Trusted Scale-Aware Convolution Network for Few-shot Object Counting
Chenliang Sun, Tong Jin, Gongxuan Zhang · 2024
This study focuses on a more general counting approach aimed at predicting the quantity and distribution of objects across arbitrary categories. Existing methods typically rely on designing similarity matching rules between samples and query images, overlooking robustness in feature extraction. To address this, we propose a novel method called Scale Aware Counting Network (SACNet), a framework for density estimation based on the pre-trained ResNet-50 model. By extracting image features and processing sample bounding box information, we obtain feature maps and compute cosine similarity, ultimately yielding predicted density maps. Additionally, in specific class object counting tasks, ground truth density maps are generated by convolving Gaussian kernels with annotated point maps, while the quality of predicted density maps is typically evaluated using L2 loss. We introduce a generalized loss method to address the imbalance in optimal transport problems and quantify the distance between predicted density maps and point maps. Extensive experiments demonstrate significant improvements achieved by our model, achieving state-of-the-art performance in category-agnostic counting.