Depth Restoration Occlusionless Temporal
Dataset
Daniel Rotman, Guy Gilboa
Electrical Engineering Department,
Technion – Israel Institute of Technology, Haifa, Israel
DROT is a depth dataset created to test depth restoration, rectification and upsampling methods
Dataset details:
- Real sensor input from Kinect 1, Kinect 2 and RealSense R200 sensors.
- RGB and depth images registered pixel-to-pixel versus high quality ground truth depth.
- For each sensor there are two viewpoints:
- RGB sensor with the registered depth image. This allows testing upsampling methods when the RGB image is substantially larger.
- IR sensor with the registered RGB image for straightforward depth restoration.
- The dataset consists five multiple frame videos, with varying types of object motion.
Vid 1 | Vid 2 | Vid 3 | Vid 4 | Vid 5 |
30 frames
Motion: turn, parallel |
21 frames
Motion: parallel |
30 frames
Motion: turn |
20 frames
Motion: parallel |
11 frames
Motion: turn, parallel, perpendicular |
Vid1_sensor_data | Vid2_sensor_data | Vid3_sensor_data | Vid4_sensor_data | Vid5_sensor_data |
Vid1_ground_truth | Vid2_ground_truth | Vid3_ground_truth | Vid4_ground_truth | Vid5_ground_truth |
Note: Depth PNG images may look completely black when opened with a simple image viewing program due to the 16bit encoding.
If you use this dataset in your work, please cite the following publication:
- Rotman and G. Gilboa, “A depth restoration occlusionless temporal dataset,” in International Conference on 3D Vision (3DV). IEEE, 2016.