Traffic3D Challenge 2025

Important Dates

Challenge Overview

Welcome to the Traffic3D Challenge 2025, a premier international competition focused on advancing 3D traffic scene understanding through point cloud processing. This challenge is part of the 13th International Conference on Mobile Mapping Technology (MMT 2025) and aims to bring together researchers and practitioners to tackle fundamental challenges in railway and road scene understanding.

The significance of this research domain is multifaceted:

  1. Advancing autonomous vehicle perception systems through robust 3D interpretation
  2. Enabling millimeter-accurate digital twin modeling of transportation infrastructure
  3. Developing next-generation intelligent transportation systems with comprehensive scene understanding capabilities across diverse environmental conditions

As the volumetric complexity and semantic diversity of traffic-related point cloud data continue to expand exponentially, there exists an urgent methodological imperative to develop computationally efficient, geometrically precise, and semantically robust point cloud processing frameworks. Both efficient algorithms and post-processing methods are welcome in this challenge. This workshop provides a structured platform for the dissemination of algorithmic innovations, methodological advances, and empirical findings in this rapidly evolving field. This challenge offers generous prizes. Students, teachers, and researchers in relevant fields are all welcome to participate.

Challenge Tracks

Below is the basic information about the three tracks. The dataset introduction page offers an overview of the dataset and download links, while the dataset competition page provides detailed competition guidelines and submission procedures.

Track 1: Railway Scene Point Cloud Semantic Segmentation

Dataset: WHU-Railway3D

  • Dataset Introduction Page: https://github.com/WHU-USI3DV/WHU-Railway3D
  • Dataset Competition Page:
  • Data Description: 4.6 billion points across 30 km of railway, covering urban, rural, and plateau environments. The semantic annotation schema comprises 11 distinct categories, including critical infrastructure elements such as rails, track beds, masts, support devices, and overhead lines.
  • Tasks: Develop robust semantic segmentation models for railway point clouds that classify 11 categories spanning railway infrastructure, built structures, and natural elements across diverse railway environments, addressing challenges of occlusion and class imbalance.
  • Evaluation Metrics: Mean Intersection over Union (mIoU), per-category IoU scores, Overall Accuracy (OA), and category-specific performance metrics for railway-critical infrastructure elements.

Track 2: Road Scene Point Cloud Semantic/Instance Segmentation

Dataset: WHU-Urban3D

  • Dataset Introduction Page: https://whu3d.com
  • Dataset Competition Page: https://www.codabench.org/competitions/7197/
  • Data Description: Integrated ALS and MLS point clouds covering over 3.6 million square meters of diverse urban typologies. The dataset is structured into more than 80 distinct urban blocks with exhaustive semantic annotations, providing complementary perspective modes (vertical and horizontal) of urban scenes.
  • Tasks: Develop algorithms for both semantic and instance segmentation of urban point clouds, focusing on accurate object delineation in complex city environments and reliable detection of varied urban elements across 19 semantic categories.
  • Evaluation Metrics: Mean Intersection over Union (mIoU), per-category IoU values, Overall Accuracy (OA), mean coverage (mCov), mean weighted coverage (mWCov), mean precision (mPre), mean recall (mRec), and mean F1-score (mF1).

Track 3: Road Scene Lane Mapping

Dataset: WHU-Lane

  • Dataset Introduction Page: https://github.com/WHU-USI3DV/LaneMapping
  • Dataset Competition Page: https://www.codabench.org/competitions/7489/
  • Data Description: High-resolution Mobile Laser Scanning (MLS) systems capturing over 98 kilometers of road infrastructure. The annotation schema provides granular classification of lane markings, differentiating between discontinuous and continuous demarcations, with instance-level identification for individual lane segments.
  • Tasks: Develop algorithms to extract lane lines and semantic attributes directly from MLS point clouds, with emphasis on creating geometrically accurate and topologically consistent lane network representations that perform well at complex intersections and with discontinuous lane markings.
  • Evaluation Metrics: Precision, recall, and F1-scores for both geometric correspondence and semantic classification accuracy. The evaluation methodology will particularly emphasize algorithmic robustness to complex intersection geometries, discontinuous lane representations, and variable road surface conditions.

Prizes

The Top 3 individuals and teams on the leaderboards that provide valid submissions will receive official certificates of achievement. Additionally, we are thankful to our sponsors for providing monetary awards to the winning teams.

Workshop Format

The workshop will implement a hybrid participation model, accommodating both in-person attendance and virtual engagement to maximize accessibility and international participation. The programmatic structure will comprise:

Schedule

Time Event
14:00-14:05 Welcome Introduction
14:05-14:35 Invited Talk (Talk 1)
14:35-15:05 Invited Talk (Talk 2)
15:05-15:35 Invited Talk (Talk 3)
15:35-16:00 Coffee break
16:00-16:10 Awarding Ceremony
16:10-16:30 Winner Talk (Track 1) + Q&A
16:30-16:50 Winner Talk (Track 2) + Q&A
16:50-17:10 Winner Talk (Track 3) + Q&A
17:10-17:30 Panel Discussion
17:30-17:35 Closing Remarks

Organizing Committee

Primary Organizer

Bisheng Yang: Professor at Wuhan University and Director of the State Key Laboratory of Information Engineering in Surveying, Mapping, and Remote Sensing (LIESMARS). His research expertise includes geographic information and point cloud processing. Professor Yang has received numerous national and international academic awards, including the prestigious Carl Pulfrich Award in 2019.

Zhen Dong: Professor and Head of 3S (GNSS/RS/GIS) Integration Department, LIESMARS, Wuhan University. His research interests include 3D computer vision, 3D reconstruction, scene understanding, point cloud processing, and their applications in intelligent transportation systems, digital twin cities, urban sustainable development, and robotics.

Co-Organizers

Xiaoxin Mi: Associate Professor, School of Computer Science and Artificial Intelligence, Wuhan University of Technology. Her research interests lie at the intersection of 3D computer vision and urban understanding, particularly including scene understanding and modeling, point cloud processing, and their applications in intelligent transportation systems (ITS).

Hong Xie: Associate Professor, School of Geodesy and Geomatics, Wuhan University. His research interests include target detection based on image deep learning, point cloud data quality improvement, point cloud information extraction, and model reconstruction.

Jian Zhou: Associate Researcher, LIESMARS, Wuhan University. His research interests include high-definition maps, computer vision, and autonomous vehicles.

Bo Qiu: M.S. Student, LIESMARS, Wuhan University. His research interests include 3D computer vision and their applications in intelligent transportation systems.

Chong Liu: Ph.D. Student, LIESMARS, Wuhan University. His research interests lie in the field of point cloud processing and intelligent transportation systems.

Zhen Cao: Ph.D. Student, LIESMARS, Wuhan University. His research interests lie in the field of 3D computer vision, point cloud completion, and scene understanding.

Yuzhou Zhou: Ph.D. Student, Department of Computer Science, University of Oxford, UK. His research interests lie in the field of 3D computer vision, specifically in 3D scene understanding and point cloud analysis.

Confirmed Speakers

Hong Xie: Associate Professor, School of Geodesy and Geomatics, Wuhan University. Topic: Multi-station Point Cloud Fusion for Complete Scene Representation.

Xiaoxin Mi: Associate Professor, School of Computer Science and Artificial Intelligence, Wuhan University of Technology. Topic: Urban Road Perception and Structural Modeling.

Yuzhou Zhou: Ph.D. Student, Department of Computer Science, University of Oxford, UK. Topic: Street Scene Modeling and Editing.

Broader Impact

The intellectual contributions and methodological frameworks developed through this challenge have the potential to catalyze significant technological and societal advancements across multiple domains:

Ethical Considerations

The datasets utilized in this challenge have been collected and annotated in strict accordance with applicable privacy legislation and regulatory frameworks. All personally identifiable information has been methodically anonymized to ensure the protection of individual privacy rights and community interests. The organizing committee will implement rigorous protocols to ensure that the dataset utilization remains exclusively within the intended research domain of point cloud-based traffic scene understanding.

For more information, please contact the organizing committee at traffic3dchallenge@outlook.com