- Cameras are color coded as follows
- Training, Validation, Test
- The baseline between cameras is 0.2 units.
- Reference camera (12) is shown in dotted circle.
- The azimuth angle theta goes from right to left.
- Elevation angle phi goes from top to bottom.
Coordinate system convention similar to the medieval_port dataset
x-forward, y-up and z-right
Azimuth and elevation angles map to pixel coordinates on the ERP image as follows:
-
at the top, at the bottom. is a vertical rotation with x-z as 0, up direction is positive. -
at the left, at the right. Horizontal rotation, clockwise direction (from x-axis to z-axis, when seen from the top) is positive.
Check the residential data loader and utils for more details. Camera poses (together with RGB images) are provided as .h5 files.
Train/test/val split 8 of the 9 views are used for training and 1 view is held out for testing.
This dataset has 14 scenes, each with 9x9 light-field images.Coordinate system convention similar to the medieval_port dataset
x-forward, y-left and z-up
Azimuth and elevation angles map to pixel coordinates on the ERP image as follows:
-
at the top, at the bottom -
at the left, at the right
There is no rotation between the images in the light field and the baseline between cameras is 0.1 units.
Train/test/val split
Each 9x9 light field is split into training, validation, and test sets, as follows:
# Note that the camera indexing here is row-major order
testing = [4, 20, 22, 24, 36, 38, 42, 44, 56, 58, 60, 76]
trainining = [0, 2, 6, 8, 18, 26, 40, 54, 62, 72, 74, 80]
validation = [1, 5, 7, 37, 41, 43, 71, 77, 79]