A software to automate the process of constructing the synthetic dataset in COCO format, training state-of-the-art deep neural networks (DNNs) from the framework Detectron2, then doing depth estimation for the scene. The results are detected objects with corresponding [x, y, z] coordinates relative to the camera.
The whole pipeline could be seen in Figure 1 Figure 1: Data processing and training pipeline
An output is shown in Figure 2 Figure 2: An example output
torch
>=1.8
detectron2
Change the main configurations in the file config.yaml
.
Configurations for each process (preparing the dataset, training the models on a dataset, and reconstructing 3D positions) are in the yaml
files in the directories dataset
, dataset_model
and reconstruct_3d
.
After you are done with the configurations, run python main.py
.