OWOD-FSL: Open-World Object Detection Via Few-Shot Learning and Dynamic Prototypes
Zhiwei Li, Zhiyu Zhang, Yang Zhou, Jianping Li, Tianyu Shen, Li Wang, Fengli Lu, Huaping Liu, Kunfeng Wang
AI summary
Problem
Current open-world object detection methods are limited by fixed-dimensional classification heads that prevent true incremental learning and rely heavily on extensive annotated data, hindering adaptability in few-shot real-world scenarios.
Approach
The proposed framework integrates a dynamically expandable prototype classification head with a biologically-inspired dual-phase learning strategy that generates initial prototypes offline and refines them through incremental optimization.
Key results
- Dynamic prototype classification head replaces fixed classifiers for scalable class expansion
- Dual-phase learning strategy combines offline prototype generation with incremental refinement
- State-of-the-art performance on M-OWODB benchmark with 51% mAP and 21.7% U-Recall
- Active sample selection and exemplar replay mitigate catastrophic forgetting
Why it matters
Enables real-world vision systems to continuously learn and adapt to novel objects with minimal data, crucial for autonomous driving and robotics.
Abstract
Open-World Object Detection (OWOD) presents a critical challenge for modern computer vision systems: detecting known classes, identifying unknown objects, and incremen- tally learning to recognize them over time. However, current approaches have two fundamental limitations: (1) the fixed- dimensional classification head inherently restricts incremental learning capabilities, and (2) heavy reliance on extensive an- notated data hinders adaptability in few-shot settings. To ad- dress these limitations, we propose OWOD-FSL that integrates dynamic prototype classification head with few-shot learning. At the core of our approach are two major contributions: a dynamic prototype classification head that supplants traditional fixed classifiers with an expandable prototype classifier for scalable class expansion, and a biologically-inspired bi-phase learning strategy that integrates offline prototype generation with incremental learning refinement. Comprehensive exper- iments on M-OWODB benchmark shows that OWOD-FSL achieves state-of-the-art performance in both unknown class recall (U-Recall) and known class mAP, significantly outper- forming existing methods.