Semi-Parallel Deep Neural Network (SPDNN) Hybrid Architecture, First Application on Depth from Monocular Camera

This repository contains the code for the method presented in the following paper:

Bazrafkan, S., Javidnia, H., Lemley, J. and Corcoran, P., 2017. "Depth from Monocular Images using a Semi-Parallel Deep Neural Network (SPDNN) Hybrid Architecture". arXiv preprint arXiv:1703.03867

As described in the Training section of the paper, four experiments are designed in this project:

Exp1: Input: Left Visible Image + Pixel-wise Segmented Image. Target: Post-Processed Depth map.

Exp2: Input: Left Visible Image. Target: Post-Processed Depth map.

Exp3: Input: Left Visible Image + Pixel-wise Segmented Image. Target: Depth map.

Exp4: Input: Left Visible Image. Target: Depth map.

To prepare the input for training:

1. Install the Caffe SegNet
2. Train the SegNet using CamVid road scene database
3. Use the trained model to segment the images of the KITTI 2012, 2015 dataset.

To prepare the target for training:

1. Estimate the depth from the KITTI stereo sets using Adaptive Random Walk with Restart algorithm
2. Post-process the initial depth maps using our post-processing method
   You can duplicate the experiments described in the paper using the codes in this repository and the prepared data.

Please cite the following papers when you are using this code:

Bazrafkan, S., Javidnia, H., Lemley, J. and Corcoran, P., 2017. "Depth from Monocular Images using a Semi-Parallel Deep Neural Network (SPDNN) Hybrid Architecture". arXiv preprint arXiv:1703.03867