On Exhibitions and Iterations

This post was originally published on the Cooper Hewitt Labs blog.

Since reopening in December 2014, we’ve found that the coming opening of an exhibition is a big driver of iteration. The work involved in preparing an exhibition involves the whole museum and is one of the most coordinated and planned-out things we do, and because of this, new exhibitions push us to improve in a number of ways.

First, new exhibitions can highlight existing gaps or inefficiencies in our systems. Our tagging tool, for example, always sees a round of bug fixes or new features before an exhibition because it coincides with a time when it will see heavy use. Second, exhibitions present us with new technical challenges. Objects in the Heatherwick exhibition, for example, were displayed in the galleries grouped into “projects,” which is also how we wanted users to collect them with their Pens and view them on the website. To accomplish this we had to figure out a way that TMS, our collections management software, could store both the individual objects (for internal purposes) and the grouped projects (which would hold all the public-facing images and text), and figure out how to see that through to the website in a way that made registrars, curators and ourselves comfortable. Finally, a new exhibition can present an opportunity for experimentation. David Adjaye Selects gave us the opportunity to scale up Object Phone, a telephone-based riff on the audio guide, which originally started as a small, rough prototype.

Last week was the opening of our triennial exhibition “Beauty,” which similarly presented us with a number of technical challenges and opportunities to experiment. In this post I’ll share some of those challenges and the work we did to approach them.

Collecting Exhibition Text

Triennial's wall text, with the collect icon in the lower-right corner Triennial’s wall text, with the collect icon in the lower-right corner

Since the beginning of the pen project we’ve been saying that the Pens don’t just have to collect objects. Aaron and Seb wrote in their paper on the project that “nothing would prevent the museum from allowing visitors to ‘collect’ individual designers, entire exhibitions or even architectural elements from the building itself in the future.” To that end, we’ve experimented with collecting shop items and decided that with the triennial we would allow visitors to collect exhibition text as well.

Exhibition text (in museum argot, “A-Panel” is the main text at the beginning of an exhibition and “B-Panel” are any additional texts you might find along the way) makes total sense as something that a visitor should be able to remember for later. It explains and contextualizes an exhibition’s goals, contents and organization. We’ve had the text on our collections since we reopened but it took a few clicks to get through from a visitor’s post-visit website. Now, the text will be right there alongside all of a visitor’s objects.

The exhibition text on a post-visit website The exhibition text on a post-visit website

The open-ended part of this is what visitors will expect when they collect an “exhibition.” We installed the collection points with no helper text, i.e. it doesn’t say “press here to collect this exhibition’s text.” We think it’s clear that the crosshairs refer to the text, but one of our original ideas was that we could have a way for the visitor to automatically collect every object in the exhibition and I wonder if that might be the implied function of the text tag. We will have to observe and adapt accordingly on that point.

Videos Instead of Images

When we first added videos to our collections site, we found that the fastest way to accomplish what we needed was to use TMS for relating videos to objects but use custom software for the formatting and uploading of the videos. We generate four versions of every video file — subtitled and not subtitled at two resolutions each — which we use in the galleries, on the tables and on the website. One of the weaknesses of this pipeline is that because the videos don’t live in the usual asset repository the way all of our images do, the link between TMS and the actual file’s location is made by nothing more than a “magic string” and a bit of guesswork. This makes it difficult to work with the video records in TMS: users get no preview and it can be difficult to know which video ID refers to which specific video. All of this is something we’ll be taking another look at in the near future, but there is one small chunk of this problem we approached in advance of the Triennial: how to make our website show the video in place of the primary image if it would be more appropriate to do so.

Here’s an example. Daniel Brown’s On Growth and Form is an animation on display in the Triennial. Before, it would have looked like this — the primary image is a still rendering that has been added in TMS, and the video appears as related content further down the page.


What we did is to say if the object is itself a video, animation or other screen-based media and we have an associated video record linked to the object, remove the primary image and put the video there instead. That looks like this:

Screen Shot 2016-02-16 at 3.33.50 PM

Like all good iterations, this one opened up a bunch of next steps. First, we need to figure out how to add videos into our main digital asset pipeline so that the guesswork can be removed from picking primary videos and a curator or image specialist can select it as “primary” the same way they would do with an image. Next, it brought up an item that’s been on the backburner for a while, which is a better way to display alternate images of an object. Currently, they have their own page, which gets the job done, but it would be nice to present some alternate views on the main object page as well.

Just a Reflektor Sandbox

It's fun! It’s fun!

We had a great opportunity to do some experimentation on our collections site due to the inclusion of Aaron Koblin and Vincent Morisset’s interactive video for Arcade Fire’s Just a Reflektor. The project’s source code is already available online and contains a “sandbox” environment, a tool that demonstrates some of the interactive visual effects created for the music video in a fun, open-ended environment. We were able to quickly adapt the sandbox’s source code to fit on our collections site so that visitors who collect the video with their Pen will be able to explore a more barebones version of the final interactive piece. You can check that out here.

Fully Loaded Labels

When we were working on the Pen prototypes, we tried six different NFC tags before getting to the one that met all of our requirements. We ended up with these NTAG203 tags whose combination of size and antenna design made them work well with our Pens and our wall labels. Their onboard memory of 144 bytes, combined with the system we devised for encoding collection data on them, meant that we could store a maximum of 11 objects on a tag. Of course we didn’t see that ever being a problem… until it was. The labels in the triennial exhibition are grouped by designer, not by object, and in some cases we have 35 objects from a designer on display that all need to be collected with one Pen press. There were two solutions: find tags with more memory (aka “throw more hardware at it”) or figure out a new way to encode the tags using fewer bytes and update the codebase to support both the new and old ways (aka “maintenance nightmare”). Fortunately for us, the NTAG216 series of tags have become more commonly available in the past year, which feature 888 bytes of memory, enough for around 70 objects on a tag. After a few rounds of end-to-end testing (writing the tag, collecting it with a pen and having it show up on the post-visit website), we rolled the new tags out to the galleries for the dozen or so “high capacity” labels.

The new tag (smaller, on the left) and the old tag (right) The new tag (smaller, on the left) and the old tag (right)

The most interesting iteration that’s been made overall, I think, is how our exhibition workflow has changed over time to accommodate the Pen. With each new exhibition, we take what sneaked up on us the last time and try to anticipate it. As the most recent exhibition, Beauty’s timeline included more digitally-focused milestones from the outset than any other exhibition yet. Not only did this allow us to anticipate the tag capacity issue many months in advance, but it also gave us more time to double check and fix small problems in the days before opening and gave us more time to try new, experimental approaches to the collections website and post-visit experience. We’re all excited to keep this momentum going as work ramps up on the next exhibitions!


Rethinking Search on the Collections Site

This post was originally published on the Cooper Hewitt Labs blog.

One of my longer-term projects since joining the museum has been rethinking how the search feature functions on the collections website. As we get closer to re-opening the museum with a suite of new technologies, our work in collaboration with Local Projects has prompted us to take a close look at the moving pieces that comprise the backend of our collections site and API. Search, naturally, forms a large piece of that. Last week, after a few weeks of research and experimentation, I pushed the first iteration live. In this post, I’ll share some of the thoughts and challenges that informed our changes.

First, a glossary of terms for readers who (like me, a month ago) have little-to-no experience with the inner-workings of a search engine:

  • Platform: The software that actually does the searching. The general process is that we feed data to the platform (see “index”), and then we ask it for results matching a certain set of parameters (see “query”). Everything else is handled by the platform itself. Part of what I’ll get into below involves our migration from one platform, Apache Solr, to another, Elasticsearch.
  • Index: An index is the database that the search platform uses to perform searches on. The search index is a lot like the primary database (it probably could fill that role if it had to) but it adds extra functionality to facilitate quick and accurate retrieval of search results.
  • Query: The rules to follow in selecting things that are appropriate to provide as search results. For users, the query could be something like “red concert poster,” but we have to translate that into something that the search provider will understand before results can be retrieved. Search providers give us a lot of different ways we can query things (ranges of a number, geographic distance or word matching to name a few), and a challenge for us as interface designers is to decide how transparent we want to make that translation. Queries also allow us to define how results should be sorted and how to facet results.
  • Faceting/Aggregation: A way of grouping results based on traits they posses. For example, faceting on “location” when you search our collection for “cat” reveals that 80 of our cat-related things are from the USA, 16 are from France, and so on.
  • Analysis (Tokenization/Stemming etc): A process that helps a computer work with sentences. Tokenization, for example, would split a search for “white porcelain vase” into the individual tokens: “white,” “porcelain” and “vase,” and then perform a search for any number of those tokens. Another example is stemming, which would allow the platform to understand that if a user searches for “running,” then items containing other words like “run” or “runner” are also valid search results. Analysis also gives us the opportunity to define custom rules that might include “marathon” and “track” as valid results in a search for “running.”

The State of Search

Our old search functionality showed its symptoms of under-performance in a few ways. For example, basic searches — phrases like “red concert poster” — turned up no results despite the presence of such objects in our collection, and searching for people would not return the person themselves, only their objects. These symptoms led me to identify what I considered the two big flaws in our search implementation.

On the backend, we were only indexing objects. This meant that if you searched for “Ray Eames,” you would see all of the objects we have associated with her, but to get to her individual person page, you would have to first click on an object and then click on her name. Considering that we have a lot of non-objects1, it makes sense to index them all and include them, where relevant, in the results. This made my first objective to find a way to facilitate the indexing and querying of different types of things.

On the frontend, we previously gave users two different ways to search our collection. The default method, accessible through the header of every page, performed a full text search on our Solr index and returned results sorted by image complexity. Users could also choose the “fancy search” option, which allows for searches on one or more of the individual fields we index, like “medium,” “title,” or “decade.” We all agreed here that “fancy search” was confusing, and all of its extra functionality — faceting, searching across many fields — shouldn’t be seen as “advanced” features. My second objective in rethinking how search works, then, was to unify “fancy” and “regular” search into just “search.”

Objective 1: Update the Backend

Our search provider, Solr, requires that a schema be present for every type of thing being indexed. The schema (an XML file) tells Solr what kind of value to expect for a certain field and what sort of analysis to perform on the field. This means I’d have to write a schema file — anticipating how I’d like to form all the indexed data — for each new type of thing we want to search on.

One of the features of Elasticsearch is that it is “schemaless,” meaning I can throw whatever kind of data I want at the index and it figures out how to treat it. This doesn’t mean Elasticsearch is always correct in its guesses — for example, it started treating our accession numbers as dates, which made them impossible to search on — so it also gives you the ability to define mappings, which has the same effect as Solr’s schema. But if I want to add “people” to the index, or add a new “location” field to an object, using Elasticsearch means I don’t have to fiddle with any schemas. This trait of Elasticsearch alone made worth the switch (see Larry Wall’s first great virtue of programmers, laziness: “the quality that makes you go to great effort to reduce overall energy expenditure”) because it’s important to us that we have the ability to make quick changes to any part of our website.

Before building anything in to our web framework, I spent a few days getting familiar with Elasticsearch on my own computer. I wrote a python script that loops through all of the CSVs from our public collections repository and indexed them in a local Elasticsearch server. From there, I started writing queries just to see what was possible. I was quickly able to come up with a lot of the functionality we already have on our site (full-text search, date range search) and get started with some complex queries as well (“most common medium in objects between 1990-2000,” for example, which is “paper”). This code is up on Github, so you can get started with your own Cooper Hewitt search engine at home!

Once I felt that I had a handle on how to index and query Elasticsearch, I got started building it into our site. I created a modified version of our Solr indexing script (in PHP) that copied objects, people, roles and media from MySQL and added them to Elasticsearch. Then I got started on the endpoint, which would take search parameters from a user and generate the appropriate query. The code for this would change a great deal as I worked on the frontend and occasionally refactored and abstracted pieces of functionality, but all the pieces of the pipeline were complete and I could begin rethinking the frontend.

Objective 2: Update the Frontend

Updating the frontend involved a few changes. Since we were now indexing multiple categories of things, there was still a case for keeping a per-category search view that gave users access to each field we have indexed. To accommodate these views, I added a tab bar across the top of the search forms, which defaults to the full-collection search. This also eliminates confusion as to what “fancy search” did as the search categories are now clearly labeled.


The next challenge was how to display sorting. Previously, the drop-down menu containing sort options was hidden in a “filter these results” collapsible menu. I wanted to lay out all of the sorting options for the user to see at a glance and easily switch between sorting modes. Instead of placing them across the top in a container that would push the search results further down the page, I moved them to a sidebar which would also house search result facets (more on that soon). While it does cut in to our ability to display the pictures as big as we’d like, it’s the only way we can avoid hiding information from the user. Placing these options in a collapsible menu creates two problems: if the menu is collapsed by default, we’re basically ensuring that nobody will ever use them. If the menu is expanded by default, then it means that the actual results are no longer the most important thing on the page (which, on a search results page, they clearly are). The sidebar gives us room to lay out a lot of options in an unobtrusive but easily-accessible way2.


The final challenge on the frontend was how to handle faceting. Faceting is a great way for users who know what they’re looking for to narrow down options, and a great way for users who don’t know what they’re looking for to be exposed to the various buckets we’re able to place objects in to.

Previously on our frontend, faceting was only available on fancy search. We displayed a few of the faceted fields across the top of the results page, and if you wanted further control, users could select individual fields to facet on using a drop-down menu at the bottom of the fancy search form. When they used this, though, the results page displayed only the facets, not the objects. In my updates, I’ve turned faceting on for all searches. They appear alongside the search results in the sidebar.


Doing it Live

We initially rolled these changes out about 10 days ago, though they were hidden from users who didn’t know the URL. This was to prove to ourselves that we could run Elasticsearch and Solr alongside each other without the whole site blowing up. We’re still using Solr for a bit more than just the search (for example, to show which people have worked with a given person), so until we migrate completely to Elasticsearch, we need to have both running in parallel.

A few days later, I flipped the switch to make Elasticsearch the default search provider and passed the link around internally to get some feedback from the rest of the museum. The feedback I got was important not just for working out the initial bugs and kinks, but also (and especially for myself as a relative newbie to the museum world) to help me get the language right and consider all the different expectations users might have when searching our collection. This resulted in some tweaks to the layout and copy, and some added functionality, but mostly it will inform my bigger-picture design decisions going forward.

A Few Numbers…

Improving performance wasn’t a primary objective in our changes to search, but we got some speed boosts nonetheless.

Query Before (Solr) After (Elasticsearch)
query=cat, facets on 162 results in 1240-1350ms 167 results in 450-500ms
year_acquired=gt1990, facets on 13,850 results in 1430-1560ms 14,369 results in 870-880ms
department_id=35347493&period_id=35417101, facets on 1,094 results in 1530-1580ms 1,150 results in 960-990ms

There are also cases where queries that turned up nothing before now produce relevant results, like “red concert poster,” (0 -> 11 results) “German drawings” (0 -> 101 results) and “checkered Girard samples” (0 -> 10 results).

Next Steps

Getting the improved search in front of users is the top priority now – that means you! We’re very interested in hearing about any issues, suggestions or general feedback that you might have — leave them in the comments or tweet us @cooperhewittlab.

I’m also excited about integrating some more exiting search features — things like type-ahead search and related search suggestion — on to the site in the future. Additionally, figuring out how to let users make super-specific queries (like the aforementioned “most common medium in objects between 1990-2000”) is a challenge that will require a lot of experimentation and testing, but it’s definitely an ability we want to put in the hands of our users in the future.

New Search is live on our site right now – go check it out!

1 We’ve been struggling to find a word to use for things that are “first-class” in our collection (objects, people, countries, media etc.) that makes sense to both museum-folk and the laypeople. We can’t use “objects” because those already refer to a thing that might go on display in the museum. We’ve also tried “items,” “types” and “isas” (as in, “what is this? it is a person”). But nothing seems to fit the bill.

2 We’re not in complete agreement here at the labs over the use of a sidebar to solve this design problem, but we’re going to leave it on for a while and see how it fares with time. Feedback is requested!

Announcing SkyDesigner! Sam Brenner joins the Labs

This post was originally published on the Cooper Hewitt Labs blog.

As part of my application for my position at the Cooper Hewitt museum, I built SkyDesigner, a web application that lets users replace the color of the sky with a picture of a similarly-colored object from the Cooper Hewitt’s collection. The “sky” idea comes from the original assignment, which was to create an application using both a weather API and the Cooper Hewitt API, but you can use SkyDesigner to swap out colors from anything you can take a picture of (meaning, it’s great for selfies). Give it a try now!

687474703a2f2f7777772e73616d6a6272656e6e65722e636f6d2f70726f6a656374732f736b792f6c69622f696d672f30312e6a7067 687474703a2f2f7777772e73616d6a6272656e6e65722e636f6d2f70726f6a656374732f736b792f6c69622f696d672f30322e6a7067

Here’s how it works: first, users take a picture. If they’re on a computer, they can use their webcam. If they’re on a smartphone, they can use the built-in camera. Android users get (in my opinion) the better experience, because Android supports getUserMedia – this means that users can start their camera and take a picture without ever having to leave the application. iOS doesn’t support getUserMedia yet, so they are sent off to the native iOS camera app to take their picture, which then gets passed back to the browser. Once I receive the picture, I load it into a canvas.

In the next step, users tap on their picture to select a color. The color’s hex code is sent straight to the Cooper-Hewitt API’s search method, where I search for similarly-colored objects that have an associated image. While waiting for a response from the API, I also tell the canvas to make every pixel within range of the selected color become transparent. When I get the image back from the API, I load it in behind the canvas and presto! It shows through where the selected color used to be. Finally, the image is titled based on the object’s creator and your current weather information.

It’s built using HTML, CSS and JavaScript. The original application had PHP to talk to the API but that’s since been ported to JavaScript since I now have the luxury of running the site on the Collections website itself where we have our own built-in API hooks.

Being a weekend project, there are some missing features – sharing is a big one – but I think it demonstrates the API’s ability to provide fresh, novel ways into a museum’s vast collection. Here’s the link again, and you can also find the source on GitHub.