Multi-Object 2DGS: Occlusion-Robust Background Removal for Multiple Object Models

Multi-Object 2DGS: Occlusion-Robust Background Removal for Multiple Object Models
Marcel Rogge, Didier Stricker
In: International Conference on Pattern Recognition Applications and Methods 2025. International Conference on Pattern Recognition Applications and Methods (ICPRAM-2025), February 23-25, Porto, Portugal, Pages 151-178, LNCS, Vol. 16555, ISBN 978-3-032-28661-1, Springer, 2026.

Abstract:

Most Gaussian Splatting methods focus on the reconstruction of entire scenes and lack the ability to focus the reconstruction on specific objects. This makes them computationally expensive and requires labor-intensive post-processing for object-centric applications. While a recent method makes object-centric reconstruction possible, it does not perform well for datasets with strong inter-object occlusions. Additionally, it does not support the reconstruction of multiple objects within a single scene without cumbersome sequential processing. We propose a substantial extension that enables object-centric reconstruction of multiple objects in parallel by leveraging instance segmentation. We also propose an improved loss term that shows resilience against inter-object occlusions. Our proposed method reconstructs multiple object models as object-centric Gaussian representations, which can be exported as meshes. Our improvements increase quality by up to 88% while reducing training time by up to 29% on scenes with a large number of objects and many inter-object occlusions.