SLFD: Spatial Location-Aware Feature Distillation for Incremental Object Detection

Yanyan Shi, Yang Chen, Xinye Yu, Dianxi Shi, Xiaokai Xia

Abstract: In practical applications, object detection models often encounter new classes of interest after initial training. Traditional approaches require retraining from scratch using both existing and new class data—an infeasible process when original training images are inaccessible. Fine-tuning the model solely on new data typically leads to catastrophic forgetting, where performance on previously learned classes degrades severely. To address this challenge, we propose a novel Spatial Location-Aware Feature Distillation (SLFD) framework for Incremental Object Detection (IOD), which redefines distillation by focusing on adaptive spatial feature selection rather than global mimicry. Our method integrates three complementary components to mitigate forgetting while enhancing new-class adaptation: 1) Spatial Transformation Feature Distillation (STFD), which identifies and transfers region-specific features from the old model, ensuring the incremental model retains spatial awareness of object locations; 2) Spatial Attention Feature Distillation (SAFD), which preserves pixel-level knowledge by emphasizing salient spatial features; and 3) Meta-Learning Optimization Integration, which leverages gradient preprocessing via meta-learning to enable rapid new-class plasticity without sacrificing prior knowledge. Extensive experiments on benchmark datasets under both single-step and multi-step incremental learning scenarios demonstrate that SLFD outperforms state-of-the-art methods, particularly in mitigating catastrophic forgetting in multi-step settings. Furthermore, evaluations in an autonomous driving scenario confirm SLFD’s practical effectiveness for real-world applications, as it achieves robust detection of new objects while preserving prior knowledge.

External IDs:dblp:journals/access/ShiCYSX25