Supplementary Material for submission 2147: Traditional Saliency Reloaded: A Good Old Model in New Shape

Transcription

1 Supplementary Material for submission 247: Traditional Saliency Reloaded: A Good Old Model in New Shape Simone Frintrop, Thomas Werner, and Germán M. García Institute of Computer Science III Rheinische Friedrich-Wilhelms-Universität Bonn, Germany frintrop@iai.uni-bonn.de In the supplementary material, we show the complete set of evaluation plots. Sec. shows some evaluations of the most important parameters of our system and Sec. 2 shows the evaluation plots on four different datasets. Influence of Parameters Fig. shows an evaluation of the most important parameters of our system, namely the choice of the layers of the pyramids (left/middle) and the choice of the center-surround ratio for creating the surround pyramid from the center pyramid (right). For each plot, we keep the other parameters which are not evaluated fixed to the optimal value. Itcanbeseenthatthesystemisnotverysensitivetobothparameters:similar values also perform quite well. Even a single layer obtains reasonable results if the layer is chosen well (layer 2 or 3), which is interesting for high-speed performance on real-time systems. However, this diminishes also the ability of the system to generalize, thus we recommend to use at least 3 layers for good performance. MSRA MSRA MSRA.2. Layer 2 (347) Layer 3 (259) Layer (54) Layer 4 (947) Layer (924) Layer 5 (.943).2.2. Layer 4 (52) Layer 5 (785) Layer 4 (6) Layer 3 (96) Layer 2 3 (7534) Layer 2 4 (724) Layer 2 5 (56) Layer 2 (458) Layer 3 5 (4777).2.2. Center 3, Surround 5 (82) Center 3, Surround 3 (52) Center 2, Surround 3 (432) Center 4, Surround 4 (23) Center 3, Surround (35) Center 5, Surround 5 (657) Center 5, Surround 3 (338) Center 5, Surround 2 (5) Center, Surround 3 (9749) Center 5, Surround (9473).2 Fig.: Evaluation of parameters for our saliency system on the MSRA- dataset. Left/middle: effect of layer choice of the pyramids. Right: Effect of Center-Surround Ratio parameter. 2 Comparative Results on Benchmarks We show the results of our VOCUS2 saliency system ( = pixel-precise; /5/$3. = pixel-precise with location 25 prior; IEEE = segment-based generated

2 2 CVPR-5 submission ID 247 with object proposals) obtained in four different datasets for salient object segmentation: MSRA k [5] in Fig. 2, ECSSD [3] 2 in Fig. 3, PASCAL-S [] 3 in Fig. 4 and SED [3] 4 in Fig. 5. We compare our method with the following saliency methods: Itti s [7], the SaliencyToolbox()[2], HZ8[6], [4], AC9 [], AC [2], [9], HSaliency [3], Yang 23 [4], and [8]. In terms of weighted F-measure [] the results show that our method that integrates segmentation () is the best in the MSRA k (Fig. 2, left), PASCAL-S (Fig. 4, left) and SED2 (Fig. 5, bottom left) datasets; it is third on SED (Fig. 5, top left) and second on ECSSD (Fig. 3, left). The weighted F-measure plots are separated into the methods that incorporate segmentation (left hand side) and those that do not (right hand side). In the precision-recall plots, our method outperforms all other pixelprecise methods, except in SED2, in which the approach is better. This can be easily explained since the SED2 dataset incorporates less center bias since each frame contains 2 objects. Among the segment-based approaches, the V2- Prop method is usually ranked 2nd or 3rd. The gap between our method and the top methods results at least partly from the limitations of the precision-recall plots as announced by []. Please note also that we intentionally optimized our code only on the MSRA- dataset and not, as some other methods, on all test datasets..2. Yang23 MSRAk AC AC9 HZ8.2. V2-prop. (546) (6227) V2 (966) (979) (4789) Yang23 (9847) (748) (3647) AC (832) AC9 (967) HZ8 (7969) (5646) (7245) MSRAk..2 Fig. 2: MSRAk Dataset results. Left: weighted F-measure from []. Right: -recall curves (AUC-values in parentheses). : pixel-precise : pixel-precise saliency map with location prior. :

3 weighted F-measure CVPR-5 submission ID CSSD.2. V2-P Yang23 V2-Ba ECSSD AC AC9 HZ8.2.. (9575) (5245) VOCUS2-Basic (333) (995) (439) Yang23 (98) (3338) (4898) AC (7224) AC9 (824) HZ8 (328) (99) (28)..2 Fig. 3: ECSSD Dataset results. Left: weighted F-measure from []. Right: -recall curves (AUC-values in parentheses). : pixel-precise : pixel-precise saliency map with location prior. :.2. Yang23 PASCAL-S AC AC9 HZ8.2. V2-prop. (583) (7724) V2 (574) (7226) (683) Yang23 (963) (5874) (383) AC (5249) AC9 (895) HZ8 (3796) (9796) (8) PASCAL-S..2 Fig. 4: PASCAL-S Dataset results. Left: weighted F-measure from []. Right: -recall curves (AUC-values in parentheses). : pixel-precise : pixel-precise saliency map with location prior. :

4 a thweig e s -df 4 CVPR-5 submission ID 247 re Yang23 Yang23 S D S D2 AC AC9 HZ8 AC AC9 HZ V2-prop. (8242) (25) V2 (2362) (746) (5836) Yang23 (3565) (4223) (552) AC (2354) AC9 (8859) HZ8 (895) (2822) (58) S D..2 S D2 V2-prop. (7) (5797) V2 (7567) (7783) (9) Yang23 (235) (4464) (74) AC (72) AC9 (3499) HZ8 (943) (73) (48)..2 Fig. 5: SED and SED2 Datasets results. Left: weighted F-measure from []. Right: -recall curves (AUC-values in parentheses). : pixelprecise : pixel-precise saliency map with location prior. : proposal maps obtained from generating object proposals and fusing them to a

5 CVPR-5 submission ID References. R. Achanta, S. Hemami, F. Estrada, and S. Süsstrunk. Frequency-tuned salient region detection. In CVPR, R. Achanta and S. Süsstrunk. Saliency Detection using Maximum Symmetric Surround. In ICIP, S. Alpert, M. Galun, R. Basri, and A. Brandt. Image segmentation by probabilistic bottom-up aggregation and cue integration. In CVPR, N. D. B. Bruce and J. K. Tsotsos. Saliency, attention, and visual search: An information theoretic approach. J. of Vision, 9(3), M.-M. Cheng, N. J. Mitra, X. Huang, P. H. S. Torr, and S.-M. Hu. Global contrast based salient region detection. TPAMI, 37(3): , X. Hou and L. Zhang. Dynamic visual attention: Searching for coding length increments. In Advances in Neural Information Processing Systems, L. Itti, C. Koch, and E. Niebur. A model of saliency-based visual attention for rapid scene analysis. TPAMI, 2(), H. Jiang, J. Wang, Z. Yuan, Y. Wu, N. Zheng, and S. Li. Salient object detection: A discriminative regional feature integration approach. In CVPR, D. A. Klein and S. Frintrop. Salient Pattern Detection using W 2 on Multivariate Normal Distributions. In DAGM-OAGM, 22.. Y. Li, X. Hou, C. Koch, J. M. Rehg, and A. L. Yuille. The secrets of salient object segmentation. In CVPR, 24.. R. Margolin, L. Zelnik-Manor, and A. Tal. How to evaluate foreground maps? D. Walther and C. Koch. Modeling attention to salient proto-objects. Neural Networks, Q. Yan, L. Xu, J. Shi, and J. Jia. Hierarchical Saliency Detection. In CVPR, C. Yang, L. Zhang, H. Lu, X. Ruan, and M.-H. Yang. Saliency Detection via Graph-based Manifold Ranking. In CVPR, 23.