Please note that this repository is longer functional and only exists for archival purposes. Since the release of this repository, several other approaches (e.g. https://arxiv.org/abs/1609.03677) have produced superior results; therefore, I recommend that you explore these methods instead. Regardless, the model.py provides a "barebones" implementation without weights or display tools.

DepthNet

DepthNet is an unofficial Tensorflow implementation of Predicting Depth, Surface Normals and Semantic Labels with a Common Multi-Scale Convolutional Architecture. Note: This repository was created for a research project, not associated with NYU, to explore the implications of residual neural networks for monocular depth estimation and smartphone-based spatial mapping. These modifications were not included in this repository for compeleteness. If you would like these modification, please email me at rcbridendev@gmail.com.

Installation

There are two ways to install NYUDepthNet - Automatic and Manual. The latter is complex to configure, so it's recommended that you use the Automatic method.

Automatic

This is the recommended way to install NYUDepthNet.

• Clone the repository.
• Install dependencies
  • Tensorflow
  • Numpy
  • Matplotlib
• Run main.py to ensure NYUDepthNet was installed correctly.

Manual

This installation method is more complex; however, it does grant increased customizability.

• Clone the repository.
• Install dependencies
  • Theano
  • Install the dependencies mentioned in the Automatic method.
• There are two methods to setup the installation:
  1. Run setup_env.py
  2. Manually download an unpack weights
     • Download weights and scripts from NYU
     • Convert weights from .pk format to .npy or to tensorflow variables. NOTE: These weights are formatted as Theano tensors, so they must be converted to Tensorflow tensors. See TheanoUnpickler.
• Run main.py to ensure NYUDepthNet was installed correctly.

Example Images