Study factors of unusual patterns found from results of DDRNet23-slim trained on DUTS-TR about ddrnet HOT 2 CLOSED

• Failed to solve it with lower down-sampling rates
• We decreased down-sampling rates to 2, 4, 8, 16 and 32
• Refer to this commit for more details
• It achieves lower MAE on DUTS-TE (0.0463 to 0.0452) at 48 epochs

^{Figure 03. Inference results of DDRNet23-slim with lower down-sampliong rates for Figure 1<\sup>}

• However, it still creates diamond patterns
• I.e. We cannot effectively solve it with lower down-sampling rates
• Then, what about the values of down-sampling rate in PoolNet and CSFNet?

^{Figure 04. Inference results of PoolNet-ResNet50-wo-edge with lower down-sampliong rates for Figure 1<\sup>}

• PoolNet-ResNet50-wo-edge down-samples the spatial size of input image by factor of 2 in the output layer (ScoreLayer)

^{Figure 05. Inference results of CSFNet-Res2Net50 with lower down-sampliong rates for Figure 1<\sup>}

• CSFNet-Res2Net50 down-samples the spatial size of input image by factor of 4 in the output layer (self.cls_layer)
• It seems CSFNet-Res2Net50 also suffers from the problem
• lowering down-sampling rates may improve some of it
• However, it requires tremendous FLOPS and time complexity

from ddrnet.

• Refer to Pixelation for more details

from ddrnet.