This project implements the ADNN architecture described in “Deep Convolutional Compressed Sensing for LiDAR Depth Completion” (http://arxiv.org/abs/1803.08949)

• Tensorflow 1.4
• Numpy 2.0
• PIL

This project uses Tensorflow’s binary tfrecord file to speed up training. Perform the following steps to set up the datasets for training and testing

1. Download the KITTI depth completion dataset from http://www.cvlibs.net/datasets/kitti/eval_depth_all.php
2. Unzip the various archives using the directions provided in the downloads
3. Change the final line of kitti_depth_to_tfrecord.py to reflect the locations of your data and the desired location for the tfrecords, then run the file.
4. Change lines 38, 43, 49, and 54 of main.py to reflect these locations as well.

In order to train the three layer model described in the paper, create an output directory and run the command

python3 main.py <output directory> --train_size 20000 --val_size 2000

Running the command with the help flag will output a description of the other training and validation options.