argoverse/argoverse-api

Official GitHub repository for Argoverse dataset

What it solves

It provides a standardized programmatic interface to access and manipulate the Argoverse dataset, which contains high-definition (HD) maps, 3D tracking data, and motion forecasting scenarios for autonomous driving research.

How it works

The API acts as a data loader and utility suite that allows researchers to interact with three primary components:

  • HD Maps: Tools to represent and query city-scale map data.
  • Tracking Dataset: Access to raw LiDAR and camera data, egovehicle poses, and track annotations.
  • Forecasting Dataset: Access to sequences of agent trajectories (cars, pedestrians) for predicting future motion.

It includes specialized utilities for coordinate transformations (SE(3) and SE(2)) to align data between the map, the vehicle, and the sensors, as well as visualization scripts for rendering bird's-eye-view maps and 3D cuboids on images.

Who it’s for

Researchers and developers working on autonomous vehicle perception, 3D object tracking, and motion forecasting.

Highlights

  • Comprehensive Data Loading: Dedicated loaders for both tracking and forecasting datasets.
  • Sensor Fusion Support: Provides calibration data to align LiDAR and camera sensors.
  • Coordinate Transform Utilities: Built-in handling of complex spatial transformations between city, vehicle, and object frames.
  • Visualization Tools: Scripts for rendering 3D bounding boxes (cuboids) and ground LiDAR points on images.

Related

  • Project
    cvg/GeoCalib

    GeoCalib is a research‑grade Python library that predicts camera intrinsics and scene gravity from a single image using a deep network plus geometric optimisation. It supports multiple lens models, partial priors, batch and multi‑camera rig calibration, and ships with easy‑install inference code, a Gradio web demo, a Colab notebook, and a live webcam demo. Pre‑trained models are trained on the OpenPano dataset (≈37 k perspective crops from HDR panoramas) and achieve state‑of‑the‑art results on benchmarks such as LaMAR, MegaDepth, TartanAir, and Stanford2D3D.

  • Project
    facebookresearch/mobile-vision

    A collection of hardware-aware efficient ConvNet models and tools, including FBNet and ChamNet, designed to optimize computer vision performance and latency on mobile devices.

  • Project
    XiandaGuo/OpenStereo

    A comprehensive benchmark and framework for stereo matching that integrates 17 datasets and multiple state-of-the-art models for depth estimation.

  • Project
    FORTH-ModelBasedTracker/MocapNET

    A real-time 2D-to-3D human pose estimator that converts RGB video feeds into standard BVH motion capture files for 3D animation.