DesktopObjects-360

The dataset introduced in "PointGauss: Point Cloud-Guided Multi-Object Segmentation for Gaussian Splatting" DesktopObjects-360, a benchmark specifically designed for 3D segmentation for radiance field (3DGS and NeRF). The key features include: 1. The dataset provides COLMAP format data, making it directly applicable to 3DGS and NeRF modeling. 2. 3D Instance annotations are available for Gaussian models. 3. Pixel-level instance segmentation annotations are provided for 2D images. 4. The instance IDs for 2D and 3D objects remain consistent across different viewpoints. The dataset contains a total of 3,364 multi-view images with fine-grained 2D and 3D instance annotations. Each scene includes 7--10 object instances (56 annotated 3D instances in total), covering common desktop items under varying layouts and occlusions. To support instance segmentation tasks, we provide 26,042 pixel-accurate 2D instance masks across all scenes, with an average of 465 masks per 3D instance, ensuring dense and consistent multi-view correspondence. The DesktopObjects-360 dataset is organized in a hierarchical directory structure, with a root folder containing six subdirectories (Desk1 through Desk6) and one PointCloud.7z. Each Desk folder follows a consistent organization scheme, as exemplified by Desk1 which contains subfolders for 1. test data (Desk1_test), 2. original images (images), 3. segmentation masks (mask), 4. visualized masks (mask_visualize), 5. pretrained model annotations (annotated\_pretrained_model(2dgs)), 6. and a class label file (class.txt). In addition, we also provide pre-trained 2D Gaussian Splatting (2DGS) models with annotations for 10 scenes from the Nerds360 dataset.