LED : Light Enhanced Depth Estimation at Night

Official implementation of the encoder-decoder in LED. Code, information and download link for the Nighttime Synthetic Drive Dataset.

Arxiv | Project Page

If you use our dataset in your research, please consider citing:

@article{deMoreau2024led,
    title = {LED: Light Enhanced Depth Estimation at Night},
    author = {De Moreau, Simon and Almehio, Yasser and Bursuc, Andrei and El-Idrissi, Hafid and Stanciulescu, Bogdan and Moutarde, Fabien},
    journal = {arXiv preprint arXiv:2409.08031},
    year = {2024},
}

Dataset

The code is meant to be used with the Nighttime Synthetic Drive Dataset. Please see the project page to download the dataset.

Full dataset size is 483Go, to facilitate download, it has been splitted in multiple zip files, separated between Pattern-illuminated and High Beam part of the dataset, each subset (train/val/test), and each type of annotation (object detection, depth, semantic segmentation...).

File structure

If all zips are downloaded and extracted the directories should have this format :

Nighttime Synthetic Drive Dataset/
├── HB
│   ├── test
│   │   └── wuppertal
│   │       └── ...
│   ├── train
│   │   ├── china
│   │   │   └── ...
│   │   ├── herrenberg
│   │   │   └── ...
│   │   └── ottosuhrallee
│   │       └── ...
│   └── val
│       └── hamburg
│           └── ...
└── Pattern
    ├── test
    │   └── wuppertal
    │       └── ...
    ├── train
    │   ├── china
    │   │   └── ...
    │   ├── herrenberg
    │   │   └── ...
    │   └── ottosuhrallee
    │       └── ...
    └── val
        └── hamburg
            └── ...

And each map contains those directories :

map
├── bounding_box_2d_loose
├── bounding_box_2d_tight
├── bounding_box_3d
├── camera_params
├── distance_to_camera
├── distance_to_image_plane
├── dynamics
├── instance_segmentation
├── ldr_color
├── normals
├── occlusion
├── semantic_segmentation
└── transforms

An example of pytorch dataset module is available in DriveSimDataset.py.

Install

1. Clone the repository

   git clone https://github.com/SimondeMoreau/LED.git
   

2. Install Python 3.11 and pip:

   conda create -n LED python=3.11 pip
   conda activate LED
   

3. Install the required dependencies:

   pip install -r requirements.txt
   

Training

To train a model, use the train.py script. Here's how to use it:

python train.py [dataset_root] [Pattern/HB]

Testing

To test a model, use the test.py script. Here's how to use it:

python test.py [dataset_root] [Pattern/HB] [experience_name]

License

Code is released under the Apache 2.0 license.

Dataset license is available here : LICENSE_Dataset. It is released for research and non-commercial purposes.