SSCA-net: Few-shot Detection without Fine-tuning for Online Garbage Detection

Few-shot object detection has garnered notable interest due to the long-tail distribution of real-world data and the recent impact of data scarcity on detector performance. Currently, it is applied in various fields such as autonomous driving, medical image detection, remote sensing image analysis, and agricultural pest detection. However, effective research on online detection for low-power waste robots is still lacking. The fine-tuning stage required by existing methods is time-consuming and challenging to deploy on low-power waste robots, especially with the ever-increasing variety of new categories. To address this issue, we introduce a new network architecture, "SSCA-net," designed for online detection without fine-tuning. Specifically, we propose a support-guided network within the existing Faster R-CNN detector framework. Our network includes a Cross-Attention Region Proposal Network (CA-RPN) for generating object proposals, which boosts the recall rate for candidate bounding boxes of new classes, and a Support Class Feature (SFF) network that captures more representative prototypes of the support classes. In addition, we incorporate SimAM global attention to improve the quality of feature maps. To demonstrate the viability of our proposed method, we conducted extensive experiments on the VOC2007 dataset and a dedicated waste dataset. SSCA-net attains comparable or even batter results then exhaustive fine-tuning methods, achieve a 10-20% improvement over the baseline without fine-tuning.