New Lookalike3D method detects identical objects in 3D scenes with 104% IoU improvement using multiview transformers

Lookalike3D introduces a multiview image transformer that classifies object pairs as identical, similar, or different in indoor scenes

3DTwins dataset contains 76k manually annotated object pairs based on ScanNet++ to support the lookalike detection task

Method leverages semantic priors from large image foundation models to distinguish between repeated objects in real-world scenes

Achieves 104% IoU improvement over baseline methods and enhances downstream 3D object understanding tasks