asumgh/avia-missing-airplanes

Missed Airplanes (https://cups.mail.ru/)

Table of contents

• Install packages and dependencies
• Project structure
• Task
• Data description
• Augmentation
• Learning
• Result

1. Install packages and dependencies

  pip install -r requirements.txt
  pip install -e .

2. Project structure

.
├── 7 folds current model                                                                               <-- First folder with models
│   ├── You should install models
│       into this catalog from 
│       https://drive.google.com/drive/folders/1REAS6CUAllJ7HyDDP3Xf9sprExrJUkhI?usp=sharing                <-- Link for download models
├── mean acc and f1                                                                                     <-- Second folder with models
│   ├── You should install models
│   │    into this catalog from
│   │    https://drive.google.com/drive/folders/1zvM4BXmOvuhwQM6dEcIieLdcZ4UnOx-n?usp=sharing               <-- Link for download models
│   ├── src                                                                                                 <-- Main functions
│   │   ├── __init__.py                                                                                         <-- Initialization
│   │   ├── augment.py                                                                                          <-- Audmentation functions
│   │   ├── dataset.py                                                                                          <-- Dataset functions
│   │   ├── global_var.py                                                                                       <-- Global variables
│   │   └── modeling.py                                                                                         <-- Train loop
│   └── Avia_base.ipynb                                                                                     <-- Training file
├── model full                                                                                          <-- Third folder with models
│   ├── You should install models
│   │   into this catalog from
│   │   https://drive.google.com/drive/folders/1nShNy0YlmQNzUc9XjBCmipqAPkW_Wtma?usp=sharing                <-- Link for download models
│   ├── src                                                                                                 <-- Main functions
│   │   ├── __init__.py                                                                                         <-- Initialization
│   │   ├── augment.py                                                                                          <-- Audmentation functions
│   │   ├── dataset.py                                                                                          <-- Dataset functions
│   │   ├── global_var.py                                                                                       <-- Global variables
│   │   └── modeling.py                                                                                         <-- Train loop
│   └── Avia_base.ipynb                                                                                     <-- Training file
├── Avia_base.ipynb                                                                                     <-- Training file with making prediction
├── README.md
├── requirements.txt                                                                                    <-- Description of dependencies
└── setup.py                                                                                            <-- Building python-packages file

3. Task

In this competition we are searching missed airplanes in satellite images. It is classification task. Metric: ROC-AUC score, 1(main test, 1000 images) + 0.001(addited data, 100000 images)

4. Data description

Training:

• 31080 images
• 7899 airplanes on it
• 40 are generated Test:
• 101000 images

5. Augmentation

I used albumentation with these components:

• One of: HorizontalFlip, VerticalFlip, RandomRotate90
• One of: HueSaturationValue, RandomGamma, RandomBrightnessContrast
• ShiftScaleRotate

6. Learning

It is blending three different-parameters models and each learn on 7 stratified folds. Each model is ResNet18(https://arxiv.org/abs/1512.03385)

7. Result

Public score = 1.0007232982844492 Privat score = 1.0009728024369016