Reflections Removal Algorithm Evaluation

In this page there is the data used for the experimentation of the paper:

D. Conte, P. Foggia, G. Percannella, M. Vento (2012): Removing object reflections in videos by global optimization. In: Circuits and Systems for Video Technology, IEEE Transactions on, 22 (11), pp. 1623 – 1633, 2012, ISSN: 1051-8215.

The dataset used for the tests is composed by four publicly available videos, all referring to indoor scenarios with reflecting floorings.

• The first video (hereinafter referred to as AVSS) belongs to the dataset published during the International Conference on Advanced Video and Signal-based Surveillance AVSS 2007; the scene shows a subway station.

  http://www.eecs.qmul.ac.uk/~andrea/avss2007_d.html

• The second video sequence (hereinafter referred to as CAVIAR) is taken from the public dataset CAVIAR, widely used for video surveillance systems evaluation; the scene shows a shopping mall.

  http://homepages.inf.ed.ac.uk/rbf/CAVIARDATA1/ – “OneLeaveShopReenter1cor” sequence

• The third and the forth video sequences (hereinafter referred to as PETS-1 and PETS-2), belong to the dataset published at the International Workshop on Performance Evaluation of Tracking and Surveillance PETS2006; in particular we have used the S1-T1-C1 and S1-T1-C3 sequences.

  http://www.cvg.rdg.ac.uk/PETS2006/

For each video, a ground truth has been produced by inspecting the objects detected in each frame and choosing by hand the expected cut line. The objects missed by the detection algorithm, as well as the wrongly detected ones (those corresponding to partial detections of the persons) have been discarded, since reflection removal methods cannot recover from such errors.

For some videos, we don’t have considered all the frames of the sequence because some parts of the video was not significant. In the following table the first and the last considered frame for each video are shown. In the table two links for each video are shown. The meaning of the two files is the following:

• the first link is a file containing the bounding boxes of the considered objects of the video before application of any reflection removal algorithm
• the second link is a file containing the ideal bounding boxes of the considered objects of the video

The files are in the viper format.