Combination of Global Maximum Pooling and Local Average Pooling for Unsupervised Fine-Grained Image Retrieval
Chang-Hsing Lee, Jau-Ling Shih, Wen-Li Su, Cheng‐Chang Lien, Chin Chuan Han · 2024
In this paper, a novel pooling approach which combines global maximum pooling (GMP) and local average pooling (LAP), called GMP-LAP, is designed to extract global and local CNN features for unsupervised fine-grained image retrieval (FGIR), without any fine-tuning or re-training of the CNN backbone model.First, each image is inputted into a pre-trained CNN model to get a number of feature maps in the last convolutional layer.GMP is first used to extract the global feature from each feature map.Meanwhile, LAP is proposed to obtain some local features from selected salient regions.GMP and LAP are then combined for unsupervised FGIR.As a result, unsupervised FGIR is realized since no manual labeling or re-training/fine-tuning of the CNN model is involved.To evaluate the performance of the proposed GMP-LAP approach, we have conducted experiments on six datasets, including the Stanford Cars, Stanford Aircrafts, CUB-200-2011, Oxford Flowers 102, Stanford Dogs, and Oxford-IIIT Pets, for unsupervised FGIR.Compared with the most widely used global pooling approaches, such as global average pooling (GAP) and GMP, the experimental results have shown that GMP-LAP can improve the retrieval performance in terms of mean average precision (MAP).