The Second Challenge on Cross-Domain Few-Shot Object Detection at NTIRE 2026: Methods and Results
Summary: arXiv:2604.11998v1 Announce Type: cross
Cross-domain few-shot object detection (CD-FSOD) presents significant challenges for current object detection systems and few-shot learning methodologies, especially when it comes to generalizing across various domains. In response to this pressing issue, the NTIRE 2026 conference hosted its second CD-FSOD Challenge, aimed at systematically assessing and fostering advancements in detecting objects within unseen target domains while operating under limited annotation conditions.
The challenge attracted remarkable interest from the research community, with a total of 128 participants registering and submitting a cumulative 696 entries. Out of these, 31 teams actively engaged in the competition, and 19 of those teams submitted valid final results. This high level of participation underscores the importance of CD-FSOD and the community’s commitment to addressing its inherent difficulties.
Challenge Overview
The NTIRE 2026 CD-FSOD Challenge was structured to evaluate the effectiveness of various strategies for few-shot object detection across different domains. Participants were encouraged to introduce innovative solutions that could potentially push the performance boundaries in both open-source and closed-source tracks. The challenge aimed to highlight the importance of adaptability and efficiency in detecting objects with minimal labeled data.
Participant Contributions
Teams participating in the challenge employed a wide array of methodologies to tackle the complexities of CD-FSOD. Below are some common strategies observed:
- Meta-Learning Approaches: Many teams utilized meta-learning frameworks to enable their models to quickly adapt to new domains using very few examples.
- Transfer Learning Techniques: Participants leveraged transfer learning to enhance model performance by pre-training on large datasets before fine-tuning on specific tasks.
- Data Augmentation: Several teams emphasized the use of data augmentation techniques to artificially increase the diversity of training samples, thereby improving model robustness.
- Domain Adaptation Methods: Various domain adaptation strategies were explored to minimize the gap between source and target domains, facilitating better generalization.
Results Analysis
The final results of the NTIRE 2026 CD-FSOD Challenge revealed significant advancements in the field. The submitted approaches showcased a range of innovative techniques, reflecting the diverse perspectives and expertise of the participating teams. A detailed analysis of the results indicates:
- Teams that effectively combined multiple strategies tended to outperform those relying on single-method approaches.
- Models exhibiting strong performance in open-source tracks often utilized collaborative frameworks that encouraged knowledge sharing among participants.
- The competition’s results indicate a clear trend towards improved accuracy and efficiency in CD-FSOD, suggesting that the community is making substantial progress in overcoming the challenges posed by this complex task.
Conclusion
The NTIRE 2026 CD-FSOD Challenge successfully highlighted the ongoing efforts within the research community to advance the capabilities of object detection systems under limited conditions. With 19 participating teams submitting valid results, the challenge not only provided a platform for showcasing innovative methods but also fostered collaboration and knowledge exchange among researchers. For more information and access to the challenge codes, visit this link.