SpaceNet5 and xView2 competition code

Code used when competing in the SpaceNet5 and Xview2 challenges; chronological order was SpaceNet5 first, and that code was modified for Xview2 afterward. See individual commits if desired, as the resulting code in the master branch is Xview2-specific and there are no releases (code was meant for internal use only).

i.e. this repo houses code that generates solutions for two specific semantic segmentation and object classification tasks.

Xview2 challenge summary

Xview2, sponsored by the Defense Innovation Unit, sought an automated method of identifying damaged buildings after natural disasters using only satellite imagery.

The given samples in this case were a pair of pre-disaster and post-disaster satellite images, and the required output was a pair of grayscale masks classifying each pixel in the input images that represented a building, as well as the level of damage sustained by said building (if any).

Xview2 Challenge site

SpaceNet5 challenge summary

SpaceNet5, sponsored by Cosmiq Works et al., sought an automated method of identifying damaged road networks so that first responders could reach necessary areas efficiently after natural disasters, also using only satellite imagery.

The given samples for this one were single satellite images of road networks. The required output was a set of coordinates representing a graph of the usable roads in the input image, as well as the speed at which those roads could be safely traveled.

Cosmiq Works

Project overview

The model takes RGB input images (satellite imagery) and outputs a one hot encoded (BATCH_SIZE, HEIGHT * WIDTH, NUM_CLASSES) array so that standard categorical crossentropy could be used for training.

The project wasn't written to serve as a basis for derivative works, but if you find it useful, it's released under a BSD license and you're free to incorporate it or build on it as you please.

If you want to use this project on your own dataset(s), the easiest way would be to make a Dataflow class in a file named flow.py, and implement __len__() and __getitem__(). Then update the shape of your samples in settings.py, and if you want to use train.py unmodified, you'll also need to implement flow.get_training_files() and flow.get_validation_files() that return a list of paths for the files representing your dataset.

You can also subclass flow.Dataflow(), but you'll need to change some calls from flow.Dataflow() to myFile.betterClass() or whatever you call your dataset objects in the existing Python files.

File descriptions

The models

• unet.py: Version of Motokimura's winning SpaceNet model (Unet style fully convolutional network)
• deeplabmodel.py: DeepLabV3+ model in Keras by Bonlime

Using the models

• settings.py: Global settings for the project (neural net input shape, mask shape, and so forth)
• train.py: Code for building and training a semantic segmentation model
• infer.py: A few functions for converting neural net output => displayable images
• score.py: Metrics for training and evaluating
• test.py: Generate solution files (PNG image data, 1024 x 1024, 8-bit grayscale, non-interlaced)
• scansolution.py: Counts pixel values in solution files (sanity checks)

The dataset (see the xBD paper for details)

• flow.py: All the code for the dataset (data objects, a few image deformations, etc).
• show.py: Code for displaying input samples and model predictions
• damage.py: some miscellaneous code related to damaged buildings
• mkbuildings.py: A little script to dump individual buildings from input samples to their own files

---

Development environment (the only platform it was really tested on):

• Linux v5.x
• python v3.7.4
• tensorflow v2.0
• all required packages installed via pip3 install