Home People Research Publications Demos
News Jobs Prospective
About Internal

Background Subtraction in Dynamic Scenes

Natural scenes are usually composed of several dynamic entities. Foreground objects often move amidst complicated backgrounds that are themselves moving, e.g. swaying trees, other objects such as a crowd, a flock of birds, moving water, waves, snow, rain and smoke-filled environments. Biological visual systems have evolved to be extremely efficient in discriminating between foreground and background objects in such dynamic scenes.

We propose a novel spatiotemporal saliency paradigm, inspired by biological vision, where background subtraction is inherent to the deployment of visual attention. In particular, background subtraction is equated to the detection of salient motion, for which we propose a solution based on the discriminant center-surround saliency discriminant center-surround saliency hypothesis.

Biological Motivation for Motion Saliency Previous approaches to background subtraction use motion similarity to identify foreground objects. However, psychophysics experiments of Nothdurft have shown that motion saliency and perceptual organization depend on measurements of local motion contrast. Here we replicate the classical motion pop-out experiment of Nothdurft to emphasize the biological motivation for the use of a center-surround framework to account for local contrast. [project]     

Background Subtraction Results The discriminant center-surround spatiotemporal saliency algorithm can be used to perform backgroun subtraction : background points are simply those having lowest center-surround saliency. We test the performance of our algorithm on 18 sequences collected on the web, many of which have significant camera motion. [demo & results]     

UCSD Background Subtraction Dataset The datasets consists of 18 video sequences. The frames of each sequence are provided in JPEG format. The groundtruth mask is also provided in the form of a 3D array variable in Matlab, where 1 indicates foreground and 0 indicates background. For some sequences, the number of frames of the groundtruth mask is smaller than the number of frames in the seqeunce. But the groundtruth is provided for frames starting from frame 1 of the sequence. [Sequences] [Groundtruth]


Publications: Spatiotemporal Saliency in Dynamic Scenes
V. Mahadevan, and N. Vasconcelos.
IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 32, no. 1, pp. 171-177, January 2010.
© IEEE [ps] [pdf]

On the plausibility of the discriminant center-surround hypothesis for visual saliency
D. Gao, V. Mahadevan, and N. Vasconcelos.
Journal of Vision, 8(7):13, 1-18, 2008.

Background subtraction in highly dynamic scenes.
V. Mahadevan and N. Vasconcelos.
In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
Anchorage, AK, 2008.
© IEEE [ps] [pdf]

The discriminant center-surround hypothesis for bottom-up saliency.
D. Gao, V. Mahadevan and N. Vasconcelos.
In Proc. Neural Information Processing Systems (NIPS),
Vancouver, Canada, 2007.
[ps] [pdf]

Contact: Vijay Mahadevan, Nuno Vasconcelos