Test Images for Wallflower Paper
This page contains the test images that we used for our Wallflower paper.* We tested several different background maintenance algorithms on seven different image sequences. For each test image in the sequence, the output of the algorithm was an image segmented into background and foreground pixels.
Each of the seven different test sequences represents a different, potentially problematic scenario for background maintenance. In general we used the first 100 images in each sequence for training. We then ran each algorithm on the remainder of the sequence up to a prespecified evaluation image. We compared the algorithm's result on the evaluation image to a hand-segmented version of the evaluation image.
We have created one .zip file for each of the seven test scenarios. Each .zip file contains the training images, test images, and hand-segmented evaluation image for comparison, all in BMP format. Each .zip file also contains a file called "script.txt" that gives a short description of the events in the sequence and specifies which images are used for training, testing, and evaluation.
The table below shows sample images and gives links to the ftp site with the .zip files.
For questions about these test images, please contact John Krumm at Microsoft Research, email@example.com.
|Scenario||First Training Image||Evaluation Image||Hand-Segmented Evaluation Image||Notes||.zip File|
|Moved Object||Chair is moved||
|Time of Day||Lights gradually brighten simulating moving sun||
|Light Switch||Person switches on light||
|Waving Trees||Tree moving in background||
|Camouflage||Static foreground occludes dynamic background||
|Bootstrapping||No pure background training available||
|Foreground Aperture||Trained on sleeping person who wakes up||
Kentaro Toyama, John Krumm, Barry Brumitt, Brian Meyers, "Wallflower: Principles and Practice of Background Maintenance", Seventh International Conference on Computer Vision, September 1999, Kerkyra, Greece, pp. 255-261, IEEE Computer Society Press.
back to top