Page 45 - Read Online

Page 514 Liu et al. Intell Robot 2024;4(4):503-23 I http://dx.doi.org/10.20517/ir.2024.29

Table 2. Comparison with existing methods in terms of #Param, FLOPs, FPS, maxF, avgF, MAE, and S in general scenarios
#Param FLOPs DUTS-TE DUT-OMRON ECSSD
Methods FPS
(M) (G) maxF↑ avgF↑ MAE↓ S↑ maxF↑ avgF↑ MAE↓ S↑ maxF↑ avgF↑ MAE↓ S↑
Heavyweight method (#Param > 10 M)
CPD [51] 47.85 59.5 42 0.861 0.805 0.043 0.866 0.794 0.747 0.056 0.818 0.930 0.917 0.037 0.905
U2Net [52] 44.02 58.8 45 0.873 0.792 0.045 0.874 0.823 0.761 0.055 0.847 0.951 0.892 0.033 0.928
UCF [53] 29.47 61.4 12 0.772 0.631 0.112 0.782 0.730 0.621 0.120 0.760 0.901 0.844 0.069 0.883
Amulet [25] 33.15 45.3 10 0.778 0.678 0.085 0.804 0.743 0.647 0.098 0.781 0.913 0.868 0.059 0.894
DSS [54] 62.23 114.6 7 0.825 0.720 0.056 0.826 0.781 0.656 0.066 0.790 0.921 0.842 0.056 0.879
PiCANet [13] 32.85 19.7 5 0.851 0.759 0.051 0.869 0.794 0.717 0.065 0.832 0.931 0.886 0.046 0.917
BASNet [55] 87.06 127.3 36 0.859 0.791 0.048 0.866 0.805 0.756 0.057 0.836 0.938 0.880 0.037 0.916
PoolNet [56] 53.63 123.4 39 0.866 0.819 0.043 0.875 0.791 0.752 0.057 0.829 0.934 0.919 0.048 0.909
MINet [14] 162.38 87.1 43 0.877 0.823 0.039 0.875 0.794 0.741 0.057 0.822 0.943 0.922 0.036 0.919
VST [57] 44.48 23.2 40 0.877 0.818 0.037 0.896 0.800 0.756 0.058 0.850 0.944 0.920 0.033 0.932
PFSNet [58] 31.18 37.6 44 0.898 0.846 0.036 0.890 0.823 0.774 0.055 0.852 0.952 0.932 0.031 0.927
ICON [59] 33.09 20.9 57 0.892 0.838 0.037 0.888 0.825 0.772 0.057 0.844 0.950 0.928 0.032 0.929
MENet [60] - - 45 0.912 0.893 0.028 0.905 0.834 0.818 0.045 0.850 0.955 0.942 0.031 0.928
TSERNet [61] 189.64 203.6 35 0.861 0.798 0.046 0.864 0.818 0.768 0.056 0.837 0.945 0.922 0.031 0.930
A3Net [62] 17.00 34.1 46 0.843 0.769 0.052 0.863 0.801 0.739 0.062 0.831 0.937 0.913 0.045 0.912
Avg-heavy 62.00 72.6 34 0.856 0.785 0.051 0.863 0.797 0.735 0.064 0.825 0.936 0.902 0.042 0.914
Lightweight method (#Param <= 10 M)
HVPNet [16] 1.24 1.1 55 0.839 0.749 0.058 0.849 0.799 0.721 0.065 0.831 0.925 0.889 0.052 0.904
SAMNet [17] 1.33 0.5 37 0.835 0.745 0.058 0.849 0.797 0.717 0.065 0.830 0.925 0.891 0.050 0.907
CSNet [15] 0.14 1.5 48 0.819 0.687 0.074 - 0.792 0.675 0.081 - 0.916 0.844 0.065 -
Ours 2.29 1.5 62 0.845 0.773 0.054 0.866 0.804 0.742 0.061 0.833 0.934 0.907 0.047 0.913
#Param FLOPs DUTS-TE DUT-OMRON ECSSD
Methods (M) (G) FPS maxF↑ avgF↑ MAE↓ S↑ maxF↑ avgF↑ MAE↓ S↑ maxF↑ avgF↑ MAE↓ S↑
Heavyweight method (#Param > 10 M)
CPD [51] 47.85 59.5 42 0.866 0.820 0.071 0.847 0.924 0.891 0.034 0.904 0.848 0.740 0.113 0.765
U2Net [52] 44.02 58.8 45 0.859 0.770 0.074 0.845 0.935 0.896 0.031 0.916 0.861 0.769 0.106 0.789
UCF [53] 29.47 61.4 12 0.757 0.726 0.116 0.806 0.888 0.823 0.062 0.875 0.805 0.737 0.148 0.763
Amulet [25] 33.15 45.3 10 0.806 0.757 0.100 0.818 0.897 0.841 0.051 0.886 0.795 0.741 0.144 0.755
DSS [54] 62.23 114.6 7 0.831 0.740 0.101 0.820 0.916 0.844 0.041 0.881 0.846 0.747 0.122 0.746
PiCANet [13] 32.85 19.7 5 0.880 0.792 0.076 0.854 0.921 0.870 0.043 0.904 0.855 0.785 0.103 0.793
BASNet [55] 87.06 127.3 36 0.854 0.771 0.076 0.838 0.928 0.896 0.032 0.909 0.849 0.744 0.112 0.772
PoolNet [56] 53.63 123.4 39 0.855 0.826 0.065 0.867 0.925 0.903 0.037 0.908 0.863 0.758 0.111 0.781
MINet [14] 162.38 87.1 43 0.882 0.843 0.065 0.855 0.932 0.906 0.030 0.914 - - - -
VST [57] 44.48 23.2 40 0.850 0.829 0.061 0.873 0.937 0.900 0.029 0.928 0.866 0.833 0.065 0.854
PFSNet [58] 31.18 37.6 44 0.881 0.837 0.063 0.876 0.943 0.919 0.026 0.933 - - - -
ICON [59] 33.09 20.9 57 0.876 0.833 0.064 0.861 0.940 0.910 0.029 0.920 0.879 0.804 0.084 0.824
MENet [60] - - 45 0.890 0.870 0.054 0.872 0.948 0.932 0.023 0.927 0.878 0.868 0.087 0.809
TSERNet [61] 189.64 203.6 35 0.857 0.782 0.062 0.840 0.930 0.904 0.036 0.910 0.850 0.746 0.109 0.775
A3Net [62] 17.00 34.1 46 0.844 0.791 0.089 0.831 0.920 0.881 0.042 0.903 0.843 0.787 0.120 0.765
Avg-heavy 62.00 72.6 34 0.853 0.799 0.076 0.847 0.926 0.888 0.036 0.908 0.849 0.774 0.110 0.784
Lightweight method (#Param <= 10 M)
HVPNet [16] 1.24 1.1 55 0.826 0.784 0.089 0.830 0.915 0.872 0.045 0.899 0.826 0.779 0.122 0.765
SAMNet [17] 1.33 0.5 37 0.812 0.778 0.092 0.826 0.915 0.871 0.045 0.898 0.833 0.780 0.124 0.762
CSNet [15] 0.14 1.5 48 0.835 0.723 0.103 - 0.899 0.840 0.059 - 0.825 0.724 0.137 -
Ours 2.29 1.5 62 0.847 0.801 0.084 0.833 0.919 0.889 0.044 0.901 0.845 0.796 0.117 0.767
The larger the mF and S, the better, the smaller the MAE, the better, and Avg-heavy represents the average of each metric of
all heavyweight methods. The best lightweight methods are in bold, and the underline indicates the metrics where SANet is better than
Avg-heavy. FLOPs: Floating-point operations; FPS: frames per second; MAE: mean absolute error.

4.2. Performance analysis
Inthissection, wecompareSANetwitheighteentypicalSODmethods, includingfifteenheavyweightmethods
and three lightweight state-of-the-art methods. This paper uses the same method to evaluate the detection
results of related models.

4.2.1 Comparison with heavyweight SOD methods in general scenarios
Table 2 shows the evaluation results of SANet and existing state-of-the-art SOD methods in terms of #Param,
FLOPs, FPS, maxF, avgF, MAE, and S. From Table 2, we can see that SANet can achieve the performance of
general heavyweight methods in the four evaluation metrics of maxF, avgF, MAE, and S. Especially in the
challenging DUT-OMRON dataset, the four performance metrics all exceed the average level of heavyweight
methods, with maxF, avgF, and S increase by 0.88%, 0.95%, and 0.97%, respectively, and MAE reduced by
4.92%. In terms of efficiency metrics, SANet has reduced parameters by 96.31%, reduced FLOPs by 97.93%,
and increased FPS by 82.35% compared to the average level of heavyweight methods.