| Recent advances in image compositing have enabled realistic object placement using context-aware models such as TopNet and PlaceNet. However, the majority of these techniques do not explicitly handle aesthetic style compatibility and depth-based geometric consistency. As a result, the inserted objects may not visually match the scene. In this research, a novel framework called INSPIRED is presented for photorealistic object placement in an indoor scene from a single 2D image. It is based on monocular depth estimation and semantic segmentation for correct scale and placement on any identified valid surfaces. To maintain aesthetic harmony, fine-tuned deep learning classifiers assess the style compatibility between the target environment and furniture object, thus avoiding any visually incompatible placements such as vintage furniture in a contemporary environment. Object extraction is performed for accurate object boundary determination and state-of-the-art inpainting techniques are used for removing existing objects and background reconstruction. The framework supports both automatic and user-guided placement and is currently implemented for living room environments. Experimental results demonstrate that combining spatial reasoning with style constraints leads to more geometrically accurate and aesthetically coherent object insertions compared to style-agnostic compositing approaches. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.