Β One poster has been accepted to SIGGRAPH ASIA
Title: Assessing Inter-Object Reflection Consistency for Photorealistic Virtual 3D Object Insertion
Authors: Jin-Woo Kong (KAIST), GeonU Kim (KAIST), Tae-Hyun Oh (KAIST)
Realistic 3D object insertion requires not only the estimation of global scene illumination but also the modeling of local interactions, such as inter-object reflections. Existing benchmarks, such as GMLight, primarily evaluate global lighting in simplified scenes and fail to capture these local reflections. To address this limitation, we introduce a new benchmark for evaluating inter-object reflection consistency in 3D object insertion. Our benchmark provides a dataset of synthetic scenes containing multiple objects, including paired images before and after insertion, object masks, and camera and object-pose metadata. We demonstrate that existing methods struggle to model reflections from neighboring objects, and show that our benchmark enables better evaluation of local illumination effects.