CrossWeave: Cross-Modal Scene and Instance Retrieval from Sparse 2D-3D Alignment

NeurIPS 2026 · Paper (coming soon) · Project page · Code

File	Image encoder	Scene R@1 ScanNet	Scene R@1 3RScan	Instance Top-1 ScanNet	Instance Top-1 3RScan
crossweave_dinov2_vitl14.ckpt	DINOv2 ViT-L/14	60.58	47.13	77.66	73.61
crossweave_tipsv2_l14.ckpt	TIPSv2 ViT-L/14	59.29	43.95	75.22	73.28

Image → point cloud, val split. Checkpoints contain only the trained adapters and SALAD. Installation, data preparation and evaluation are described in the code repository.