MPF-Net: multi-projection filtering network for few-shot object detection

Deep learning-based object detection has made tremendous progress in the field of intelligent vision systems. However, one of its major complaints is the high demand for large amounts of experimental data. Few-shot object detection (FSOD) aims to identify novel objects with only a few training samples. Existing techniques don’t fully explore the potential mapping relationships between support and query features due to the limitation of global matching contrast. In this work, we propose a multi-projection filtering network (MPF-Net) to exploit feature relevance and aggregate the information between multiple scales, ensuring an optimal global representation. Furthermore, we take the lead in proposing a feature contrast filtering paradigm for the classification and regression subtasks in order to fully utilize fine-grained features for contrastive match. The multi-visual contrast approach motivates our model to gracefully handle a variety of difficult detection challenges such as scale discrepancies, occlusions, and feature confusions. MPF-Net accurately perceives features at various scales and adaptively excavates category information. Extensive experiments on PASCAL VOC and MS COCO datasets have demonstrated that our detectors significantly improve upon baseline detectors, especially for extremely low-shot settings (average accuracy improvement is up to 3.5% in 1-shot scenarios and 2.5% in 2-shot scenarios). In general, we propose a novel strategy to construct the few-shot feature space and achieve remarkable results.