The Second Challenge on Cross-Domain Few-Shot Object Detection at NTIRE 2026: Methods and Results
The NTIRE 2026 challenge drew 128 participants and 696 submissions to tackle a core computer vision problem.
A consortium of 74 researchers, led by Xingyu Qiu and Yuqian Fu, has published the official report for the second Cross-Domain Few-Shot Object Detection (CD-FSOD) Challenge held at the NTIRE 2026 workshop. The challenge tackled a critical bottleneck in practical computer vision: training an AI model to detect objects in a completely new visual domain using only a handful of labeled examples. This problem, known as cross-domain few-shot learning, is essential for deploying robust vision systems without costly data collection and annotation for every new environment.
The challenge proved to be a significant draw for the research community, attracting 128 registered participants who made a total of 696 submissions. Of these, 31 teams actively competed, with 19 submitting valid final results across both open-source and closed-source tracks. The report details the wide range of innovative strategies explored by participants, which collectively pushed the performance frontier for this difficult task. The high participation underscores the field's recognition of CD-FSOD as a major hurdle for real-world applications like autonomous driving and robotics, where models must adapt quickly to unseen conditions.
- Organized by a 74-author consortium for the NTIRE 2026 workshop, benchmarking progress on a key computer vision problem.
- Achieved high community engagement with 128 registered participants, 31 active teams, and 696 total submissions.
- Focused on Cross-Domain Few-Shot Object Detection (CD-FSOD), a critical capability for adapting AI models to new environments with minimal data.
Why It Matters
Advances in CD-FSOD are crucial for deploying adaptable, data-efficient vision systems in real-world scenarios like robotics and autonomous vehicles.