Time-Lapse Videos of a Building Site

Jorge L. Reyes and Robert Fisher

School of Informatics - University of Edinburgh

August 2008

 

Abstract

 

This research explores the construction of Time-Lapse Videos from cluttered consecutive images.  It is focused on the analysis of the evolution of a construction project (The Informatics Forum). A capture process of 3 years produced a large image database which provides the basis for this work. Images were affected by different artifacts such as environmental and lighting changes, pedestrians, machinery and vehicles. Mechanisms have been developed to automatically render the images and reduce these adverse elements from the scenes to produce more 'truthful' videos which more accurately describe the building construction, thus contrasting with traditional techniques which only use captured raw images spaced at fixed intervals to produce the output video.

 

Reduction of noise, jitter removal and normalization of color levels are the methods employed in the preprocessing stage to clean the image database and provide a suitable starting point for the implementation of the time lapse video creation algorithms. Groups of images captured on the same day are processed to produce the frames of the final video which represent the background of the scene. Foreground elements and sudden weather transitions are removed. Two different methods to solve this are evaluated in this research. First, a multivariate median filter is employed as a deterministic approach to solve the 'day image' generation and processing. Then we employ a filter based on non-parametric kernel density estimation and a neural network classifier. Results show improvements on the time lapse videos. A visible reduction of the foreground elements and a higher stability in the image colors present a smoother version of the building construction.


On this page you can find: what we think are the best reconstructed videos, videos from different stages and algorithms and the source images from which the different stages were made. You can also find an MSc dissertation that describes the work in more detail.


The image data was obtained from three cameras (Doric:left: Corinthian:middle, Ionic:right) over about a 50 metre total baseline. The data is for 599 days over the period 12 October 2005 through 30 May 2008.

 

     Best Reconstructions (about wmv format, 30 MB each): Doric, Corinthian, Ionic

     J. L. Reyes-Ortiz, Probabilistic Time Lapse Video, MSc Dissertation, School of Informatics, Univ of Edinburgh, 2008.
Full text: PDF

 

 



Some Sample Images
DoricCorinthianIonic
Before
After


Video Gallery (avi format - about 45 MB each)

 

Traditional Approach

Time lapse video from original images

Corinthian

Doric

Ionic

Multivariate Median

Without image preprocessing

Corinthian

Doric

Ionic

Color normalization and noise reduction.

Corinthian

Doric

Ionic

Jitter removal, color normalization and noise reduction.

Corinthian

Doric

Ionic

Non-parametric kernel density estimation

Color normalization and noise reduction.

Corinthian

Doric

Ionic

Jitter removal, color normalization and noise reduction.

Corinthian

Doric

Ionic

Multivariate Median of Temporally Neighboring images

MVM images, neighborhood of 6 consecutive frames.

Corinthian

Doric

Ionic

KDE images, neighborhood of 6 consecutive frames.

Corinthian

Doric

Ionic

 

Source Images - one per day (at 12:00) for 599 days from 3 cameras. Some days don't have all 3 cameras. (Warning 250-350 MB zip files.) Pre-processing consists of: removal of camera jitter between frame capture, outlier pixel color replacement and some color normalization to account for weather, illumination and camera color-balancing effects. The 'Aligned' images are computed by a pixel-wise Median/Kernel' filter process over all 18 'Pre-processed' images from a day (every 10 minutes from 11:00-14:00). The 'Neighborhood of 5 frames' computes a pixelwise temporal median from +/- 2 frames about the current frame, using the 'Aligned' images as input.

 

Raw Images

Noon Images

Pre-Processed Images

Noon Images

Processed Images

Multivariate Median: Aligned

Multivariate Median: Neighborhood of 5 frames

Kernel Density Estimation: Aligned

Kernel Density Estimation: Neighborhood of 5 frames

 

Acknowledgements

Funding for the cameras came from the Edinburgh University small projects fund. The key contributors of effort to the project were: Scott Blunsden, Bob Fisher, Andrew Ferguson, Tim Lukins and Jorge Reyes-Ortiz.