Aerial Imagery Pixel-level Segmentation

Codebase for the MSc thesis paper Aerial Imagery Pixel-level Segmentation

Creating environments for the DroneDeploy fastai/keras benchmark and DeepLabv3+ codebase

It is crucial to take note of your own GPU driver environment. For this work the following environment was available (High Performance Cluster at Eindhoven University of Technology):

• GCC & G++: 5.4.0 (20160609)
• Nvidia driver: 450.51.06
• CUDA driver: 11.0
• CUDA compilation tools (including nvcc): release 10.2, V10.2.89
• Using TensorFlow through Conda automatically installs the latest CuDNN version in your local environment.

fastai environment and preparations

This is for use with the dd-ml-benchmark implementation

1. conda create --name fastai_gpu pytorch torchvision cudatoolkit=10.1 -c pytorch
2. conda activate fastai_gpu
3. conda install -c fastai fastai=1.0.61
4. conda install opencv typing wandb scikit-learn
5. pip install image-classifiers

keras environment and preparations

This is for use with the dd-ml-benchmark implementation

1. conda create --name keras_gpu keras tensorflow-gpu=1.15
2. conda activate keras_gpu
3. conda install opencv typing wandb scikit-learn
4. pip install image-classifiers

DeepLabv3+ environment and preparations

This is for use with the models/research/deeplab Tensorflow implementation

1. conda create --name tf1_gpu tensorflow-gpu=1.15
2. conda activate tf1_gpu
3. conda install -c conda-forge pillow tqdm numpy
4. pip install tf_slim
5. From tensorflow/models/research/ directory run: export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim

Original data

Can be found here: https://github.com/dronedeploy/dd-ml-segmentation-benchmark/blob/master/libs/datasets.py

List of altered files for DroneDeploy dataset experiments

DeepLabv3+ Tensorflow research codebase https://github.com/tensorflow/models/tree/master/research/deeplab extensions

• convert_rgb_to_index.py (altered to strip 3 dimensional segmentation labels to 1 dimensional)
• build_dd_data.py (altered for DroneDeploy compatiblity)
• data_generator.py (altered for DroneDeploy compatiblity)
• train-dd-full.sh, eval-dd.sh, vis-dd.sh (dataset adaptations inspired by this GitHub repo (https://github.com/heaversm/deeplab-training))

DroneDeploy benchmark codebase https://github.com/dronedeploy/dd-ml-segmentation-benchmark extensions

• custom_training.py (implementation Focal loss function for fastai u-net)
• custom_training_keras.py (implementation Focal loss function for Keras u-net)
• images2chips.py (added test images conversion to tiles for DeepLabv3+ compatibility)
• scoring.py (added mean IOU and IOU score per class metric)