R-DiP: Re-ranking Based Diffusion Pre-computation for Image Retrieval

In image retrieval tasks, although efficient methods based on pre-computing information related to retrieval and effective methods utilizing re-ranking have been proposed, developing a method that achieves both efficiency and effectiveness at the same time, remains challenging. To develop an efficient and effective image retrieval method, we propose a simple-yet-effective novel image retrieval framework; R-DiP (Re-ranking based Diffusion Pre-computation). It incorporates an effective re-ranking model into the pre-computation step of an existing efficient method, namely, Offline Diffusion that pre-computes the diffusion process in the offline step and provides a simple linear combination-based retrieval in the online step. Experimental results on standard benchmarks shows that R-DiP performs comparable to the State-Of-The-Art (SOTA) image retrieval method, namely SuperGlobal, while maintaining the efficiency of Offline Diffusion. Notably, in million-scale datasets, R-DiP improves the mAP (mean Average Precision) by about 2.0%, and reduces the speed by about 75% on average, surpassing SOTA methods. These results indicate that R-DiP is a promising solution to the efficiency-effectiveness trade-off in image retrieval, that offers the flexibility to incorporate any advanced re-ranking method in the future.