This is a submission to the Iris Challenge. Taken from the challenge instructions:

There are 31 images in the dataset. They are all taken from drones in a range of different environments.Provide us with as much contextual awareness about these environments as you possibly can. We want to automatically understand everything about the scene in the way a human pilot might.

This repo contains two parts:

• Object detection model for clouds, sun, houses, and trees
• Horizon line detector

Detector

The detector (detector/*) uses the TF Object Detection API. The procedure is as follows:

• Find images on the internet, use LabelImg to label them (I used a tiny dataset of 20 images)
• Generate TFRecords files for training using detector/generate_dataset.ipynb
• Train a model from a pre-trained coco OD model using detector/train_model.ipynb
• Run the saved model on the the challenge test_images using detector/run_model.ipynb

Horizon Angle

The horizon angle regression model (horizon/*) uses OpenCV built-ins. The procedure is as follows:

• Preprocess image (gray, blurring)
• Use Canny Edge Detection on image, dilate the resulting edges
• Use Hough Line Transform to get lines, fine tune parameters as needed
• Average the resulting Hough Lines to get the horizon

Below are some hand-picked results. (Find more in output_images/*)

• Python 3.5
• TensorFlow 0.14
• Jupyter Notebooks
• OpenCV 3.3

Hugo Ponte