NYU Depth V2
RGB-D dataset collected with a Kinect sensor, composed of images annotated indoors with depth, useful for 3D vision, semantic segmentation, and robotic tasks.
1449 annotated RGB+ depth images, 407,000 raw images, Parquet format (~3 GB per split)
Apache 2.0
Description
NYU Depth V2 is a multimodal dataset containing RGB images aligned with depth maps recorded via Kinect, in various indoor environments. It includes 1449 pairs of annotated images and over 400,000 unlabeled raw images. Each object in the labeled scenes is categorized with a class and an identifier (e.g. cup1, chair3).
What is this dataset for?
- Train segmentation models or 3D detection in an indoor environment
- Develop perception systems for home robotics or AR/VR glasses
- Studying the reconstruction of 3D scenes from RGB-D images
Can it be enriched or improved?
Yes, you can enrich this dataset with additional annotations (3D boxes, flat surfaces, directions of movement), combine unannotated data with semi-supervised methods, or adapt scenes to localized environments (e.g. European vs American apartments).
🔎 In summary
🧠 Recommended for
- Indoor robotics projects
- Augmented reality applications
- 3D vision research
🔧 Compatible tools
- Open3D
- PyTorch3D
- Detectron2
- Hugging Face Datasets
- OpenCV
💡 Tip
Consider combining this dataset with synthetic data to overcome diversity limitations in real scenes.
Frequently Asked Questions
Does the dataset only contain interior scenes?
Yes, all images come from indoor environments like apartments, classrooms, or offices.
What sensor was used to capture this data?
The images were captured with the Microsoft Kinect sensor, combining RGB image, depth, and accelerometer data.
Can this dataset be used for 3D segmentation tasks?
Yes, it is one of its main uses thanks to the dense annotations per object and the presence of aligned depth.
