This dataset is used in S.-H. Lee, J.-H. Kim, K. P. Choi, J.-Y. Sim, and C.-S. Kim, "Video saliency detection based on spatiotemporal feature learning," submitted to Proc. IEEE ICIP 2014. Video sequences have the resolution of 480 x 270 and consist of around 800 frames. The binary ground-truth maps are manually obtained for every 8 frame. This dataset can be used only for research purposes.

Dataset Download

 
We compare the saliency detection maps, which are obtained by GBVS [1], QUAT [2], ISS [3], CEN [4], SRE [5], and ROCT [6]. We use "Car," "Court," "Campus," and "Toy" sequences as the training sequences, and "Ball," "Stair," "Square," and "Road" sequences as the test sequences in [6].

Comparison on Training Sequences

 

Comparison on Independent Test Sequences

 

References

[1] J. Harel, C. Koch, and P. Perona, "Graph-based visual saliency," in Proc. Adv. Neural Inf. Process. Syst., Dec. 2006, pp. 246-257.
[2] B. Schauerte and R. Stiefelhagen, "Quaternion-based spectral saliency detection model and its applications in image and video compression," in Proc. European Conf. Comput. Vis., Oct. 2012, pp. 116-129.
[3] Y. Li, Y. Zhou, L. Xu, X. Yang, and J. Yang, "Incremental sparse saliency detection," in Proc. IEEE ICIP, Nov. 2009, pp. 3093-3096.
[4] Y. Li, Y. Zhou, J. Yan, and J. Yang, "Visual saliency based on conditional entropy," in Proc. Asian Conf. Comput. Vis., Sep. 2009, pp. 246-257.
[5] H. J. Seo and P. Milanfar, "Static and space-time visual saliency detection by self-resemblance," J. Vision, vol. 9, no. 12, pp. 1-27, Nov. 2009.
[6] S.-H. Lee, J.-H. Kim, K. P. Choi, J.-Y. Sim, and C.-S. Kim, "Video saliency detection based on spatiotemporal feature learning," submitted to Proc. IEEE ICIP 2014.