FRDet: Few‐shot object detection via feature reconstruction

State-of-the-art object detection models rely on large-scale datasets for training to achieve good precision. Without sufficient samples, the model can suffer from severe overfitting. Current explorations in few-shot are mainly divided into meta-learning-based methods and fine-tuning-based methods. However, existing do not focus how feature maps should be processed present more accurate regions of interest (RoIs), leading many non-supporting RoIs. These RoIs increase burden subsequent classification even lead misclassification. Additionally, catastrophic forgetting is inevitable both models. Many classify directly low-dimensional spaces due insufficient resources, but this transformation data space confuse some categories To address these problems, Feature Reconstruction Detector (FRDet) proposed, a simple yet effective fine-tune-based approach detection. FRDet includes region proposal network (RPN) based channel attention called Multi-Attention RPN (MARPN) head reconstruction Head (FRHead). MARPN utilizes suppress classes spatial enhance support Attention RPN, resulting fewer Meanwhile, FRHead features reconstruct query RoI through closed-form solution, allowing comprehensive fine-grained comparison. The was validated PASCAL VOC, MS COCO, FSOD, CUB200 achieved better results.