Immersive 3D experience at the CNAM: witness construction of the Statue of Liberty

140 years ago, the Statue of Liberty was unveiled in New York. To mark this anniversary, the “Lady Liber­té” exhib­i­tion will open in Septem­ber 2026 at the Musée des Arts et Méti­ers in Par­is. Vis­it­ors will be taken behind the scenes of the con­struc­tion of this colossal struc­ture, designed and built in France. Com­bin­ing phys­ic­al sets and three-dimen­sion­al anim­a­tions, the exhib­i­tion offers an immers­ive exper­i­ence akin to vir­tu­al real­ity, but without the need for a head­set. It thus illus­trates the tech­no­lo­gic­al advances in the 3D mod­el­ling sec­tor, which has been sig­ni­fic­antly revital­ised, in par­tic­u­lar by the rise of 3D Gaus­si­an Splat­ting, a fam­ily of meth­ods enabling highly real­ist­ic renderings.

Donated by the French people to mark the cen­ten­ary of US inde­pend­ence (though delivered a full 10 years late!) and designed by the sculptor Auguste Bartholdi in col­lab­or­a­tion with the engin­eer Gust­ave Eif­fel in Par­is, the Statue of Liberty is a feat of engin­eer­ing. Com­pris­ing a cop­per shell rest­ing on an intern­al met­al frame­work, it stands approx­im­ately 46 metres tall. How was such a colossal monu­ment, yet one with such pre­cise phys­ic­al pro­por­tions, created?

Vis­it­ors to the “Lady Liber­té” exhib­i­tion will be invited to wander through a series of suc­cess­ive rooms in which they will observe the stages of the statue’s con­struc­tion, from the draft­ing of the plans to the pre­par­a­tion of the ship set­ting sail for New York laden with the vari­ous parts of the struc­ture. Through­out the exhib­i­tion, vis­it­ors will recog­nise cer­tain icon­ic objects, such as the hand hold­ing the torch, recre­ated in 3D.

And the exhib­i­tion takes the immers­ive exper­i­ence even fur­ther: “In one of the rooms, the real set­ting of Bartholdi’s work­shop blends with a pro­jec­ted back­drop: the sculptor appears and addresses vis­it­ors from a mezzan­ine via a 3D pro­jec­tion using a sys­tem com­bin­ing two screens,” explains Hubert Naudeix, founder of Aristeas, the com­pany respons­ible for the exhibition’s 3D recon­struc­tions and anim­a­tions. To cre­ate this scene, an act­or was filmed in cos­tume, then Auguste Bartholdi’s face, recre­ated from pho­to­graphs, was integ­rated into the video using arti­fi­cial intel­li­gence techniques.

“It’s like a vir­tu­al real­ity exper­i­ence, except you don’t have a head­set; you’re partly in real set­tings. The video, light­ing and sound must be per­fectly syn­chron­ised,” he con­tin­ues. Fur­ther on, vis­it­ors will find them­selves “as if inside the statue: they will see all the met­al plates being put in place, assembled…” adds Hubert Naudeix.

The “Lady Liber­té” exhib­i­tion thus com­bines vari­ous 3D mod­el­ling tech­niques. For example, two main tech­niques were used to recon­struct objects in 3D: pho­to­gram­metry and laser scan­ning. Pho­to­gram­metry util­ises a series of pho­to­graphs taken from dif­fer­ent angles. Dis­tinct­ive points are iden­ti­fied across the images, enabling the pos­i­tion of the cam­er­as to be estim­ated and a point cloud to be recon­struc­ted, and sub­sequently a tex­tured mesh.

The under­ly­ing geo­met­ric prin­ciple is that of tri­an­gu­la­tion. Laser scan­ning, on the oth­er hand, relies on the use of LiD­AR (Light Detec­tion and Ran­ging), which meas­ures the dis­tance between the sensor and vis­ible sur­faces by ana­lys­ing the return of a laser sig­nal. It thus pro­duces a point cloud rep­res­ent­ing the vis­ible sur­faces of the object or scene.

Today, these tech­niques can be com­bined with arti­fi­cial intel­li­gence tools that optim­ise and accel­er­ate the 3D recon­struc­tion pro­cess. Nev­er­the­less, pho­to­gram­metry and laser scan­ning have cer­tain lim­it­a­tions. For example, “pho­to­gram­metry recon­structs geo­metry based on image cor­res­pond­ences. It can pro­duce highly accur­ate mod­els, but the res­ult depends on the qual­ity of the images, the tex­ture of the sur­faces and the abil­ity to identi­fy the same details across mul­tiple images,” explains Mari­us Preda, seni­or lec­turer at Télé­com Sud­Par­is (IP Paris).

Fur­ther­more, “trans­lu­cent, shiny or highly reflect­ive sur­faces com­plic­ate the recon­struc­tion, as their appear­ance var­ies depend­ing on the view­ing angle and light­ing, and char­ac­ter­ist­ic points are more dif­fi­cult to identi­fy con­sist­ently,” he con­tin­ues. As for the laser tech­nique, it is “often more geo­met­ric­ally pre­cise, but data acquis­i­tion can be time-con­sum­ing and remains best suited to stat­ic objects or scenes; move­ment dur­ing cap­ture makes recon­struc­tion much more dif­fi­cult,” adds the researcher.

A new approach is now mak­ing it pos­sible to sig­ni­fic­antly improve the visu­al ren­der­ing of 3D scenes: 3D Gaus­si­an Splat­ting. “A real revolu­tion is under­way!” enthuses Hubert Naudeix, who is eager to start using this solu­tion, which is cur­rently in the exper­i­ment­al phase. Com­bin­ing cut­ting-edge com­puter graph­ics meth­ods with machine learn­ing tech­niques, the 3D Gaus­si­an Splat­ting meth­od achieves “optim­al visu­al qual­ity with a high level of detail,” explains the founder of Aristeas.

A level of visu­al real­ism that sur­passes the usu­al stand­ards of main­stream 3D. How? By using a dif­fer­ent rep­res­ent­a­tion of the scene. Where­as pho­to­gram­metry often pro­duces a mesh of tri­angles to which a tex­ture is applied, 3D Gaus­si­an Splat­ting rep­res­ents the scene as a set of small “spots” called 3D “Gaus­si­an prim­it­ives”. Each is described in par­tic­u­lar by its pos­i­tion, size, ori­ent­a­tion, opa­city and col­our. It there­fore does not remove all geo­metry: rather, it replaces the tri­an­gu­lar mesh with a volume rep­res­ent­a­tion optim­ised for rendering.

“It’s exactly the same prin­ciple as pixels in an image: you don’t see the pixels when you look at it, but if you zoom in, you see them side by side,” explains Hubert Naudeix. “The her­it­age object is thus visu­ally recon­struc­ted, and in a way that is very faith­ful to real­ity, par­tic­u­larly in terms of col­our prop­er­ties and tex­tures,” adds Mari­us Preda. Whilst the prin­ciples of “splat­ting” and “Gaus­si­an prim­it­ives” have long exis­ted in com­puter graph­ics, 3D Gaus­si­an Splat­ting is enjoy­ing a resur­gence of interest thanks to the work of Inria’s Graph­Deco research team¹, which, in 2023, pro­posed a meth­od enabling real-time rendering.

Today, com­pan­ies in the audi­ovisu­al and video game sec­tors are rush­ing to test and devel­op their own solu­tions based on this innov­at­ive tech­no­logy. And with a view to demo­crat­ising access to advanced 3D tech­no­lo­gies, at Télé­com Sud­Par­is², Mari­us Preda and his team are exper­i­ment­ing with these approaches as part of the “Museum Meta­verse” plat­form, to improve the visu­al qual­ity of 3D her­it­age mod­els and make these tech­no­lo­gies more access­ible to part­ner museums.

Through the “ICON” pro­ject, launched this year as a con­tinu­ation of the “Museum Meta­verse” pro­ject and sched­uled to run for three years, the research­er aims to “accel­er­ate the cre­ation of con­tent based on Gaus­si­an Splat­ting, in line with sceno­graphy require­ments.” This brand-new approach could address the major chal­lenge facing museums: pre­serving the authen­ti­city of the cul­tur­al exper­i­ence in a vir­tu­al environment.

Olympe Delmas