height_estimation

Estimate objects' heights in Street View images

One of the challenges in generating realistic images of floods is to control the level of the flood, in order to match climate predictions. We propose to introduce height information in our model in order to generate and respect the geometry of the scene.

Here we present the two types of approaches that we have tried so far to estimate height in street scene images : "geometrical approaches" which consist in recovering the 3D metric geometry of the scene, and "end-to-end approaches" which consist in predicting height from an image input.

You can refer to this page to understand how height estimation fits in our goal of generating images of the impacts of climate change.

Description

pinhole contains a notebook to show how to recover 3D coordinates from an image with a depth map and camera parameters using the pinhole camera model.

src contains the pipeline for height estimation in an image using geometrical approaches.

• Clone the content of this [repo]((https://github.com/skhadem/3D-BoundingBox) into the 3DBoundingBox folder.
  git clone https://github.com/skhadem/3D-BoundingBox BoundingBox3D

• Download the weights of the 3D bounding box estimator from that same repo. To do so, go into the BoundingBox3D/weightsfolder and run get_weights.sh.

• Set up srcas a package run: python setup.py develop

Have a look at the demo_notebook.ipynb to see how all the pieces come together.

hourglass is a first model for end-to-end height estimtion

simulated_world contains code to generate the height maps dataset from the images and depth maps from our simulator.