Somewhere on disk, have an argoverse2/
folder so that the downloaded files live inside
argoverse2/train
argoverse2/val
argoverse2/test
Please note that when downloaded from the cloud, these files may have a different top level directory format (their stored format keeps changing); you can solve this by moving the files or symlinking the appropriate directories into a different tree.
Generate the train and val supervision labels to
argoverse2/train_sceneflow_feather
argoverse2/val_sceneflow_feather
To generate these supervision labels, use the generation script in data_prep_scripts/argo/create_gt_flow.py
. We have uploaded a prebuilt DockerHub image for running the generation script; it can be run using ./launch.sh
.
We provide the Argoverse 2 NSFP Pseudolabels for the Sensor split in the S3 bucket
s3://argoverse/assets/av2/scene_flow/sensor/
and for our subsets of the Lidar split
s3://argoverse/assets/av2/scene_flow/lidar/subsample/
To get started, we provide a directly downloadable tiny demo dataset (5.5MB).
argoverse2_tiny
contains four subfolders:
argoverse2_tiny/val
: a single sequence with the single frame pairargoverse2_tiny/val_sceneflow_feather
: the supervised ground truth for this frame pairargoverse2_tiny/val_nsfp_flow_feather
: the NSFP pseudolabels for this frame pairargoverse2_tiny/val_supervised_out
: the output of the forward pass of FastFlow3D, a supervised scene flow estimator.
Download Waymo Open v1.4.2 (earlier versions lack map information) and the scene Flow labels contributed by Scalable Scene Flow from Point Clouds in the Real World from the Waymo Open download page. We preprocess these files, both to convert them from an annoying proto file format to a standard Python format and to remove the ground points.
Do this using
data_prep_scripts/waymo/rasterize_heightmap.py
-- generate heightmaps in a separate folder used for ground removaldata_prep_scripts/waymo/extract_flow_and_remove_ground.py
-- extracts the points into a pickle format and removes the groundplane using the generated heightmaps
We have uploaded a prebuilt DockerHub image for running the Waymo conversion scripts.
We have also provided a directly downloadable tiny demo dataset (3.1MB).
waymo_open_processed_flow_tiny
contains two subfolders:
training
: a single frame pair of waymo datatrain_nsfp_flow
: the flow labels for the framepair
TODO