Final Writeup

https://docs.google.com/document/d/1EP_nxi3XBUT7qdDtB7CaWM7yGUBvRZ2EXbXx93PPVCw/edit?usp=sharing

Writeup #2

Introduction

We are implementing a research paper which has the intention to address climate change by influencing people’s behaviors. It uses a GAN model to make landscapes appear flooded in order for people to visualize the impacts of climate change. We will use google street view data in addition to local images so that the Brown community will see areas they are familiar with.

Challenges

So far, the biggest challenge we have encountered so far has been finding ways to break the project down into manageable pieces. We based our project off of a research paper that takes in GPS coordinates and produces a flood image based on Google Maps’ street view, which is too large a task for the amount of time we have (for example, to produce a train set they manually painted each pixel of images to generate floods). Instead of attempting to implement the entire project, our initial plan was to use their code up until a certain point and use the produced images as the input to our model, which would be only a small piece at the end of their project. That way we could test our model’s produced flood images against their produced flood images (finding a way of testing our model was also a challenge we ran into, because aside from accuracy being generally dependent on how realistic we think the flooded images look or manually painting pixels, we were unsure how to test it) so this approach solved that problem.

Insights

We have generated our labels for our data set using the model in the paper. Additionally, we have a concrete plan of our model structure.

Plan

We are roughly on track with the project, and we plan on dedicating more time to experimenting with different metrics, loss functions, and model architectures.

Writeup #1

Who

Dan Fiume, Katie Orchard, Tyler Zickmund

Introduction

We are implementing a research paper which has the intention to address climate change by influencing people’s behaviors. It uses a GAN model to make landscapes appear flooded in order for people to visualize the impacts of climate change. We will use the data set Cityscapes but also hope to pass in local images so that the Brown community will see areas they are familiar with.

link

Related Work

Climate visuals: A mixed methods investigation of public perceptions of climate images in three countries This research paper investigates the role that visual communication plays in climate change response. Four discussion groups in three different countries were conducted to identify responses to different imagery. They found that images of solutions produced positive responses and less polarization than familiar climate images, which were met with cynicism. They also found that images of ‘impacts’ produced greater intentions to change one’s behavior. This paper illustrates the importance of the project we are reimplementing in contributing to climate solutions. link

Data

We are using a semi-standard data set, Cityscapes, link, particularly the set of classes dealing with street view images of objects and buildings. This dataset is large and will require some preprocessing to access the features we are interested in. We also plan to give the model some new images of streets in Providence and the Brown area to test its generality.

Methodology

We plan on using a GAN to take in an image and generate a new image that makes the original image appear flooded. The model will consist of two major parts: a mask layer/model that will learn how to take an image and identify which pixels/areas should appear to be flooded, and a painter layer/model that learns how to make those pixels/areas appear flooded. We believe the most difficult aspect will be implementing the metrics to get the models to train well, as there is not a lot of test data of flooded images that to learn.

Metrics

“Accuracy” in the traditional sense is not as absolute in our case. While our model will have accuracy which will be reflective of how well our model learns the mask of where to generate flood water and what it should look like, the real test comes from how realistic the image looks. We plan on developing some sort of human survey to either compare our results to other models or gauge how realistic the photos look. This is similar to what the authors of the paper we are implementing did.

For our goals, our base goal is to get the model working with the given data set. Our target goal is to incorporate a new data set of images of streets and landscapes around Brown and Providence. Our stretch goal is to incorporate some of the other capabilities of the paper, such as using the same architecture to generate images with more smog or with residue of a wildfire.

Ethics

What broader societal issues are relevant to your chosen problem space? Climate change is the biggest social issue relevant to our problem space. One research paper highlights the importance of showing people climate impacts in order to change their behavior, which is the goal of this project. Other aspect: issues about the goal of this project On a broader scale, we acknowledge that many communities that are most impacted by climate emergencies are already aware and scared of these impacts, and that these communities are often the ones with the least power to prevent it. Similarly, we acknowledge that changing individual behaviors will not be able to prevent climate impacts quickly enough and that there must be changes made at a policy level; we hope that these images will bring climate issues closer to mind for people who are not actively impacted and that they will put pressure on political actors to pass climate legislation.

Division of Labor

One group member will focus more on creating the model to learn the mask of the image, while the other group member will focus more on the model to paint the image.

Built With

• tensorflow