README.md

Depth Estimation for Mono Camera

This is the repository of the project for Deep Learning with TensorFlow 2.0 class.

Team members:

• Ejup Hoxha

• Stanislav Sotnikov

• Santosh Suwal

Final Report

• Final Report

• Presentation (via google drive link)

Reading Materials

• Original U-Net paper describes the basic architecture we use for the depth prediction.

• Monodepth 2 A novel method of monocular depth estimation. Desribes a self-supervised method, see occlusion loss.

• D3V0 Describes depth and pose estimation network.

Deployment Instructions

Requires python3.8, pip3, pipenv to fetch dependencies automatically.

1. Clone repository and install dependencies.

git clone git@github.com:CSCCNY/final-project-monodepth-ccny.git
cd final-project-monodepth-ccny
pipenv install  # Installs dependencies
pipenv shell  # Enter virtual environment

Data Sources

All benchmarks we performed over the course of the project were trained on NYU v2 dataset. Then evaluated on NYU v2, Middlebury, DIML Indoor, and DIML Outdoor. Only the dedicated test portion of NYU v2, DIML Indoor, and DIML Outdoor was used for evaluation. In case with Middlebury, the entire dataset was used (less than 20 image - depth pairs).

• NYU v2 -- Default Indoor Set.
• Middlebury -- Still Objects.
• DIML -- Indoor and Outdoor.