DD3D: Is Pseudo-Lidar needed for Monocular 3D Object Detection?

August 8, 2020

Have you ever wondered how autonomous vehicles and 3D object detection technologies accurately perceive their surroundings? Enter the world of DD3D, where we utilize innovative techniques like Pseudo-Lidar to enhance monocular 3D object detection. In this guide, we’ll explore how to set it up, work with datasets, conduct experiments, and troubleshoot common issues. Let’s embark on this technical journey!

Installation

To ensure a smooth establishment of the DD3D environment, we recommend using Docker for reproducibility. Below are the steps to get started:

Open your terminal (best suited for Ubuntu 18.04).
Clone the DD3D repository:

git clone https://github.com/TRI-ML/dd3d.git

Navigate into the directory:

cd dd3d

Build the Docker image (if using CUDA 10.2):

make docker-build

If you require CUDA 11.1:

make docker-build DOCKERFILE=Dockerfile-cu111

Check your Nvidia driver version and CUDA compatibility to select the proper Dockerfile.

Datasets

For our experiments, datasets need to be organized correctly. Here’s how to set them up:

KITTI Dataset

Download the KITTI 3D dataset from the KITTI website.

To structure your dataset, follow this organization:

DATASET_ROOT
KITTI3D
train and test folders, calibrated images, and labels

nuScenes Dataset

Similarly, download the nuScenes dataset from the nuScenes website. The desired folder structure is outlined above.

Running Experiments

Configuring and running experiments requires the use of hyra for effective management. The command to start an experiment is:

bash scripts/train.py +experiments=your-experiment-file

Understanding the Code with an Analogy

Think of our installation and dataset preparation process as cooking a recipe. Firstly, you gather all your ingredients (cloning the repository, setting up Docker, downloading datasets). Next comes the cooking process (running experiments), where your careful measurements (dataset structure) guarantee the final dish (accurate 3D object detection). Each step plays a crucial role in creating a successful meal!

Troubleshooting

Encountering issues during installation or experimentation? Here are some tips to guide you:

Ensure Docker is installed correctly and running.
Double-check your Nvidia driver version against the required CUDA version.
Verify dataset structures are accurately maintained.
If running out of GPU memory, consider using gradient accumulation. More info can be found here.

For more insights, updates, or to collaborate on AI development projects, stay connected with fxis.ai.

Conclusion

At fxis.ai, we believe that such advancements are crucial for the future of AI, as they enable more comprehensive and effective solutions. Our team is continually exploring new methodologies to push the envelope in artificial intelligence, ensuring that our clients benefit from the latest technological innovations.