An Incremental Few Shot Approach for Litter Detection

Litter detection currently lacks large-scale public datasets. Previous work faces difficulties in deployment. In this paper, we introduce an incremental few-shot learning approach for litter detection. We convert multi-class classification into binary classification, enabling incremental learning without the need of fine-tuning. Feature maps are efficiently extracted through the proposed visual attention mechanism. We reduce computational complexity by building class prototypes to calculate similarity. Our model is of great generalizability. Comparison with existing few-shot incremental learning work on multiple datasets shows that our method outperforms previous work in detecting both base and novel classes by a significant margin, and requires much less inference time.