Open‑world object detection (OWOD) demands that a model not only detect known categories but also discover unknown objects and learn them incrementally once annotations appear. In remote‑sensing images, object categories often exhibit latent hierarchies that Euclidean embeddings fail to capture, limiting unknown recall and incremental performance. This work introduces HyRS‑OWOD, which leverages hyperbolic geometry to represent hierarchical relations and boost OWOD capabilities.
To improve unknown recall, a two‑step discovery pipeline is proposed. The Decoupled Objectness Learning (DOL) module disentangles foreground perception from semantic cues, allowing foreground proposals to be separated from background. Then, Hyperbolic Uncertainty Learning (HUL) uses the radius of hyperbolic embeddings as an uncertainty signal for known‑unknown discrimination.
For incremental learning, a Hyperbolic Metric Learning (HML) strategy is devised to enlarge inter‑class separability in hyperbolic space, facilitating the integration of new categories while mitigating catastrophic forgetting.
Experiments on three remote‑sensing benchmarks show consistent gains in unknown recall and incremental learning over state‑of‑the‑art OWOD approaches.
Review