Kaggle Amazon
contest
Dmitry Pranchuk
Labels Distribution
Co-occurence Matrixes
Weather Labels Most Common Labels
Rare Labels
Target Metric
Dataset
JPG (RGB) and TIFF (GeoTIFF = RGB + near infrared)

~40k train set

~60k test set (40k + 20k)
First Model
Keras

3 conv + 2 fc layers

Adam (lr = 1e-3, 1e-4, 1e-5)

sigmoid

binary loss

0.909 on LB
Augmentation
original 256x256
random crop 224x224
random rotate by 0, 90, 180 or 270 degree
random horizontal or/and vertical flip
Thresholds Optimization
Finetuning
Ensemble
1) 5-layers CNN (0.925)

2) VGG-16 (0.927)

3) VGG-19 (0.928)

4) ResNet-50 x 3 (0.925-0.927)

5) ResNet-50 (sigmoid for non-weather + softmax for weather) (0.923)

6) ResNet-50 (weather) + ResNet-50 (non-weather) (0.925)

7) LightGBM on the ResNet-50 features (0.920)

4-threshold voting

0.93042 on Public, 0.92915 on Private
Public/Private Correlation
Leaderboard Shake-up
Panorama
Haze Removal Using Dark Channel Prior
He, K., Sun, J., Tang, X.: Single Image Haze Removal Using Dark Channel Prior

IEEE Trans. Pattern Anal. Mach. Intell. (2011)
Lessons Learned
1) Manually check your submits

2) Save experiments results

3) Try not to start too late

4) Optimize your code

Kaggle Amazon Contest