Dataset Replay

The nvblox/apps/datasets folder contains sample GXF apps to use Nvblox with offline datasets. Currently, two datasets can be downloaded:

• Replica: This is a synthetic dataset containing RGB-D frames and ground-truth poses. It can be downloaded using wget https://cvg-data.inf.ethz.ch/nice-slam/data/Replica.zip. Unzip the archive to access the dataset.

• Redwood: This is a real dataset recorded with an Asus Xtion camera and estimated camera poses. Download the RGB-D sequences and camera poses from here and extract them to a corresponding folder with the sequence name in lowercase (eg. Lobby -> lobby).

Here is an example folder structure for the office sequence of the Replica dataset:

            

            
datasets/
└── replica/
    └── office0/
        ├── results/
        └── traj.txt
        

Here is an example folder structure for the lobby sequence of the Redwood dataset:

            

            
datasets/
└── redwood/
    └── lobby
        ├── pose_lobby
        └── rgbd_lobby
            └── lobby
                ├── depth
                └── image
        

Use the following command to run the reconstruction:

            

            
docker run -it --gpus all --network=host \
  -v <DATASET_PATH>:<DATASET_PATH> \
  nvcr.io/nvstaging/isaac-amr/nvblox_dataset \
  --path <DATASET_PATH> --name <DATASET/SEQUENCE>
        

• <NAME/SEQUENCE> specifies the slash-separated dataset name and sequence, for example redwood/apartment or replica/office01

• <DATASET_PATH> is the root path to where the all the datasets are located. If the sequence is located in /my/datasets/redwood/apartment, this shoud be set to /my/datasets

• <IMAGE_TAG> refers to the Docker image tag (e.g. master-aarch64).

For development, it is useful to run apps without generating Docker images:

            

            
bazel run //extensions/nvblox/apps/datasets --path <DATASET_PATH> --name <NAME/SEQUENCE>
        

If you are using Dazel, make sure the datasets are mounted. Modify your ~/.dazelrc file to include the following:

            

            
DAZEL_VOLUMES = [volume_host:volume_dazel]
        

Then restart Dazel with dazel stop.

• Color visualization will not work because of the image stride.