Sparse LiDAR Depth Completion

Code for DAN-Conv: Depth Aware Non-local Convolution for LiDAR Depth Completion.

Requirments

pytorch >= 1.6.0
tqdm
fire

Prepare training data

You need to download the entire raw KITTI dataset to train.

Warning: it weighs about 175GB, so make sure you have enough space to unzip too!

Our default settings expect that you have converted the png images to jpeg.

The path containing KITTI datasets should be organized like this:

/path/to/kitti
    raw/
        2011_09_26/
            2011_09_26_drive_0001_sync/
                image_02/
                image_03/
    test_depth_completion_anonymous/
        image/
        velodyne_raw/
    val_selection_cropped/
        image/
        velodyne_raw/
        groundtruth_depth/

Build kNN lib

cd nearest_neighbors
make

Train on Single GPU

Run

python train.py --data_path /path/to/kitti --split full

Train on multiple GPUs

Modify nproc_per_node in script train_multi_gpus.sh to the number of GPUs you have and adjust the batch size.

Run