People-Bakground Segmentation with Unequal Error Cost
This page presents sample results of our proposal of a people-background segmentation approach.
The main contribution is a people-background segmentation approach with unequal error cost between classes in order to ensure that there are no people (or body parts) classified as background (bias on background). The proposed approach first performs the detection of multiple body parts and then the detection confidence maps are extended according to the space that should occupy the detected parts and combined in one final background confidence map. Finally, the corresponding background segmentation mask is generated.
Related publication
A. García-Martín, A. Cavallaro, J. M. Martínez. People-background segmentation with unequal error cost.In Proc. of the IEEE International Conference on Image Processing, Orlando (FL, USA), 30-3 September-October 2012.
Contact Information
Álvaro García-Martín -
show
email
Dataset links
TUD-Campus and TUD-Crossing are available here
TRECVID dataset is available here
PETS2006 dataset is available here
PETS2009 dataset is availabe here
AVSS
dataset is
available here
In order to compare the different approaches results, we show the results in terms of ROC curves with different false positives penalty factors: 1, 2, 4 and 10 and as we are interested in knowing the operating range of the algorithm in which we have no false detections (pixels who belong to a person incorrectly classified as background) so we also show the average number of false positive detections according to the chosen binarization threshold.
ROC curves |
||||
False positive penalty factor |
1 |
2 |
4 |
10 |
TUD Campus |
||||
TUD Crossing |
||||
TRECVID |
||||
PETS2006 |
||||
PETS2009 |
||||
AVSS |
False positives curves |
||
TUD Campus |
TUD Crossing |
TRECVID |
PETS2006 |
PETS2009 |
AVSS |
The available code here