WIP: Currently cleaning up the code

full-scene-reconstruction

Full Scene Reconstruction of a 3D scene (geometry, instance ids, semantic labels, and color) from an RGB Image. Based on Panoptic 3D Scene Reconstruction From a Single RGB Image: code | paper

Install

1. Install Docker following the instructions on the link and nvidia-docker (for gpu support).

2. Clone this repo and its submodules

   git clone --recursive -j8 git@github.com:jdgalviss/full-scene-reconstruction.git
   cd full-scene-reconstruction

3. Build Docker Container

   docker build . -t scene_reconstruction

Run Inference

1. Run Docker container

   source run_docker.sh 

2. Download pretrained model and put it in the panoptic-reconstruction/data folder

3. Run inference inside docker container

   python tools/test_net.py

Run Training

2. Run Docker container. Make sure you set the in the run_docker.sh file to the correct path where you downloaded and extracted the data.

   source run_docker.sh 

3. Train Inside docker:

   python tools/train_full_reconstruction.py --config configs/front3d_train_3d.yaml --output-path output/

Run Evaluation

1. Run Docker container

   source run_docker.sh 

2. Run evaluation: Make sure you set the correct path to the trained model in front3d_evaluate.yaml

   python tools/evaluate_net.py --config configs/front3d_evaluate.yaml --output output/

3. Run jupyter lab inside Docker

   jupyter lab --ip=0.0.0.0 --port=8888 --allow-root --no-browser

FAQ

If docker build fails due to missing cuda libraries, edit the file in /etc/docker/daemon.json so it looks like this:

{
    "runtimes": {
        "nvidia": {
            "path": "nvidia-container-runtime",
            "runtimeArgs": []
        }
    },
    "default-runtime": "nvidia"
}