Meta R-CNN : Towards General Solver for Instance-level Few-shot Learning

Better Class Feature Representation for Few-Shot Object Detection: Feature Aggregation and Feature Space Redistribution

“…(1) LSTD [58] proposes a regularization method based on transfer knowledge and background depression regularizations to enhance the fine-tuning effect; (2) Meta YOLO [20] learns feature representation by reweighting module to reassign feature weights; (3) MetaDet [59] solves the problem of few-shot classification and localization simultaneously through a weight prediction meta-model; (4) CME [21] balances the novel class margins by class margin loss and feature interference; (5) Meta R-CNN [25] obtains the class attention vector through the predictor-head remodeling network (PRN) module to remodel the ROI feature; (6) Viewpoint [26] performs efficient feature similarity calculation through feature subtraction; (7) DCNet [27] introduces adaptive context awareness into the feature aggregation module, to gain better global features and local features; (8) FSCN [60] introduces a novel few-shot classification refinement mechanism to improve the final classification; (9) FsDetView+ISAM+QSAM [61] generated an individual prototype for each sample to extract the unique characteristics of each support sample; (10) CAReD [62] maximizes the inter-class distance and minimizes the intra-class distance through contrastive learning.…”

“…Two-stage object detection methods are more suitable for few-shot object detection tasks because of their characteristics, and methods based on Faster R-CNN usually perform better than one-stage methods [63]. However, the proposed BFR approach still outperforms MetaDet [59], Meta R-CNN [25], and Viewpoint [26], which are based on Faster R-CNN. Meanwhile, BFR performs better than DCNet [27], FSCN [60], CAReD [62], etc.…”

“…Meta YOLO [20] simulates the few-shot scenario on the base classes, enabling it to quickly learn feature representations for novel classes. Meta R-CNN [25] combines meta-learning with Faster RCNN [10] and proposes meta-learning over RoIs. Li et al [21] believed that there is a contradiction between the classification and feature representation in novel classes, which have inconsistent requirements for the distribution of base class features.…”

“…The support images are convolved by the feature extraction module (FE S ), which uses darknet-53 as the backbone to obtain the feature map. In existing studies, global max pooling or global average pooling are mostly used for feature compression [20,21,[25][26][27]. The global max pooling extracts the most representative information in the channel as the local information, and the global average pooling extracts the average information of the channel as the global information.…”

“…It is difficult to obtain class feature prototypes well conforming to the real feature distribution when the number of samples is extremely scarce. Existing methods usually use simple feature compression to encode support set samples [20,21,[25][26][27], and obtain class feature prototypes by averaging encodings [20,21,25,[28][29][30][31]. Therefore, the class prototype embedding is highly sensitive to support set noise and less robust.…”