stadium_cnn

A sample project working with keras to analyze satellite imagery.

Introduction

As a study project following the read of deep learning with keras book by f.chollet, I've come up with the idea of using transfer learning with semi-automatic annotated datasets. In particular, gathering lat,lng coordinates of landmarks within a city which their position is already known and annotated and then extracting their satellite image representation.

I've think that soccer stadiums could be a good fit for this idea because stadiums possibly stand-out from the city-scape in terms of size, color and shape.

I know finding stadiums automatically does not add any real-world value but the purpose of this project is to gain insights and xp on extracting value from satellite images and leveraging ML to do so.

Project

Built using a conda env + keras packages. The project is divided into 3 notebooks:

• extract_stadiums.ipynb ( Feature Extraction ) We take satellite images and generate a dataset for training.
• train_explained.ipynb ( Training ) We use the dataset to train a RNN capable of classifying satellite images.
• discover_explained.ipynb ( Discover ) We use the trained RNN to automatically find new samples in unseen data.

sample extracted stadium image:

Results

• DEC 2019: the trained model was able to detect some stadiums in unseen data but it also produced a larger set of false positives.

some succesful detections

some false positives

Next steps

• Augment dataset( very few positive samples)
  • Extract more stadium images from other cities.
  • Create images with partial sections of stadiums instead of a full centered image.
  • scale, rotate, warp positive samples.
• Reduce noise in training set:
  • Positive samples contain images of both the stadium and its surroundings (e.g assuming the parking outside the stadium its a stadium itself). Explore the idea of leaving only the stadium section in the training data and assess effects. (Can we train on partial images and then evaluate on mixed images? how would this affect the performance?)
• Better model diagnostics
  • See data points where model makes mistakes.
• RNN architecture:
  • Incude resnet50 as part of the model architecture and look into training a few top layers ( would require moving into google colab)