Evaluation Dataset
  • The MSRA10K dataset consists of 10,000 images along with pixel-wise ground truth masks. [Download]
  • The DUT-OMRON dataset consists of 5,168 images along with three types of ground truth, namely pixel-wise, bounding box and eye-fixation masks. [Download]
  • The SOD dataset is a multiple objects datasets built upon Berkeley segmentation dataset. [Download]
  • The iCoSeg dataset is another multiple objects dataset and originally a image co-segmentation dataset. [Download]
  • The ECSSD dataset is a complex scene saliency datsetset, consisting of 1; 000 images along with pixel-wise ground truth masks. [Download]

Evaluation Metric and Script (.m) [Download]  
  • Precison-Recall curver (PR)
  • Receiver Operating Characteristic (ROC) curve 
  • F-measure curve 
  • Overlapping Ratio (OR) curve
  • Area Under the ROC Curve (AUC)
  • Overlapping Ratio (OR) score 
  • Weighted F-measure (WF) score 
  • Mean Absolute Error (MAE) 
  • Average Precison (AP) 
  • Average Recall (AR) 
  • Average Fmeasure (AF) 

Evaluation Method