Awesome

Platform	Build Status
Ubuntu 20.04.3

Waymo Open Dataset -> Argoverse Converter

Repo Overview

Simple utility to convert Waymo Open Dataset raw data, ground truth, and detections to the Argoverse format [ paper, repo ], run a tracker that accepts Argoverse-format data, and then submit to Waymo Open Dataset leaderboard.

Achieves the following on the Waymo 3d Tracking Leaderboard, using run_ab3dmot.py from my argoverse_cbgs_kf_tracker repo.

Data Format Overview

Waymo raw data follows a rough class structure, as defined in Frame protobuffer. Waymo labels and the detections they provide also follow a rough class structure, defined in Label protobuffer.

Argoverse also uses a notion of Frame at 10 Hz, but only for LiDAR and annotated cuboids in LiDAR. This is because Argoverse imagery is at 30 Hz (ring camera) and 5 Hz (stereo). Argoverse data is provided at integer nanosecond frequency throughout, whereas Waymo mixes seconds and microseconds in different places. Argoverse LiDAR points are provided directly in the egovehicle frame, not in the LiDAR sensor frame, as .PLY files.

A Waymo object defines a coordinate transformation from the labeled object coordinate frame, to the egovehicle coordinate frame, as an SE(3) comprised of rotation (derived from heading) and a translation:

object {
  box {
    center_x: 67.52523040771484
    center_y: -1.3868849277496338
    center_z: 0.8951533436775208
    width: 0.8146794438362122
    length: 1.8189797401428223
    height: 1.790642261505127
    heading: -0.11388802528381348
  }
  type: TYPE_CYCLIST
}
score: 0.19764792919158936
context_name: "10203656353524179475_7625_000_7645_000"
frame_timestamp_micros: 1522688014970187

Argoverse data is provided similarly, but in JSON with full 6 dof instead of 4 dof transformation from labeled object coordinate frame to egovehicle frame. A quaternion is used for the SO(3) parameterization:

{
  "center": {"x": -25.627050258944625, "y": -3.6203567237860375, "z": 0.4981851744013227}, 
  "rotation": 
    {"x": -0.000662416717311173, 
    "y": -0.000193607239199898, 
    "z": 0.000307246307353097, "w": 0.999999714659978}, 
    "length": 4.784992980957031, 
    "width": 2.107541785708549, 
    "height": 1.8, 
    "track_label_uuid": "215056a9-9325-4a25-bbbd-92d445d60168", 
    "timestamp": 315969629119937000, 
    "label_class": "VEHICLE"
},

Whereas Waymo uses "context.name" as a unique log identifier, Argoverse uses "log_id".

Installation

Before you use this code, you will need to download a few packages. To install the waymo_open_dataset library, use the commands here to install the pre-compiled pip packages. You will also need to download argoverse. You can use this command to download the package.

Guide to Repo Code Structure

• waymo2argo/waymo_dets_to_argoverse.py: Convert provided Waymo detections to Argoverse format. Use shards to not exceed Colab RAM.
• waymo2argo/dump_waymo_persweep_detections.py: Given sharded JSON files containing labeled objects or detections in random order, accumulate objects according to frame, at each nanosecond timestamp. Write to disk.
• waymo_data_splits.py: functions to provide list of log_ids's in Waymo val and test splits, respectively.
• waymo2argo/waymo_raw_data_to_argoverse.py: Extract poses, LiDAR, images, and camera calibration from raw Waymo Open Dataset TFRecords.
• run_tracker.sh: script to run AB3DMOT-style tracker on Argoverse-format detections, and write tracks to disk.
• create_submission_bin_file.py: Given tracks in Argoverse format, convert them to Waymo submission format.

Converting Waymo Raw Data to Argoverse Format

To convert the Waymo dataset to Argoverse format, you will need to run

python waymo_raw_data_to_argoverse.py --waymo-dir /path-to-waymo-data/ --argo-dir /path-to-write-argo-data/

e.g.

python waymo2argo/waymo_raw_data_to_argoverse.py --save-labels true --argo-dir waymo_data_in_argoverse_form_2021_10_06 --waymo-dir /srv/datasets/waymo_opendataset/waymo_open_dataset_v_1_0_0/training

After running this script, you will also need to run

python argoverse/utils/make_track_label_folders.py /path-to-write-argo-data/

to create the track_labels_amodal folder. There is more information about that here.

Usage Instructions for Waymo Leaderboard

1. Download test split files (~150 logs) from Waymo Open Dataset website which include TFRecords.
2. Download provided detections from PointPillars Progressive Population-Based Augmentation detector, as .bin files.
3. Convert to Argoverse format using scripts provided here in this repo.
4. Run tracker
5. Convert track results to .bin file
6. Populate a submission.txtpb file with metadata describing your submission (example here).
7. Run create_submission binary to get tar.gz file. Binary is only compiled using Bazel. I used Google Colab.
8. Submit to Waymo eval server.

Submission process overview is here.

References

@InProceedings{Chang_2019_CVPR,
author = {Chang, Ming-Fang and Lambert, John and Sangkloy, Patsorn and Singh, Jagjeet and Bak, Slawomir and Hartnett, Andrew and Wang, De and Carr, Peter and Lucey, Simon and Ramanan, Deva and Hays, James},
title = {Argoverse: 3D Tracking and Forecasting With Rich Maps},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2019}
}