Posted by Florian Wardell | 4 comments
Paris in 26 Gigapixels
Have you ever been to Paris? Have you ever walked along the Seine, climbed on the Eiffel tower, have you ever tried to get lost in the Marais?
Well, if you haven’t, do not worry. From now on, you can roam Paris from your computer, either with Google Street View, or more recently, with Paris 26 Gigapixel, a single gigantic, huge, monstrous, gargantuous, colossal picture.
Of course, it’s not the same as being there, but thanks to the picture’s gigantic size – 26 Gigapixels – you can literally spend a day looking at the city. A gigapixel, in case you’re wondering, equals a billion pixels. The 26 gigapixel image, the largest ever created, could be printed on a 6500 sq.ft support, roughly the size of two football fields, just to put things into perspective.
But the spectacular result is just as impressive as the realization of the project itself, which was executed in three distinct steps.
On September 8th 2009, the team, lead by Arnaud Frich one of the most famous panoramic photographers in France, and Martin Loyer a photographer specialized in cityscapes, gathered at the tour St. Sulpice, in the 6th arrondissement. With them, two Canon 5D Mark II (21.1 MP each), mounted on a custom-made panoramic head. They chose to use two cameras as to reduce shooting time, and were set as follows:
- 300mm f4.0 with a 2x tele-converter (equivalent 600mm f8.0)
- Manual focus with Live View (zoom and pan) to get a very precise focus control.
- Priority aperture F13 (to have a bigger depth of field)
- ISO 800
- Speed 1/800 in order to reduce the heat haze
- Recording in RAW
- Compact Flash 16GB
- Triggering by motorized heads.
It was decided that the images would overlap by 30%, and that there would be 138 columns and 17 rows, which totals is 2346 images. The total shooting time was about 5 hours, which was much longer than the 2:30 hours they originally planned: one of the cameras failed, slowing everything down. Once all the pictures were gathered and saved, they proceeded to the next step.
Arnaud Frich explains: “The stitching itself remains a bit complicated with a project of that size. Even if the use of a motorized panorama head helps a lot during the shooting, the stitching still doesn’t give perfect results because of certain points : first, robustness of the motorized panorama head (we had a break after each row to manually correct the focus) and second, because of the repeating patterns in some buildings.”
Not only this, but they also found that they missed a few shots: “We could notice a shift in the rows. It comes from a shooting issue: we missed 3 images during the shooting at the edge of one row and at the beginning of the next one. It’s the kind of stuff you don’t notice live [...], but only when you come back at the office. By the way: it was really a nightmare to figure out which images were missing as they were not named after the actual shooting order. Our luck was that the missing images were at the edges of the panorama so that the consequences were really low. We cropped that zone on the left part of the panorama. Yes! We could have 2 more columns in that panorama because we used only 136 out of 138 shot columns.” This could have been a 27 gigapixel image.
“The first optimization of the project with all the images was quite good but not perfect because of some bad links remaining in the editor. And it’s easy to guess why: repeating patterns in a urban landscape are really common; there is nothing more similar than a window and the one just nearby. This can lead to alignment issues in detected control points between neighbor images and that’s what we can see in the following screenshot [...].”“This screenshot shows a standard detection without using the Clauss plugin of Autopano Giga. You can see a lot of red zones where all false links have been found. In practice, coping with these false links is quite quick. What was far longer is to cope with out-of-focus zones (for example with the top of the roof of St-Sulpice tower). No miracle here, you need to manually move around the images with the move mode in the editor. A full day of work was needed to do that because even if each operation was really easy to do, it took time for each move operation to happen on a panorama of that size. [...] But we still need to cope with color correction. The lighting of the scene has changed a lot during such a long shooting so we needed to be careful about that. By the way: yes, if you look accurately in the image, you can see that shadows moved in that picture.”
The final raw image was 354159 x 75570 = 26,763,795,630 pixels big. Once every detail was fixed and taken care of, the photographers proceeded to the last step.
Is the last step before publishing the image on the web. Basically, it’s the process of merging all the images into one by following the stitching instructions discussed above. Two years ago, a 13 gigapixel image of Harlem took 48h hours to render. So how long did it take to render an image twice as big? The answer is 3 hours and 14 minutes, an incredible time which was achieved thanks to borrowed hardware: an SR2600UR Intel Server System that included 2 Intel Xeon processors 5500 series ans 6 SSD hard drives of 160 GB each (16 cores, 24 GB RAM, 1 TB SSD). Yes, this thing could run Crysis.
You can read more about this in the case study written by Intel.
But that was the easy part: because of legal reasons, they had to mask all privacy invading sights in the picture (we’re talking about Paris, after all), and tweak the colors a little bit: remember, the shooting took more than 5 hours, a time-frame during which light can change dramatically.
Because photoshop only supports images up to 300,000 by 300,000 pixels, they had to re-slice the pictures using an internally developed tool that will soon become commercially available, edit them, and then re-stitch them.
The final task was to upload everything onto the servers, which wasn’t easy: their website is made of more than 600,000 files!
The website is nice by the way, you have the choice between a flash and HD version, an automated tour, information regarding the monuments, and the music from “Amelie Poulain” in the background, unfortunately.
This is the kind of project I’m fascinated about. If you think about it, this is absolutely pointless, right? So much work for a single, ridiculously big picture. Why? Because we can.
You’ll see, one day point and shoot cameras will have a native resolution of 50 gigapixels.
I hope you like the picture. It’s not so much art as it is an incredible technical achievement. Oh, and if you have time to kill, why don’t you try to find the 10 surprises that the photographers have hidden in the picture? I already found the U.F.O.s.