1,449 densely labeled pairs of aligned RGB and depth images representing 464 diverse indoor scenes. This toolbox facilitates Python-based processing of the dataset's semantic labels, depth maps, and raw video frames originally captured via Microsoft Kinect.
Use Cases
- Train monocular depth estimation models using the aligned RGB and depth map pairs.
- Develop semantic segmentation architectures using the 894 object categories found in the dense labels.
- Implement 3D scene reconstruction pipelines utilizing the camera intrinsics and spatial depth data.
- Benchmark surface normal estimation techniques using the pre-computed surface normal maps.
Strengths
- 1,449 labeled pairs of RGB and depth images with per-pixel semantic annotations.
- 407,024 unlabeled frames across 464 indoor scenes for self-supervised learning.
- Includes camera intrinsic parameters and surface normal maps for 3D geometric analysis.
- Provides access to 894 distinct object categories within the dense labels.