Future work for transfer learning on salient object detection

[swoook]

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

• We trained DDRNet23-slim on DUTS-TR and it achieved 0.47 MAE on DUTS-TE
• And we'd like to improve the accuracy of DDRNet23-slim on salient object detection

Describe the solution you'd like

• Research BoFs and BoSs for related tasks, especially segmentation

Describe alternatives you've considered

• None

Additional context

• None

[swoook]

1. Enable an intermediate supervision (auxiliary loss)
2. Modify the architecture so that DAPPM has similar effect for images with size of (512, 512)
   • Refer to #7 for more details
3. Many other losses verified on famous benchmark datasets for semantic segmentation tasks
   • Refer to JunMa11/SegLoss (github) for more details
4. Study the trade-off between input resolution and number of channels

[swoook]

• Confirmed 1 (auxiliary loss) doesn't improve MAE (0.048 at 24 epochs)