This is a system that covers all aspects of a 3D object detection pipeline. It essentially consists of three main phases: Data preprocessing, training and inference. The individual steps are described in more detail in the respective directories. But first you have to place your raw data in this project.
So far, this project supports two sensor devices: Intel RealSense and StereoLabs ZED. Exemplary data is provided for both devices. In the case of the RealSense, the data has not yet been extracted and is stored in .bag
files. In case of the ZED, the data has already been extracted (RGB images, depth maps, point clouds).
Navigate to raw_data/
and place the exemplary data in the respective directory.
Note that the containerized training environment and inference API run on a Linux distribution using the NVIDIA runtime environment. Therefore, the host system itself must be a Linux distribution and have CUDA and cuDNN installed. For more information, refer to the corresponding README files.