Speaker
Description
AI models are becoming increasingly important in forestry operations, enabling robotic systems to automatically detect and classify forest features such as stems, logs, rocks, and roads from vehicle-mounted images. This automation enhances operational efficiency, reduces human error, and supports sustainable management through real-time monitoring and analysis. However, the development of such models is constrained by the limited availability of data, which is frequently incomplete or partially labeled, thereby limiting model generalization and adaptability.
To address these challenges, we propose an incremental learning strategy for object detection using the transformer-based detection model DETR, which allows the number of classes to be expanded over successive training phases. By fine-tuning DETR on new forestry classes while retaining knowledge of previously learned categories, our approach enables continuous adaptation to evolving datasets without requiring full retraining. The DETR model leverages the capabilities of state-of-the-art foundation models by using a fixed backbone, while transfer learning is applied to the detection head. This strategy allows for rapid adaptation to new datasets with minimal labeled data, maintaining high detection performance across multiple forestry classes.
Overall, this approach demonstrates the potential of combining transformer-based detection models with incremental learning to create flexible, efficient, and scalable solutions for automated forestry feature detection.
| Keywords | object-detection; deep-learning; incremental-learning; detection-transformer |
|---|
