GPR1200: A Benchmark for General-Purpose Content-Based Image Retrieval

Even though it has extensively been shown that retrieval specific training of deep neural networks is beneficial for nearest neighbor image search quality, most these models are trained and tested in the domain landmarks images. However, some applications use images from various other domains therefore need a network with good generalization properties - general-purpose CBIR model. To best our knowledge, no testing protocol so far introduced to benchmark respect general quality. After analyzing popular test sets we decided manually curate GPR1200, an easy accessible but challenging dataset broad range categories. This subsequently used evaluate pretrained different architectures on their qualities. We show large-scale pretraining significantly improves performance present experiments how further increase by appropriate fine-tuning. With promising results, hope interest research topic CBIR.