1000+ high-resolution indoor scenes with 3D laser scans, panoramic images, 33MP DSLR sequences and iPhone RGBD streams. All registered together, with fine-grained semantic and instance annotations. Includes public semantic and novel view synthesis benchmarks.

Joint 3D object reconstruction and part co-segmentation of identical and near-identical objects using a multiview image transformer. Introduces the 3DTwins dataset with 76k manually annotated object pairs from ScanNet++, achieving 104% IoU improvement over baselines.

Fine-grained object and part-level dense captioning of 3D scenes with Cider score improvement of 17% and 124% over baselines, using semantic and textual consistency losses, Large-scale dataset with 190k object and part-level captions of 34k 3D objects in the ScanNet++ dataset.

Indoor 3D scene generation as sequence of object tokens using Transformers. Generates object location, semantic class, size, rotation and relevant properties by training on the SUNCG dataset. State-of-the-art scene generation validated by a user study and inference speed.

Molecular property prediction using graph neural networks with synthetic coordinates -- distances and angles, as features edge features. State of the art on ZINC and coordinate-free QM9 by incorporating synthetic coordinates in the SMP and DimeNet++ models