RSIPAC2022
RSIPAC 2022 @Track3
Track 3: Semantic Segmentation
TODO: upload the tuned class weights for CE loss and the DICE loss after the final phase.
Update 2022/11/20
- We uploaded the final version of our model configuration. We used a cas-training approach: the model is initially trained @512x512 and then tuned @768x768.
Update 2022/11/19
- The score is the weighted average F1-Score over all classes, thus the small classes will contribute little to the final score. So we can drop 3-4 classes for simplification. In our solution, we dropped the photovolts, the airports and the railway-station. As the percentage of the natural-bare-soil is also very low, we believe this could also be dropped.
- To solve the class imbalance problem, we tuned the class weight manually using the confusion matrix.
- Multi-scales training and inferencing could also improve the performance.
Update 2022/11/17
Main features of Basic Model:
- CE + DICE loss (1:2)
- Multiple augmentations with Albu to avoid overfitting
- TTA
- SWA
- Drop some small samples, such as the photovolts, the airports and the railway-station. We only classify the objects into 15 classes with the class 0 to be the playground.