Incremental Few-Shot Object Detection with High-Efficiency via Expandable-RCNN
Ramakrishnan Raman, Vikram Kumar, Dimple Saini, Dhaval Rabadiya, Smruti Patre, Ramakrishnan Meenakshi · 2024
In the dynamic field of computer vision, the challenge of Few-Shot Object Detection (FSOD) requires robust models that can efficiently learn from a limited number of examples. Our paper introduces the Expandable-RCNN, an innovative approach that integrates incremental learning with FSOD to address inefficiencies and scalability challenges in existing methodologies. This novel architecture avoids catastrophic forgetting by dynamically incorporating new knowledge while preserving previously learned information. We employ a unique incremental learning strategy, enabling the Expandable-RCNN to swiftly adapt to new object classes using minimal examples and maintain high detection accuracy across existing categories. Our evaluation on standard benchmarks demonstrates that Expandable-RCNN surpasses traditional FSOD techniques in detection performance and efficiency, achieving state-of-the-art results. The model’s adaptability and scalability make it exceptionally suitable for real-world applications where data availability frequently changes. Results indicate that Expandable-RCNN significantly improves precision, recall, and F 1 -score, evidencing its potential to reshape FSOD paradigms. These findings highlight the critical importance of incorporating incremental learning capabilities within object detection frameworks, paving the way for future advancements in adaptive, efficient, and scalable detection systems in various dynamic environments.