3D SCENE RECONSTRUCTION FROM RGB IMAGES

Object recognition is significantly improving, allowing us to better understand and extract information from images. This paper presents a novel method for 3D scene reconstruction using a single RGB image, based on a known 3D model database. We use every detected object in an image to further process its pose and a corresponding 3D model by leveraging existing datasets of both 3D models and labeled images to understand and simulate perception correctly. This method produces a clean and lightweight representation of a scene. State-of-the-art research, implementation details, and evaluation results are presented.