This repository implatement the OS3D for the open-set 3DOR (3D object retrieval) task. The main code is motivated by the Example Code of the SHREC'22 Open-set 3D Object Retrieval. The OS3D is implemented by directly concatenating the features extracted by those modality-specify networks (MVCNN(ResNet18 as the backbone) for the multi-view modality, PointNet for the point cloud modality, and ShapeNets for the voxel modality). Then, the concatenated features are adopted to evaluate the retrieval performance on those unknown categories of the retrieval set on the OS-MN40-core, OS-NTU-core, and OS-ESB-core datasets. Detailed performance can be found in the "Open-set 3DOR Leaderboards".
- Python 3.9
- Pytorch 1.11.0
- Open3D 0.15.2
The definition of mAP, NDCG, ANMRR can refer to this book.