APPROACH.md

Approach

Data Cleaning process

1. Data cleaning
   • Removed inconsistent labels
   • Removed broken masks
   • Removed wrong masks (e.g. inverted masks, wrong masks...)
2. Trained segmentation model on clean data
3. Generated pseudo-labels for sample that got removed in step 1

Models

• Used in final models:
  • 3x encoder-decoder models in ensemble:
    • Encoder: resnet34, decoder: unet
    • Encoder: efficientnet-b1, decoder: unet
    • Encoder: resnext50_32x4d, decoder: unet
• Tried out:
  • Encoder: resnet34, resnet50, resnet101, efficientnet-b4, xception, se_resnet50
  • Decoder: unet, FPN

Loss function

• Used in final models:
  • Cross-Entropy Loss
• Tried out:
  • Binary Crossentropy Loss (worse performance)

Optimizer

• Used in final models:
  • Adam
• Tried out:
  • Rectified Adam (worse performance)

Learning Rate Scheduler:

• Cosine Annealing LR

Metric

• JaccardIndex (IOU)

Augmentations

• Used in final models:
  • Horizontal flip, vertical flip
• Tried out:
  • Copy&paste augmentation
  • ShiftScaleRotate
  • Random crop
  • Multi scale

TTA

• Used in final models:
  • unfortunatly i only could use TTA1 (hlip) without over-extending the 4 hour submission time limit
• Tried out:
  • TTA3 showed best performance (TTA3 = vflip, hflip, vflip+hflip)

Post-Processing

• Used in final models:
  • none
• Tried:
  • Doublet scheme
    • set min_score_threshold (e.g. 0.4) and top_score_threshold (e.g. 0.6)
    1. Determine if image contains clouds (values > top_score_threshold)
    2. Add values around the clouds borders, which are > min_score_threshold
  • Morphological close
    source: OpenCV
  • Morphological dilation
    source: OpenCV
  • Remove areas < 150px
  • Convex hull, approx. polygon shape, approx. rectangle shape
    source: Kaggle notebook

Inference Pipeline

• Inference on each model:
  • Make 2 predictions on each image:
    1. Prediction without aug.
    2. Prediction after aug.
  • Take average of both predictions ((pred_1 + pred_2) / 2)
• Add each resulting mask of each model to each other and take the average of that ((model1_pred + model2_pred + model3_pred) / 3)

Lessons learned and future approaches

• Proper model validation is key for success
• Iterate faster to have more time for post-processing
• Try out convNext and SwinTransformer models
• Implement efficient TTA3 solution
• Find edge cases to use post-processing techniques mentioned above
• Iterate faster to have more time with post-processing & ensembling
• Teaming up