Cross-Domain Few-Shot Classification of Tire Pattern Images

Background

As urbanization progressively advances and the number of vehicles steadily increases, tire pattern classification has gained substantial applications in public safety domains, specifically in areas like traffic accident processing and criminal investigations. However, practical implementations often encounter challenges due to the limited tire pattern samples collected on-site, cross-domain issues between tire surface patterns and indentation images, and problems related to unclear or incomplete indentation images. Consequently, the cross-domain few-shot classification of tire pattern images emerges as a critical challenge, deserving of thorough investigation.

Data

The dataset for the challenge is based on CIIP tire pattern image database, with a total of 75 classes. Each class contains tire surface pattern images with different proportions and angles, and tire indentation images in practical application scenarios.

Task

In the challenge, out of the 75 tire pattern classes, data from 49 classes will be released for training and validation purposes, including tire surface pattern images, tire indentation images, and their corresponding class labels. The remaining 26 classes are designated as the test set. We design 2000 5-way 5-shot/1-shot cross-domain classification tasks. For each task, the support set comprises 5/1 randomly sampled tire surface pattern images from 5 random classes within the test set, while the query set encompasses a total of 10 test tire indentation images from the same 5 classes. Participants need to predict the category of samples in the query set based on the given support set. It's important to note that the class label space of the query set and the support set remains consistent but differs from that of the training set.

Performance Evaluation

Classification accuracy (%) of 5-way 5-shot/1-shot tasks.

Requirements

  • Participants can be college students, graduate researchers or professionals.
  • 1-6 team members and 1-2 advisers in each group.
  • Participants need to submit testing results that conform to the required format and algorithm descriptions.
  • The winning teams are expected to revise their algorithm descriptions into conference papers according to the conference template format, which can be included in the conference proceedings.
  • Each winning team must present the winning work in a special session in the conference via a standard registration process.

Milestones of the Challenge

  • May 10-June 16, 2024 Registration and Data Release.
  • June 1-Sep. 10, 2024 Teamwork Phase.
  • Sep. 10, 2024 Submission Deadline.
  • Sep. 11-Sep. 15, 2024 Performance Evaluation and Ranking.
  • Sep. 15, 2024 Winner Notification.
  • Sep. 15-Sep. 30, 2024 Camera-Ready Submission by Winning Teams.

Prizes

One 1st prize, two 2nd prize, 3 third prize. Each winning team will receive a certification as well as prize in cash, and will be asked to present their work. Winners will be announced in the conference social event.

Organizers

  • Xi'an University of Posts and Telecommunications, Xi’an, China
  • Nanyang Technological University, Singapore

Challenge Organizing Chairs

  • -Ying Liu, Xi'an University of Posts and Telecommunications, Xi’an, China
  • Kezhi Mao, Nanyang Technological University, Singapore

Links

  • Registration form and Tire pattern image data of the Multimedia Grand Challenge at ICARCV2024 can be downloaded from the following website: http://www.xuptciip.com.cn/show.html?news-ICARCV2024
  • Data will be provided by Center for Image and Information Processing, Xi'an University of Posts and Telecommunications, Xi'an, China, 710121.

We look forward to meeting you at the Grand Challenge!

For any inquiry, please email us at: ciip_xupt@xupt.edu.cn