Argoverse Competitions 2025

Event: CVPR 2025 Workshop on Autonomous Driving · Duration: 25 min · ▶ Watch on YouTube

Abstract

This workshop session provides an overview of the Argoverse Competitions for CVPR 2025, covering three key challenges: (Un)supervised Scene Flow, Scenario Mining, and Multi-Agent Motion Forecasting. Speakers detail the datasets, evaluation methodologies, and highlight state-of-the-art approaches and winning solutions from the past year. The session emphasizes the community-driven nature of Argoverse and the continuous efforts to push the boundaries of autonomous driving research through robust datasets and challenging tasks.

Speakers

  • Ishan Khatri — Stack AV (formerly Argo AI)
  • Neehar Peri — Argoverse

Talks (3)

  • 00:02:34 — Ishan Khatri: (Un)supervised Scene Flow
    • Discusses advancements in scene flow evaluation metrics, moving from EPE to Bucket Normalized EPE to better assess performance on dynamic objects, and highlights new methods like ΔFlow and Floxels.
  • 00:10:05 — Neehar Peri: Scenario Mining
    • Introduces the new scenario mining challenge, focusing on finding interesting multi-agent interactions using natural language queries and multimodal data, and presents the RefAV dataset and a baseline leveraging LLMs.
  • 00:10:05 — Neehar Peri: Multi-Agent Motion Forecasting
    • Reviews the evolution of the Argoverse motion forecasting challenge over seven years, emphasizing the shift to multi-world forecasting for predicting joint future states of all actors, and highlights the winning method IMR.

Key Takeaways

  • Argoverse datasets are continuously evolving with new challenges and community support, including a new Scenario Mining challenge for 2025.
  • Scene flow evaluation has advanced significantly, with new metrics like Bucket Normalized EPE addressing limitations of traditional EPE by focusing on dynamic objects and normalizing errors across classes.
  • Novel methods like ΔFlow and Floxels demonstrate significant progress in both supervised and unsupervised scene flow, offering improved accuracy and computational efficiency.
  • The Scenario Mining challenge explores the use of LLMs for program synthesis to identify complex multi-agent interactions from multimodal data, highlighting current limitations in semantic ambiguity and expressiveness.
  • Multi-agent motion forecasting is moving towards predicting joint future states (multi-world forecasting) for all actors, crucial for safe autonomous vehicle planning, with continuous year-over-year performance improvements.

Methods / Models / Datasets Mentioned

  • Argoverse
  • Argoverse 2
  • Waymo Open
  • nuScenes
  • ONCE
  • Average Endpoint Error (EPE)
  • Threeway EPE
  • Bucket Normalized EPE
  • ΔFlow
  • FastFlow3D
  • DeFlow
  • NSFP
  • ZeroFlow XL 5x
  • EulerFlow
  • Floxels
  • RefAV
  • Referential Multi-Object Tracking
  • HOTA (Higher Order Tracking Accuracy)
  • HOTA-Temporal
  • Log Balanced Accuracy
  • Timestamp Balanced Accuracy
  • RefProg
  • LLM (Large Language Model)
  • VLM (Vision Language Model)
  • Le3DE2E Detector (2023)
  • Link
  • SECOND
  • FocalFormer3D
  • Weighted Box Fusion (WBF)
  • IMR (Iterative Mode-World Weighted Regression for Multi-Agent Trajectory Prediction)
  • EACON-IMR
  • Lite-QCNet
  • LOF
  • QCNet

Topics

Autonomous Driving · Argoverse · Scene Flow · Scenario Mining · Motion Forecasting · LiDAR · Datasets · Evaluation Metrics · Large Language Models · Multi-Agent Systems

Notes

Open for commentary — connections to other work, critiques, follow-up reading.