Hot Topics in Computer Vision

Addressing Label Scarcity in the Semantic Analysis of Earth Observation Imagery

We implement the data augmentation method InAugment by Arar et al.

What does this project contain

• Preprocessing scripts
• InAugment implementation
• Notebooks for classifiers

Requirements

Tested with:

PYTHON 3.10

Installation

pip install -r requirements.txt

Usage

1. Download xview2 train, test and holdout data.

2. Create a data directory in the root directory and put all three into the data directory.

3. Put the xview2.txt inside the data folder

Make an ordered dataset

This takes the raw images from the xview2 data and orders them into the required directory structure. Type python3 src/data_preprocessor.py -h for help. This also equals to the Baseline dataset

python src/data_preprocessor.py -src ./data -out ./dataset  

Perform InAugment

mode: random

Generates a specified amount of augmented images from a given original data pool.

• type python3 src/in_augment.py -h for help
• to change InAugment hyperparameters change CONSTANTS in inaugment.py
• source folder should be the unaltered unorganized data
• output folder defines the augmented image directory
• set datasize to the size of the original data pool i.e. 280, 700, 1400, 2800
• set number to the amount of images to sample, for further steps a certain amount of images need to be generated

python src/in_augment.py -src "./data/train/" -out "./augmented_datasets/DATASET_NAME/" -ds 280 -num 1000

mode: single-instance/linear

This generates the dataset for the single-instance/linear model

• type python3 src/in_augment.py -h for help
• to change InAugment hyperparameters change CONSTANTS in inaugment.py
• source folder should be the unaltered unorganized data
• output folder defines the augmented image directory
• set datasize to the size of the original data pool i.e. 280, 700, 1400, 2800
• performs InAugment once per image, resulting augmented dataset is as large as datasize

python src/in_augment.py -src "./data/train/" -out "./augmented_datasets/DATASET_NAME/" -ds 280

Mix datasets

Combine regular and augmented data to create half-half, majority original, majority augmented dataset

• go into augmented_data_preprocessor.py and change constants as wanted. Consider that both source datasets need to have enough images
• original data path points to the regular ordered images
• augmented data path points to the augmented images
• new mixed path is the output path of the mixed dataset
• original data size refers to the size of the original data pool, i.e. 280, 700, etc.
• augmented data to original data ratio describes how many augmented images are being placed into the dataset in relation to the data size. Ratio = 2.0 => twice amount of augmented images than regular/original

python src/mixed_dataset_preprocessor.py  

Balance datasets

Take images from the augmented images and build the balanced dataset. Source data needs to have enough images to fill each class.

• go into balanced_data_preprocessor.py and change constants as wanted. Consider that both source datasets need to have enough images
• augmented data path are the source images we use to balance the dataset with
• new balanced path is the output path of the balanced dataset
• label instances refers to the amount of images per class

python src/balanced_dataset_preprocessor.py

Average Building Damage Classification

Under notebooks/baseline.ipynb is a jupyter notebook for the baseline model.

Under notebooks/inAugment.ipynb is a jupyter notebook that trains on any given model.

The datasets for the models need to be in the structure:

dataset_dir:

.
├── 0
├── 1
├── 2
├── 3
└── 4

Where the numbers represent different JDS labels:

• 1: No Damage
• 2: Minor Damage
• 3: Major Damage
• 4: Destroyed
• 0: No Buildings (Misc/Filler label)