Symbiosis-Inspired Knowledge Distillation for Incremental Object Detection

Incremental object detection (IOD) aims to extend detectors to new categories while retaining previously acquired knowledge. Existing methods often adopt a class incremental learning perspective, separating feature spaces to sharpen decision boundaries. However, this separation-oriented paradigm may overlook object symbiosis in detection, where co-occurrence and occlusion introduce spatial and semantic dependencies that benefit from shared representations. Ignoring these dependencies distorts the shared representations, exacerbates confusion between old and new classes, and accelerates catastrophic forgetting. To address this, we propose Symbiosis-Inspired Knowledge Distillation (SIKD), which explicitly leverages object symbiosis at two complementary levels. Spatial Symbiosis Distillation (SpSD) focuses on symbiotic regions where the old model responds with high overlap to objects in the new task. It preserves generalizable old class cues, suppresses class-specific bias and redundancy, and distills the refined evidence to the new model at matched spatial locations with slot-aligned supervision. Semantic Symbiosis Distillation (SeSD) maintains class level structure by forming confidence weighted prototypes for old classes and aligning their inter class soft ranks over the old class logits, which stabilizes the semantic topology during adaptation. Extensive experiments demonstrate the effectiveness and superiority of the proposed method.

Object detectors often need to recognize new object categories after they have already been deployed. In incremental object detection, the model is trained with labels for new categories only, while images may still contain old objects that are no longer annotated. As a result, the model can treat old objects as background or confuse them with newly learned categories. Existing methods usually rely on the old model to find reliable old-object signals, but they often focus on clean predictions that have little overlap with new annotations. Our key observation is that high-overlap responses from the old model are not always noise, since they may reflect useful relationships between objects, including shared context, semantic similarity, and occlusion. We propose Symbiosis-Inspired Knowledge Distillation, which turns these responses into useful training signals. The method refines spatial evidence from these regions and preserves the relative semantic structure among old categories. This helps the detector learn new categories while better retaining old ones. Experiments on standard benchmarks show that our method reduces confusion between old and new categories and improves incremental detection performance.