KITTI Dense Depth

This is a simple implementation to in-painting sparse depth map of KITTI Dataset for monocular depth estimation task training or result visulization.

Requirements

These code has been tested with Python 2.7 and Matlab-linux on Ubuntu 16.04 LTS.

Dataset preparation

For in-painting depth maps for training task or result visulization, you should download the KITTI raw dataset with the offical split by running the bash script in the specific data directory:

sudo bash download_kitti_raw_data.sh

You would want to follow the traditional split of Eigen, the training list file eigen_train_files.txt and testing list file eigen_test_files.txt can be found in ./utils.

In-painting

• Firstly, you can generate the projected depth ground-truth by running the python code provided by:

https://github.com/danxuhk/StructuredAttentionDepthEstimation

Do not change the file structure of the kitti raw data. Modify the kitti raw data path data_path and save path gt16bit_dir in save_16bitpng_gt.py, then

python save_16bitpng_gt.py

• Secondly, run the script extract_depth.m in Matlab before you modify the index file and save path of the code. Taken RGB image and projected depth groud-truth as input, this script can generate the dense depth map with relative depth value.

• Finally, you can remap them to a colorful result for visulization:

import cv2
im = cv2.imread('image/path')
im_color = cv2.applyColorMap(im, cv2.COLORMAP_JET)

Example

original rgb image:

Sparse depth map that only about 16% pixels available:

Colorful dense map: