Mona's $100M underground library lets books roam free — computer vision keeps track

A $100M underground library at Mona in Tasmania had to work as a real library — a team of librarians, a thousand visitors a day, 30,000+ books — while letting books be rearranged at will to spark new connections.

Abandon classification entirely: computer vision reads bare book spines so any librarian can reshelve books anywhere, and lights plus an app reveal the curatorial links.

A computer-vision system trained on natural book spines tracks every book's shelf position in near real time; touching a displayed book lights up related titles ('live bays'), and Mona's O guide gives walking directions like 'third row from the top, 8th book to the left'.

Per the project team's account, over 85% of museum visitors came through Phrontisterion, versus around 25% for the old Mona library; librarians keep re-curating the shelves as the system follows along.

Mona in Tasmania was getting a $100 million underground library, Phrontisterion, briefed by founder David Walsh in one long text message: 'We can, at whim, move art around to say what we like to say with art... Why can't we do this with books? Because, if we move them around, we can't find them anymore.' He wanted 'a system where we can move books arbitrarily, but still know where they are', with librarians as curators. Art Processors, born out of Mona's 2011 founding, and Open Field Studio took it on — a working library with a thousand visitors a day, not an installation.

The team rejected the first computer-vision proposal, which needed markers glued to spines and years of preparation: 'We would have spent years manipulating and adjusting the collection to make it work with the technology.' The fixed requirement became 'the book identifies itself'. After roughly a year of configuration, a system reading natural book spines delivered the needed accuracy with far less training. Prototypes scaled from 100 books on an IKEA shelf, to 500, to the whole old library of about 5,000 — and what they built to demo tracking started suggesting the visitor experience itself.

The shelf coordinates did double duty: projectors cast light on related books ('live bays' — touch a cover, its neighbours in meaning glow), while Mona's O guide extends associations across all 30,000+ books. Per the team's account in Museum-iD, over 85% of museum visitors came through Phrontisterion once it opened, against around 25% for the old Mona library — with Egyptian coffins, David Bowie's handwritten Starman lyrics and librarians free to keep re-shelving the collection around them.

David Walsh's brief reframed the problem as the museum's own superpower: art gets rearranged to make arguments, so books should too — provided nothing is ever lost.

Marker-based tracking would have bent the collection to the technology for years; requiring 'the book identifies itself' inverted the dependency and cut training data needs.

Roughly a year of iteration over three growing prototypes (100, 500, 5,000 books) let the accuracy bar stay 'reasonable' while the experience design matured with it.

Shelf coordinates doubling as projection addresses turned a tracking system into a storytelling layer without extra infrastructure.

When the technology demand ('books must identify themselves') is held fixed, the tooling can stay provisional: iterate on natural spines rather than bend the collection to markers.

The library is open and its arrangement keeps changing: books can be brought together for what they say to one another and moved on when there is something else to say. The account by Edward Blake (Open Field Studio) and Julien de Sainte-Croix (Art Processors) was published by Museum-iD in October 2026.