Object detection based on few-shot learning via instance-level feature correlation and aggregation

Few-Shot Object Detection: A Comprehensive Survey

“…Wang et al [48] (IFC) first used a self-attention module on top of average-and max-pooled query features to separately mine local semantic and detailed texture information. Afterward, with a new feature aggregation mechanism based on a learnable soft-threshold operator [80], redundant information can be shrunk while enhancing feature sensitivity and stability for both novel and base categories.…”

“…Wang et al [48] (IFC), Kobayashi [47] (SPCD), and Huang et al [49] (ARRM) also used a cosine loss, but in ARRM, an additional margin is added to further increase discrimination and reduce misclassification.…”

“…IFC [48] does not build on top of transformers but utilizes an interactive self-attention module to capture the discriminating features from scarce novel categories. Moreover, a novel feature aggregation mechanism is introduced, which aims at shrinking redundant information while enhancing feature sensitivity and stability for both novel and base categories.…”

“…Currently, dual-branch meta learning approaches improve much by using attention for aggregating features of both branches [48], [54], [64]. By aggregating before the RPN [55] or using a proposal-free transformer as detector [64], the issue of missing proposals for novel categories is effectively solved.…”

“…Transfer learning approaches currently improve much by guiding the gradient flow to be able to train as many components of the detector as possible [114]. In both types of approaches, a current trend is the use of metric learning concepts by modifying the loss function to enable better category separation [48]. More trends are highlighted as part of the comparison in Section VIII.…”