Answers are grounded in this paper and cite their sources.
Few-shot learning raises problems like supervisory collapse due to class imbalance and pseudo-labels unreliability, which seriously hinder the performance improvement of few-shot object detection. To solve the above problems, the paper proposed a model of few-shot object detection based on graph interaction multivariate prediction fusion. Firstly, the feature extraction network with a semantic attention mechanism is designed to adaptively adjust the new class features so that to obtain high-quality new classes pseudo-labels. Second, visual features are converted into graphical features by constructing graph interaction mechanisms using implicit relationships between objects to optimize and improve the accuracy of pseudo-labels. Finally, the multi-classifier object detection mechanism is used to achieve accurate classification and localization of few-shot objects. The simulation was designed to validate the performance of the proposed model in the MS-COCO dataset and PASCAL VOC dataset and compare it with existing state-of-the-art and representative models. The experiments prove that the detection results of the proposed model in this paper outperform other models in terms of visual effects and objective data.
No takes yet. Share an insight, caveat, or question.
Meng et al. (2023) studied this question.