Segmentation Dataset

This repository contains a description of the DroneDeploy Segmentation Dataset and how to use it. It also contains example code to get a working segmentation model up and running quickly using a small sample dataset. See below for details of the full dataset and suggested improvement directions.

Quickstart

Follow these steps to train a model and run inference end-to-end:

git clone https://github.com/dronedeploy/dd-ml-segmentation-benchmark.git
cd dd-ml-segmentation-benchmark
pip3 install -r requirements.txt

# optional: log in to W&B to track your experiements
wandb login

# train a Keras model
python3 main_keras.py

# train a Fastai model
python3 main_fastai.py

This will download the sample dataset and begin training a model. You can monitor training performance on Weights & Biases. Once training is complete, inference will be performed on all test scenes and a number of prediction images with names like 123123_ABCABC-prediction.png will be created in the wandb directory. After the images are created they will be scored, and those scores stored in the predictions directory. Here's what a prediction looks like - not bad for 50 lines of code, but there is a lot of room for improvement:

Dataset Details

The dataset comprises a number of aerial scenes captured from drones. Each scene has a ground resolution of 10 cm per pixel. For each scene there is a corresponding "image", "elevation" and "label". These are located in the images, elevation and labels directories.

The images are RGB TIFFs, the elevations are single channel floating point TIFFs (where each pixel value represents elevation in meters), and finally the labels are PNGs with 7 colors representing the 7 classes (documented below).

In addition please see index.csv - inside the downloaded dataset folder - for a description of the quality of each labelled image and the distribution of the labels.

To use a dataset for training, it must first be converted to chips (see images2chips.py). This will create two directories, images-chips and label-chips, which will contain a number of 300x300 (by default) RGB images. The label-chips are also RGB but will be very low pixel intensities [0 .. 7] so will appear black as first glance. You can use the color2class and category2mask function to switch between the two label representations.

Here is an example of one of the labelled scenes:

Each color represents a different class.

Color (Blue, Green, Red) to Class Name:

(075, 025, 230) : BUILDING
(180, 030, 145) : CLUTTER
(075, 180, 060) : VEGETATION
(048, 130, 245) : WATER
(255, 255, 255) : GROUND
(200, 130, 000) : CAR
(255, 000, 255) : IGNORE

• IGNORE - These magenta pixels mask areas of missing labels or image boundaries. They can be ignored.

Possible Improvements

---

The sample implementation is very basic and there is immediate opportunity to experiment with:

• Data augmentation (datasets_keras.py, datasets_fastai.py)
• Hyperparameters (train_keras.py, train_fastai.py)
• Post-processing (inference_keras.py, inference_fastai.py)
• Chip size (images2chips.py)
• Model architecture (train_keras.py, train_fastai.py)
• Elevation tiles are not currently used at all (images2chips.py)