On the Evaluation of Background Subtraction Algorithms without Ground-Truth

In video-surveillance systems, the moving object segmentation stage (commonly based on background subtraction) has to deal with several issues like noise, shadows and multimodal backgrounds. Hence, its failure is inevitable and its automatic evaluation is a desirable requirement for online analysis. In this paper, we propose a hierarchy of existing performance measures not-based on ground-truth for video object segmentation. Then, four measures based on color and motion are selected and examined in detail with different segmentation algorithms and standard test sequences for video object segmentation. Experimental results show that color-based measures perform better than motion-based measures and background multimodality heavily reduces the accuracy of all obtained evaluation results.