michal-nahlik/kaggle-clouds-2019

Kaggle - Understanding Clouds from Satellite Images - multiclass image segmentation

Understanding Clouds from Satellite Images

Multiclass segmentation of cloud patterns in satellite images.

Code for Kaggle's Understanding Clouds from Satellite Images Challenge. This solution has scored 0.65455 (Dice) on private leaderboard and ranked 161st place (top 11%).

Description

• PyTorch

• Ensemble of 4 pretrained models from Segmentation Models (Unet, FPN) and FastFCN (EncNet, DeepLabV3)

• Optimizer: RAdam (lr=0.005)

• Metric: Dice

• Loss: Focal + log dice

• Scheduler: ReduceLROnPlateau

• Preprocessing: Resize from 1400x2100 to 320x480

• Postprocessing: Threshold, Remove small masks, Draw convex hull mask

• Albumentations augmentations

• Trained on 80% of data (stratified based on mask areas)

• 35 epochs, models with best Dice score were selected for ensemble

Results

Local dice: 0.64465
Public leaderboard score: 0.65574
Private leaderboard score: 0.65455

Model raw output dice

Model	Backbone	Local Dice
FPN	efficientnet-b4	0.46
Unet	efficientnet-b4	0.444
EncNet	resnet50	0.445
DeepLabV3	resnet50	0.436

Class dice score after ensemble and postprocessing

Class	Fish	Flower	Gravel	Sugar
Local Dice	0.578186	0.750014	0.642398	0.608013

Example of mask and model output