A Real-time Pseudo Dense Depth Guided Depth Completion Network (Non-official)

This repository provides a non-official implementation of the paper: DenseLiDAR: A Real-time Pseudo Dense Depth Guided Depth Completion Network

Jiaqi Gu. et al, DenseLiDAR: A Real-Time Pseudo Dense Depth Guided Depth Completion Network. ICRA 2021

We express our gratitude to Jiaqi Gu et al. for their groundbreaking work on DenseLiDAR. This implementation is inspired by their paper presented at ICRA 2021. We also appreciate the contributions of the open-source community and the resources provided by PyTorch.

This project was conducted as part of the Internship program at Mobiltech and VIP-Lab at Gachon University.

Implementation by : 구도연, 김다혜, 조재현 [VIP-Lab at Gachon University.]
Project Supervision : 정선재 [Mobiltech.]

Disclaimer

Please note that this implementation is not the official code provided by the authors of the paper. Therefore, performance metrics obtained using our code may differ from those reported in the original paper due to differences in implementation details, parameter settings, and hardware.

DenseLiDAR Architecture

DenseLiDAR Depth Completion Results on paper

Note: Our implementation results are not the same.

Our Results

You can check our implementation results of Depth Completion/3D Visualization on this link.

Predictions

3D Visualizations

Requirements

Ubuntu 20.04 LTS
Python 3.8
CUDA 10.2, 11.08

Installation

git clone https://github.com/suunnnnnj/DenseLiDAR.git
cd DenseLiDAR
pip install -r requirements.txt

Dataset

kitti lidar dataset download

Visit the KITTI homepage and download the two datasets below on the depth completion page.

Download annotated depth maps data set (14 GB)
Download projected raw LiDAR scans data set (5 GB)

kitti_raw dataset download

cd datasets/kitti_depth
wget https://github.com/youmi-zym/CompletionFormer/files/12575038/kitti_archives_to_download.txt
wget -i kitti_archives_to_download.txt -P kitti_raw/
cd kitti_raw
unzip "*.zip"

datasets
├──data_depth_annotated
|     ├── train
|     └── val
├── data_depth_velodyne
|     ├── train
|     └── val
└── kitti_raw
      ├── 2011_09_26
      ├── 2011_09_28
      ├── 2011_09_29
      ├── 2011_09_30
      └── 2011_10_03

Dataset configuration

Download all three data, configure it as in the file structure above, and then run the commands below to configure the dataset to be used for learning.
Once completed, you will have the following file structure.

python submodules/dataset/dataset_settings.py

datasets
├──data_depth_annotated
|     ├── train
|     └── val
├── data_depth_velodyne
|     ├── train
|     └── val
├── kitti_raw
|     ├── train
|     └── val
├── pseudo_depth_map
|     ├── train
|     └── val
└── pseudo_gt_map
      ├── train
      └── val

Usage

Training

python train.py --data_path [YOUR_DATASET_PATH] --epochs [EPOCHS] --checkpoint [CHECKPOINT] --batch_size [BATCH_SIZE] --gpu_nums [YOUR_GPU_NUMS]

Arguments

--data_path: your dataset path | default: None
--epochs: number of epochs to train | default: 40
--checkpoint: number of epochs to making checkpoint | default: 5
--batch_size: number of batch size to train | default: 1
--gpu_nums: number of gpus to train | default: 1

Example

python train.py --data_path datasets/ --epochs 40 --batch_size 16 --gpu_nums 4

Demo

python demo.py --model_path [YOUR_MODEL_PATH] --image_path [YOUR_IMAGE_PATH] --sparse_path [YOUR_LIDAR_POINT_PATH] --pseudo_depth_map_path [YOUR_PSEUDO_DEPTH_MAP_PATH] --output_path [YOUR_SAVE_PATH]