DenseDepth-Pytorch

A simple PyTorch Implementation of the "High Quality Monocular Depth Estimation via Transfer Learning" paper.

The paper can be read here.

Official Implementation can be found here.

Training

• The model was trained on the NYU Depth V2 dataset.

• There is no need to manually download the dataset. The file download_data.py handles this. To use this, simply run the following command from the project root,

$ python densedepth/download_data.py

• The model was trained for 7 epochs (~ 7/8 hours). In contrast, the authors train for 1 M epochs for 20 Hours.

• Since the model was trained on Google Colab, it was trained periodically when Nvidia T4 or P100s were available. Training on a single 12 GB Tesla K80 takes too long. In contrast, the authors use a cluster of 4 Tesla K80s.

• Train Loss at the end of the 7th epoch was ~0.17.

Quick Start

To immediately start training without any set-up, go to the notebook DenseDepth_training.ipynb (use this link if it doesn't work) and open it in Google's Colaboratory, and simply run all the cells!

The notebook by default, asks you to link your Google Drive so as to save the checkpoints, but you can skip this step and instead save it in runtime memory.

Usage

All the required files are in the densedepth directory.

• Step 1: Clone the repository

git clone https://github.com/pranjaldatta/DenseDepth-Pytorch.git

• Step 4: Go into Project root by,

$ cd DenseDepth-Pytorch

• Step 3: To train,

python desedepth/train.py --epochs 10 \
                          --batch 4 \
                          --save "directory where checkpoints are to saved" \
                          --device "cuda" \
                          --checkpoint "path to checkpoint from which to resume training"

• Step 4: To test,

python densedepth/test.py --checkpoint "<path to checkpoint>" \
                          --device "cuda" \
                          --data "examples/"

For help,

python densedepth/train.py --help

Results

Image	Prediction

For more examples, check here.