Leveraging Colour Segmentation for Upper-Body Detection
AbstractThis paper presents an upper-body detection algorithm that extends classical shape-based detectors through the use of additional semantic colour segmentation cues. More precisely, candidate upper-body image patches produced by a base detector are soft-segmented using a multi-class probabilistic colour segmentation algorithm that leverages spatial as well as colour prior distributions for different semantic object regions (skin, hair, clothing, background). These multi-class soft segmentation maps are then classified as true or false upper-bodies. By further fusing the score of this latter classifier with the base detection score, the method shows a performance improvement on three different public datasets and using two different upper-body base detectors, demonstrating the complementarity of the contextual semantic colour segmentation and the base detector.
Stefan Duffner, Idiap Research Institute, Switzerland / LIRIS, INSA de Lyon, France
Jean-Marc Odobez, Idiap Research Institute, Switzerland
PaperS. Duffner and J.-M. Odobez, Leveraging Colour Segmentation for Upper-Body Detection, In Pattern Recognition, Vol. 47 (6), pp. 2222-2230, 2014.
DatasetsWe used three evaluation datasets:
Note: these datasets can only be used for research purposes in the computer vision domain!
AnnotationThe upper-body annotation can be downloaded here. Each line corresponds to a test image, having the form:
filename number_of_upper_bodies top_left_x1 top_left_y1 width1 height1 top_left_x2 top_left_y2...Please cite the paper above if you use one of our datasets or our annotation!
These are the spatial PIM prior models that have been used for the four classes: skin, hair, clothing, background. White represents a probability of 1.0 and black a probability of 0.0.
ResultsWe compared the proposed approach with the base detector's performance (Calvin Detector: Eichner, Ferrari) and a similar approach to ours (D. Ramanan, Using segmentation to verify object hypotheses, CVPR, 2007). Here are the precision/recall curves:
More results can be found in the paper.