Self-similarity Enhancement Via Shape-embedding For Few-shot Counting
Guiyuan Xie, Chensheng Yi, Jin Zhan, Weijian Li, JiaCheng Qiu, Weili Tian, Zhaokang Guan · 2024
This study introduces a pioneering framework termed SEcount, which amins to address the challenge of few-shot object counting across diverse categories.While traditional object counting approaches rely heavily on supervised learning and extensive manual annotations, SEcount employs a Class-Agnostic Counting strategy enhanced by Vision Transformers (ViT).Unlike prior methods, which struggle with feature loss in dense or occluded scenarios, SEcount leverages the strengths of ViT model to effectively identify self-similarities among internal regions of the image and between the image and a few example samples.Furthermore, SEcount introduces a Shape Enhancement Module that restores critical boundary and semantic features during patch-based processing.These capabilities are particularly advantageous for class-agnostic counting tasks, in which the model needs to accurately count objects of any category using minimal training data, Our experiments on the FSC-147 dataset demonstrate that SEcount outperforms recent methods, with a relative reduction of 15.1% in Mean Absolute Error (MAE) on the test set and a 12.3% improvement in MAE on the validation set.