Transfer learning for salient object detection

[swoook]

Is your feature request related to a problem? Please describe.

• We'd like to employ the DDRNet23-slim pre-trained on the Cityscapes for salient object detection
• How should we perform transfer-learning for salient object detection?

Describe the solution you'd like

• Pick a strategy by referring to Transfer learning from pre-trained models (towardsdatascience) (korean)

Describe alternatives you've considered

• None

Additional context

• DUTS-TR is a dataset for salient object detection
• We'd like to train DDRNet on this dataset for real-time salient object detection
• Refer to #6 from swoook/ucnet (github) for more details

[swoook]

• Recall that DDRNet is trained and benchmarked on Cityscapes
• It has 2975 images for training
• On the other hand, DUTS-TR has almost 10k images
• However, those two datasets are quite different
• I.e. Our case is Quadrant 1 (large dataset size, low dataset similarity)
• Thus, we'd like to choose strategy 1: Fine-tuning

[swoook]

• Failed to train DDRNet23-slim on DUTS-TR
• Specifically, we confirmed two main problems

1. The model over-fits to the DUTS-TR
2. Its validation accuracy oscillates a lot (even 0.10 after 20 epochs!)

• Then, do they also occur in PoolNet or CSFNet?
• It seems batch-normalization is a critical factor in these problems
• If we turn the parameters in batch-normalization on, they also occur in both of them
• After freezing them, we achieved even lower MAE than ones stated in the papers
• Refer to #2 from swoook/poolnet (github) for more details
• Then, let's apply this on DDRNet23-slim too

[swoook]

• Confirmed slightly better results, but not enough
• It turns out we need one more modification
• It's better to employ training settings from the backseason/PoolNet
• Achieved 0.047 MAE on DUTS-TE at 24 epochs

[swoook]

• Refer to this commit for more details