Ever since the Seattle’s Space Needle installed an HD 360 webcam on the top of the needle, I have been fascinated by the footage captured.
Over the past few months, I put together a time-lapse video of what the 360 webcam captured over the last 3 years. Check it out below, and continue reading for more details about it and to learn how it was made.
How was it made?
I started with two full panoramas a day for the last two years, more than 2000 panos. Then, the sequence was stabilized, as the camera shakes and moves over time, either by being knocked, or because of the wind and other forces of nature. The final step was to smooth temporally the sequence, to remove the variation due to weather and lighting conditions. The video below visualizes the process, but see the last section of this FAQ for more technical details.
What is the ghost cruise ship in front of downtown?
In the summer months, large cruise ships dock into the Seattle’s waterfront for one or two days a week. Because the cruise ship is only about a quarter of the time there, the temporal smoothing represents it as it was a 75% transparent, and therefore, it looks like a ghost cruise ship!
Is Mount Rainier visible in the time-lapse?
Yes! But barely. The atmospheric conditions have to be very clear in order to see Mount Rainier from the Space Needle, which is more than 80 miles away. The temporal regularization produces a “moving average” effect, which allows one to see it better in sunny periods. However, the mountain almost disappears from the time-lapse during the cloudy periods.
Why are the shadows of some buildings moving ?
The shadows cast by the buildings around Denny Ave are clearly visible during the summer months, as Seattle is very sunny. They move as the position of the sun at the capture times, 10:30 am and 2:30 pm, rises in the sky around the summer solstice.
What are other cool bits in the time-lapse?
There are countless details that are interesting to watch! Here are some of my favorites:
Have you done other time-lapse videos?
Yes, a few years ago I presented a method to generate time-lapse videos from Internet photos. To create the Seattle time-lapse, I used the same technique to smooth the sequence temporally. Check out the video below!
How many photos were used to make the time-lapse?
I used 2166 panoramas that correspond to two photos a day, taken at 10:30 am and 2:30 pm, every day for the last three years. Some of the photos are missing, probably because the camera was temporarily offline.
What is the original size of the time-lapse in pixels?
The time-lapse video is the same size as each panorama from the 360 webcam, i.e. 8283 pixels wide by 1080 pixels tall, about 9 megapixels.
How were the images processed?
The processing pipeline involves two steps:
- First, the raw panoramas images are motion stabilized using an optical flow technique, as the camera shakes and moves throughout the sequence due to wind and other natural forces. I chose a reference panorama and warped slices of all images to the reference panorama using CPM optical flow package (paper, github).
- Second, the stabilized sequence is temporally smoothed to remove flicker due to weather and lighting conditions. I use the technique shown in my previous work on time-lapse mining from internet photos (paper).
The processing happens in slices of 512 pixels (17 in total), and each slice takes takes 20 minutes for motion stabilization and one hour for temporal regularization, that is implemented in pycuda and runs on a GPU.