Data
Although images from MSRA-1000 have a large variety in their content, background structures are primarily simple and smooth. To represent the situations that natural images generally fall into, we extend our Complex Scene Saliency Dataset (CSSD) in [1] to a larger dataset (ECSSD) [2] with 1000 images, which includes many semantically meaningful but structurally complex images for evaluation. The images are acquired from the internet and 5 helpers were asked to produce the ground truth masks. Several examples with their corresponding masks are shown above.
Evaluation
We carried fix-threshold experiments on our ECSSD dataset for several state-of-the-art methods. The precision-recall and f-measure curves are plotted in the following figures (a) and (b). The top two curves correspond to HS [1] and CHS [2] (orange one) – our extended model.
 |
 |
||
| (a) | (b) |
Results
For each example below, we show results by HS [1] and CHS [2] – our extended model.
Downloads
|
ECSSD (1000 images) |
CSSD (200 images) |
References
[1] Qiong Yan, Li Xu, Jianping Shi, Jiaya Jia. Hierarchical Saliency Detection. CVPR 2013.
[2] Jianping Shi, Qiong Yan, Li Xu, Jiaya Jia. Hierarchical Image Saliency Detection on Extended CSSD. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), accepted.