineteenth century, is an event that draws a world audience. Each fair is held in a unique location, where government agencies, corporations, civic groups, and smaller organizations from around the world arrive, build extraordinary pavilions and set up exhibitions. The 1939 World’s Fair in Queens, New York was significant, as it established a recurring theme for the event — “The World of Tomorrow.”
Over the course of two seasons, 44 million people attended the fair, catching glimpses of a possible future utopia. The exhibition transformed what was considered to be a “valley of ashes” in Queens into a dream world. Perhaps most notable, though, were the souvenirs. Mementos of the fair flooded U.S. and world markets and homes by the millions. Every visitor, no matter their economic status, brought something back (Brooke, 2018).
The World’s Fair collectibles were souvenirs commissioned by large corporations such as U.S Steel and were produced in massive quantities (Brooke, 2018). The offerings ranged from plastic dinosaur dolls to car models and kitchenware to pins and much more. The items displayed in the Queens Museum today are of this mass-produced heritage. Unlike many museum collections, these artifacts do not have the aura of a singular, precious artifact. Following the simple reason that the Queens Museum has collected and displayed one copy, while the rest of the collection is scattered around the globe, the museum copy of these artifacts cannot be enough to tell the full story. In order to get a sense of the event’s global impact, perhaps we need to expand the classic museum display and find new ways to curate and connect the items.
We decided to follow one specific type of item from the entire collection — collectible plates — as an example that reflects broadly on the way in which collections evolve beyond the museum’s borders. The Queens Museum displays around 30 different types of plates that were produced to commemorate the New York World’s Fairs of 1939 and 1964. Our research revealed that the collection was partially donated and some parts were acquired from private owners throughout the years. Even though the museum does not claim its collection to be a complete one, we decided to treat it as such. Choosing to focus on one type of item would allow us to research and experiment more with the techniques and technologies that are fit to our purpose.
With the goal of revealing new methods for categorizing and archiving the plates, we decided to classify the artifacts in this collection according to visual properties such as shape and color.
Before starting, we set the following rules: this experiment will be displayed on an online and accessible platform showcasing a dynamic image of the collectibles and the people that engage with them, open for others to add their own. In short, we set out to create a democratic collection.
Establishing these rules quickly led us to the choice of the technology, and therefore, the approach for implementation.
1. Machine learning
We started to explore the use of machine learning algorithms in order to be able to identify the relevant items when sifting through images found on the internet. Machine learning is one of the most influential technologies of our times; it affects every aspect of human life. Our lives are being redefined by algorithms, robots and sensors, systems that change the way we engage with reality. This technology allows us to compute a vast amount of data and to identify and analyze hidden patterns.
Machine learning has been put to use in the museum field over the past years, as massive archives have been digitized. Most use cases for museums are inward-facing — analyzing and classifying the collection owned by the institution (Ciecko, 2017). We offer to apply this technology in order to expand the museum collection to include items outside of the building, looking outwards and analyzing the collection through global communities.
In this case study we mainly explored the use of image recognition technology to match and categorize similar plates scattered online. We started by manually searching online for the specific types of plates the museum holds, using their collection as an index. We focused on eBay and Etsy and created a small data set for training as a start. This initial batch included a little over 300 images. For each type of plate, we gathered around 15–20 different images to implement into the TensorFlow training model. At this stage, we wanted to confirm our initial assumptions about the use of image recognition technology as an approach for our goal of plate classification.
2. Scraping the internet
Our second step was to develop an online scraping tool for selected websites which we found to be popular platforms for these items — eBay and Etsy. This tool locates relevant images based on keyword variation, such as “world’s fair plate”, “WF NYC plates” and so on to compare with the type of plate in the original index we manually created.
The scraper runs daily and saves the following content: date, location, description, price and images. We started getting some interesting results from our keyword search, but the system seemed to fail in matching the found plates to the index we created.
3. Synthetic data and 3D scanning
At this stage, we understood that we needed to increase the number of images in our dataset in order to train the machine learning system on more sources. Better training would provide more reliable matches between the online images (found via the scrapper tool) and the correct plate type.
We started to research the process of creating a synthetic dataset. A synthetic dataset is a dataset that uses one source, either a 2D image or 3D model, to generate a large number of images of the same object in a variety of conditions (lights, environments, angles and etc). This process eases the learning curve for the algorithm and improves its performance. The greatest advantage of creating synthetic data lies in the fact that it allows us to use a small number of sources, such as the one that usually exists in museums, to generate hundreds of outputs.
Our approach for the creation of the synthetic dataset was to 3D scan the Queens Museum plate collection (26 plates to be exact) and to simulate different conditions within a 3D software (Unity3D, in our case). 3D scanning is a method used by a growing number of institutions to archive their collections, as it offers the most comprehensive information regarding the object’s measurements and detail. We are particularly inclined to use 3D scanning since it allows us to not only archive but enable further engagement with the artifacts.
That said, 3D scanning is not the most effective method of documentation for all types of objects. For example, overcoming the reflection of the plates was a huge challenge, as the reflections confuse the 3D scanning software and the scans often fail. We opted for scanning over 3D modeling despite several downsides because we saw a value in the leanings gained from this process, learning which could be applied in the future on other collections. We were also excited by the fact that this will render a 3D collection, open for the museum and the public to use.
We used Photogrammetry, a 3D scanning method that uses a set of images from different angles to compute a 3D reconstruction. With the help of the museum staff, we took the plates out of the vitrine into a softbox where we set up a turntable with a green screen background, to provide fast and easy cleanup in the post-production process. Some of the plates needed adjustments to their image textures, remodeling, holes fixed and other issues caused by scanning reflective materials.
Once we had a 3D representation of all 26 plates, we inserted them into Unity3D and rendered each one of them in three different lighting settings (white, yellow and blue) and four different backdrops (black, white, wooden planks and a floral tablecloth). These choices were inspired by the images people tend to upload to Etsy and Ebay to represent the items they are selling.
4. Web integration
Stage four included retraining the machine learning system with the final training set, which included both the original images and the synthetic ones and provided much more accurate results. We started to explore the online results that came through.
At the same time we developed a website: www.platesofthe.world
The website allows users to explore the museum’s collection along with the online items that were found around the internet. The interface invites two modes of explorations: a grid view and a map view. The grid view presents the museum’s collection, with the plates that matched it online via the scraper system. The map mode allows exploration of the plates based on their current location in the world. Right now the majority of plates are located in the U.S, but this could change over time or if we were to search for different objects, websites and APIs.
The user can click on each plate in order to get more information about it. Currently, we are developing an addition page that will allow users to upload their own plates. This will feed back into the ML model with the goal of making it more accurate over time.
This experiment aims to challenge the traditional roles of the museum and the hobbyist collector by proposing an expansion of the museum’s visible storage to include items scraped from websites. We were excited about the idea of tracing the objects back to where they were initially sold (back to the Queens Museum collection), and about how this action presents a more comprehensive view of a collection. In retrospect, Queens Museum didn’t have a lot of back-end information to provide us with. For example, we hoped to know the number of items that were produced from each plate design in order to understand how far we should push our machine learning system. We gave up on this idea after learning from museum staff that they didn’t have this information. We believe that this case study can expand and locate more information, based on the specific needs of each collection and the museum.
Our scraping also yielded instances of plates that the Queens Museum does not have in their display. This brought up another potential use of the system we built which is to locate items that the museum does not own yet but might want to add to their collection.
This experiment expands the understandings of old and new collections, draws attention to the interdependence of collecting and sheds light on how collections form knowledge of cultural heritage on a larger scale. As we shifted the collection across platforms, throughout the development, there were moments when one kind of technological observation, such as scraping the internet, gives way to another, such as smart image recognition. There were also moments in which we learn about what we consider to be a “real” part of the collection and what we perceive to be a public form of culture, and the benefit that lies in merging the two.
Without getting into too many technical details we would like to share two insights we gained from combining 3D scanning with machine learning algorithms in this project:
We chose 3D scanning the plates over the option of 3D modeling them. This decision was influenced by our prior experience in the scanning field. Looking back, however, it may have been better to opt for modeling, considering the time it took to process the scan results and the relatively simple shape of the plates. Furthermore, as the machine learning algorithms observe all the details in the training data, the imperfect results of the scan can even damage the learning process by providing false information about the object. However, we are overall pleased with the quality of the 3D data set we generated of the museum collection and are excited about potential future applications for it.
One of the questions we entered with was — can the museum track the presence, impact and the community around these artifacts worldwide and in the digital sphere? And how do such actions challenge the audience-collector relationship? It will take longer to reflect and answer these questions but we are happy to have created this experiment that can live in the world and engage the audience in this way.
Our next step in this experiment is underway will be presented as part of the Queens International Biennial. For this phase we used a different approach to image recognition, image classification and text to image, to explore how can we program a machine that “observes” the plate collection and generates a new plates on its own. Through this extended experimentation and research, we hope to deepen our understanding of the process by which humans collect memories, a process that highlights how knowledge and identity are constructed.
More about this part of the work on the Qi2018 website.
Opening October 7, 2018.