Symbiosis-Inspired Knowledge Distillation for Incremental Object Detection

arXiv cs.AI · 18 hours ago

Researchers proposed Symbiosis-Inspired Knowledge Distillation (SIKD), a method for incremental object detection that leverages spatial and semantic relationships between objects to prevent catastrophic forgetting when learning new categories. The approach uses two complementary distillation techniques: Spatial Symbiosis Distillation to preserve old class information in regions with high model overlap, and Semantic Symbiosis Distillation to maintain class-level structure through confidence-weighted prototypes. The method addresses the limitation of separation-oriented approaches that ignore object co-occurrence and occlusion dependencies, demonstrating improvements in retaining previously learned knowledge while adapting to new object categories.