Comparing Figures of Merit and Image Datasets for Evaluation of Salient Region Detection Algorithms

G.B. Borba, H.R. Gamba (Brazil), O. Marques, A. Colic, and V. Adzic (USA)


Image datasets, salient regions detection, salient regions evaluation, visual attention model.


There is a widely recognized need for rigorous quantitative performance assessment in many topics within computer vision, which leads to the adoption of appropriate figures of merit, image datasets, and benchmarks (e.g., in the form of ground truth). In this paper, we present a quantitative evaluation of different approaches for the salient regions detection task using different public image datasets and figures of merit. We briefly describe the suitability of the preselected datasets and the candidate figures of merit – recall, precision, F1 and area of overlap – in the context of the current task. Finally, the results of experiments using the salient region detection algorithms and different datasets are presented and commented. From the obtained data, we suggest one most adequate dataset and discuss why the current figures of merit, although acceptable and suitable for the task, are not ideal.

Important Links:

Go Back