An AI method for rebuilding 3D scenes from incomplete point-cloud observations reports low collision rates while keeping completed geometry aligned with its surroundings. In the first reported benchmark block, the COOL row recorded ColO 0.002 and ColS 0.003. In the second, it recorded ColO 0.027 and ColS 0.000. ColO is the paper's object-object collision rate, while ColS is its object-scene collision rate.
The work tackles a specific reconstruction problem: whether a partial observation of an object can be completed into geometry that remains aligned to the scene's metric coordinates without colliding with surrounding geometry. COOL uses instance geometry for metric alignment and background geometry as local scene context during completion.
How the completion is guided
During inference, COOL adds an explicit collision loss and uses joint guidance plus resampling to escape local optima. Collision handling is therefore part of the generation procedure described by the authors.
The benchmark work uses synthetic 3D-Front indoor scenes and real-world Scan2CAD RGB-D and CAD scenes. The reported 3D-Front split contains 8,000 training scenes and 1,000 test scenes. For Scan2CAD, the paper reports disjoint splits, 10 selected views per instance, and 90,000 training samples from 1,400 scenes.
The reported benchmark scores
On the first reported quantitative block, COOL's row shows CDS 0.001, FSS 0.989, CDO 0.011, FSO 0.904 and IoUB 0.895. The same row reports ColO 0.002 and ColS 0.003. The figures cover the paper's reconstruction, layout and collision measures.
On the other reported quantitative block, the row shows CDS 0.015, FSS 0.894, CDO 0.036, FSO 0.745 and IoUB 0.641. Its collision figures are ColO 0.027 and ColS 0.000.
These are descriptive benchmark comparisons, not inferential estimates. The supplied analysis reports no confidence intervals or formal hypothesis tests for the reported differences, so the numbers show how the listed rows performed in these experiments without a quantified uncertainty range.
A component comparison examined the collision-handling design. In a single-sample ablation, the collision-guided row reported ColO 0.095, compared with 0.203 without guidance, while ColS was 0.003 versus 0.015. Object-level and spatial-layout metrics remained close in that comparison.
In the comparison labeled zero-shot, the row marked Ours(zs) reports CDS 0.029, FSS 0.841, CDO 0.068, FSO 0.587, IoUB 0.549, ColO 0.041 and ColS 0.000.
Tests of masks and multiple views
The mask-source test used two protocols. The default protocol uses ground-truth instance masks, while a perception-noise protocol substitutes masks predicted by segmentation models without changing the remaining process. With SAM masks, the reported values were CDS 0.010, FSS 0.985, CDO 0.024, FSO 0.810 and IoUB 0.727. With PanoRecon, they were 0.002, 0.994, 0.016, 0.810 and 0.766; with ground-truth masks, 0.001, 0.995, 0.015, 0.816 and 0.767.
The paper also reports a multi-view study comparing inputs from 1, 3 and 5 frames, along with reconstructed-mesh point clouds. Performance was reported to rise as more views were integrated, but the supplied results do not give a numerical effect size for that pattern. A qualitative ScanNet case study describes integrating complete object geometries into incomplete scene reconstructions as a real-to-sim use case.
Where the evidence stops
The paper notes that geometry-conditioned generation may lose fine-scale shape detail compared with image-conditioned generators. It also says computing limits kept training to relatively small datasets, so broader generalization was not fully explored.
COOL's reported figures support a narrower conclusion: in the tested settings, the framework produced scene-aligned completions and recorded lower collision rates in the comparisons, but the results do not settle how it will perform on larger or more varied data. The work is an arXiv version 1 preprint dated 29 September 2026.
Paper data and sources
Original title: Collision-Aware and Observation-Aligned Object-Centric Scene Reconstruction from Point Cloud
Authors: Yuxuan Xie, Xuan Yu, Rong Xiong, Yue Wang
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-29
DOI: Not available
Original paper · Full text