DeepLiDAR (Python3, Pytorch 1.4.0)

Thanks to JiaxiongQ and valgur

This repository is the implementation for DeepLiDAR: Deep Surface Normal Guided Depth Prediction for Outdoor Scene from Sparse LiDAR Data and Single Color Image. There are some difference between author's repo and mine.

(1) Rewrite the code referenced from author's repo with python3.6 and newest version of pytorch.

(2) Clarify the structure of KITTI depth completion data

(3) Make it easier to reproduce my result (not author's result, because by considering computational resources and time, there are some differences between author's implementation and mine)

Reduce model parameters from 143,981,012 to 47,750,544. (Original model is too large to put onto single GPU, author used 3 GeForce GTX 1080 Ti GPUs)
Smaller training image (128 x 256). (Author uses 256 x 512)
Use less data to train (20000 images). (Author used 85898 images, and it takes about 3 days to train)
Slight implementation difference

(4) Add comments on the code and make it more flexible

(5) Add tensorboard visualization for every epoch

Requirements

Python 3.6.8

pip3 install -r requirements.txt

Data Preparation

Download the KITTI Depth Dataset from their website.
Use the following scripts to extract corresponding RGB images from the raw dataset.

./download/rgb_train_downloader.sh
./download/rgb_val_downloader.sh

The overall code, data, and results directory is structured as follows
- data_depth_annotated: ground truth data (dense depth)
- data_depth_velodyne: sparse data (LiDAR)
- data_rgb: RGB image
- data_depth_annotated: Used to train surface normals, and generated from data_depth_annotated by using generate_normals.py

.
├── KITTI_data
|   ├── data_depth_annotated
|   |   ├── train
|   |   |   ├── 2011_09_26_drive_0001_sync
|   |   |       ├── proj_depth
|   |   |           ├── groundtruch
|   |   |               ├── image02
|   |   |               ├── image03
|   |   ├── val
|   |   |   ├── (the same as train)
|   ├── data_depth_velodyne
|   |   ├── train
|   |   |   ├── 2011_09_26_drive_0001_sync
|   |   |       ├── proj_depth
|   |   |           ├── velodyne_raw
|   |   |               ├── image_02
|   |   |               ├── image_03
|   |   ├── val
|   |   |   ├── (the same as train)
|   ├── data_rgb
|   |   ├── train
|   |   |   ├── 2011_09_26_drive_0001_sync
|   |   |       ├── image_02
|   |   |       ├── image_03
|   |   ├── val
|   |   |   ├── (the same as train)
|   ├── data_depth_normals
|   |   ├── (the same as data_depth_annotated)
|   └── depth_selection
|   |   ├── test_depth_completion_anonymous
|   |   ├── test_depth_prediction_anonymous
|   |   ├── val_selection_cropped

Path setting

Please set the path in the env.py first

SAVED_MODEL_PATH = './saved_model' # save model in this directory
KITTI_DATASET_PATH = /PATH/TO/KITTI_data/ # path to KITTI_data as structured in the above
PREDICTED_RESULT_DIR = './predicted_dense' # path to save predicted figures (used in test.py)

To generate data_depth_normals

Credit to https://github.com/valgur/surface-normal

First, install the surface_normal library from valgur/surface-normal.

pip install -v git+https://github.com/valgur/surface-normal.git

Second, run the following script to generate data_depth_normals

python3 generate_normals.py

Pretrained Model Download

See the next section

Usage

Train and validation

python3 main.py -b <BATCH_SIZE> -e <EPOCH> -m <SAVED_MODEL_NAME> -l <MODEL_PATH> -n <NUM_DATA> -cpu
    -b <BATCH_SIZE>
        batch size used for training and validation
    -e <EPOCH>
        the number of epoch for training and validation
    -m <SAVED_MODEL_NAME>
        the model name (be saved in SAVED_MODEL_PATH)
    -l <MODEL_PATH>
        specified the model path if you want to load previous model
    -n <NUM_DATA>
        the number of data used for training. (set -1 if you want to use all the training data (85898))
    -cpu
        if you want to use CPU to train

There are three different stages of training model.

(N) Train surface normal
(D) Train depth of color pathway and normal pathway
(A) Train the whole network (fix surface normal network)

We test the model with 3 different settings

(A) Train N stage for 15 epoch, train D stage for 15 epoch, and then train A stage for 15 epoch (download)

(B) Train A for 12 epochs (due to early stop with patience 10, no update parameter of deepLidar.normal (random)) (download)

Test

Test on depth_selection/val_selection_cropped data

python3 test.py -m <MODEL_PATH> -n <NUM_DATA> -cpu
    -n <NUM_DATA>
        the number of data used for testing. (set -1 if you want to use all the testing data (1000))
    -m <MODEL_PATH>
        the path of loaded model
    -cpu
   
        if you want to use CPU to test
    -s
        if you want to save predicted figure in PREDICTED_RESULT_DIR

The following results are testing on depth_selection/val_selection_cropped data

Setting	RMSE (mm)
A (download)	1191.6127
B (download)	1182.6613
C (download)	1026.8722

Test a pair of inputs

Run a pair of rgb and lidar image as input, and then save the predicted dense depth

python3 test_a_pair.py --model_path </PATH/TO/PRETRAIN_MODEL> --rgb <PATH/TO/RGB_IMAGE> --lidar <PATH/TO/LIDAR_IMAGE>
                       --saved_path </SAVED_FIGURE/PATH>
    --model_path <MODEL_PATH>
        the path of pretrained model  
    --rgb <PATH>
        the path of rgb image
    --lidar <PATH>
        the path of lidar image
    --saved_path <PATH>
        the path of saved image

Tensorboard Visualization

tensorboard --logdir runs/

Experiment of setting A

Input data (rgb image, lidar image)

Groundtruth (surface normal, dense depth)

Masked predicted result (masked surface normal, masked dense depth)

Predicted result (surface normal, dense depth)

Citation

If you use this method in your work, please cite the following:

@InProceedings{Qiu_2019_CVPR,
author = {Qiu, Jiaxiong and Cui, Zhaopeng and Zhang, Yinda and Zhang, Xingdi and Liu, Shuaicheng and Zeng, Bing and Pollefeys, Marc},
title = {DeepLiDAR: Deep Surface Normal Guided Depth Prediction for Outdoor Scene From Sparse LiDAR Data and Single Color Image},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2019}
}

mac137 / DeepLiDAR-1