Efficient Object Placement via FTOPNet

Abstract

Image composition involves the placement of foreground objects at an appropriate scale within a background image to create a visually realistic composite image. However, manual operations for this task are time-consuming and labor-intensive. In this study, we propose an efficient method for foreground object placement, comprising a background feature extraction module (BFEM) designed for background images and a foreground–background cross-attention feature fusion module (FBCAFFM). The BFEM is capable of extracting precise and comprehensive information from the background image. The fused features enable the network to learn additional information related to foreground–background matching, aiding in the prediction of foreground object placement and size. Our experiments are conducted using the publicly available object placement assessment (OPA) dataset. Both quantitative and visual results demonstrate that FTOPNet effectively performs the foreground object placement task and offers a practical solution for image composition tasks.