CFENet: Context-aware Feature Enhancement Network for efficient few-shot object counting

Abstract

Few-shot object counting (FSOC) is designed to estimate the number of objects in any category given a query image and several bounding boxes. Existing methods usually ignore shape information when extracting the appearance of exemplars from query images, resulting in reduced object localization accuracy and count estimates. Meanwhile, these methods also utilize a fixed inner product or convolution for similarity matching, which may introduce background interference and limit the matching of objects with significant intra-class differences. To address the above challenges, we propose a Context-aware Feature Enhancement Network (CFENet) for FSOC. Specifically, our network comprises three main modules: Hierarchical Perception Joint Enhancement Module (HPJEM), Learnable Similarity Matcher (LSM), and Feature Fusion Module (FFM). Firstly, HPJEM performs feature enhancement on the scale transformations of query images and the shapes of exemplars, improving the network’s ability to recognize dense objects. Secondly, LSM utilizes learnable dilated convolutions and linear layers to expand the similarity metric of a fixed inner product, obtaining similarity maps. Then convolution with a given kernel is performed on the similarity maps to get the weighted features. Finally, FFM further fuses weighted features with multi-scale features obtained by HPJEM. We conduct extensive experiments on the specialized few-shot dataset FSC-147 and the subsets Val-COCO and Test-COCO of the COCO dataset. Experimental results validate the effectiveness of our method and show competitive performance. To further verify the generalization of CFENet, we also conduct experiments on the car dataset CARPK.