The PBVS 2025 Multi-modal Aerial View Image Challenge – T invites researchers to advance the field of multi-modal image translation. This challenge focuses on developing high-quality image translation methods using a unique dataset of spatially aligned SAR-EO, EO-IR, and SAR-IR pairs. Participants will address the challenge of conditioned image generation, with results evaluated based on fidelity and perceptual similarity metrics.

Challenge Overview

The goal of the Translation track is to design a solution capable of producing high-quality and high-fidelity multi-modal image translations. Participants will leverage spatially aligned multi-modal data, with temporal alignment provided where possible. Evaluations will utilize established metrics, including L2 Norm, Frechet Inception Distance (FID), and Learned Perceptual Image Patch Similarity (LPIPS). Creativity and technical innovation are strongly encouraged, with top solutions judged on accuracy, novelty, and reproducibility.

Key Dates

• 2025.01.19: Release of training data (inputs/outputs) and validation data (inputs only).
• 2025.01.21: Validation server opens.
• 2025.02.21: Final test data release (inputs only).
• 2025.03.02: Deadline for test results, fact sheets, and code submissions.
• 2025.03.04: Preliminary results and paper submission deadline.
• 2025.06.11: PBVS Workshop at CVPR 2025, results, and award ceremony.

Awards and Opportunities

Top participants will receive awards for their solutions and will be invited to present their work at the 21st IEEE Workshop on Perception Beyond the Visible Spectrum (PBVS), held in conjunction with CVPR 2025. Winning teams are also encouraged to submit their methods as papers for workshop presentation.

Evaluation Metrics

• L2 Norm: Measures pixel-level accuracy.
• Frechet Inception Distance (FID): Assesses image quality and distribution similarity.
• LPIPS: Evaluates perceptual similarity of generated images.

Submission Guidelines

Participants are required to submit:

1. Results of their translation methods for evaluation.
2. Fact sheets detailing their approach and methodology.
3. Code and executables to ensure reproducibility.

All submissions must adhere to the guidelines provided on the competition page.

Resources and Support

• Scripts for reproducibility and evaluation will be provided.
• For questions, participants can use the forum on the competition page or contact the organizers directly at mavoc.pbvs@gmail.com (Justice Wheelwright).

For full details, visit the competition page: PBVS 2025 Challenge – T (Translation).

This challenge is an opportunity to contribute to the advancement of multi-modal image translation methods and gain recognition at a premier conference in computer vision. We look forward to your participation.