Improved Channel-Wise Semantic Alignment for Few-Shot Object Detection

Few-shot Object Detection (FSOD), to detect new classes using a small amount of data with bounding box annotations, has recently attracted great research interest in the community. However, limited data during network training often results in weak features, which significantly restricts network performance improvement. To tackle this problem, researchers reweight query features using support features to effectively detect query objects. However, during this process, the quality of support features impacts query features. Additionally, the variations in the scales and appearance of the same object and objects from other classes lead to misaligned semantic information in support features and query features. To address these problems, we propose a novel Channel-wise Semantic Alignment Network (CSANet) to establish reliable and concise connections. CSANet reweights support features using global contextual information to suppress redundant information, enhancing the quality of support features. By analyzing sparse channel relations, semantic alignment between query and support features is achieved, resulting in more robust query features. Our network shows superior performance on the PASCAL VOC and MS COCO datasets.