December 6, 2016

Population Density Estimation with Deconvolutional Neural Networks

Workshop on Large Scale Computer Vision at NIPS 2016

By: Amy Zhang, Xianming Liu, Tobias Tiecke, Andreas Gros

Abstract

This work is part of the Internet.org initiative to provide connectivity all over the world. Population density data is helpful in driving a variety of technology decisions, but currently, a microscopic dataset of population doesn’t exist. Current state of the art population density datasets are at ~1000km2 resolution. To create a better dataset, we have obtained 1PB of satellite imagery at 50cm/pixel resolution to feed through our building classification pipeline.

Our contribution is a fast building classification pipeline, that can run through a country in 8 hours on Facebook infrastructure. It is composed of an edge detection method for fast bounding box proposals, and a weakly supervised deconvolutional neural network that is trained for pixel-level classification, then mean pooled over the bounding box to output a probability of a building or buildings present in the bounding box. We train a global model and obtain precision and recall of > 90% in most countries. Countries with poorer results we use active learning techniques to re-sample data and fine-tune a new model. We also developed a weakly supervised footprint segmentation model that processes larger images with more context and produces a mask of location and shape of each building and a denoising network to clean up poorer quality source data.