UASOL:
A Large-scale High-resolution Stereo Dataset

The dataset is divided in folders. Each folder contains the data for a particular sequence. Inside each folder, it is provided a log file, a manifest file and another folder with the actual RGB color pairs and the corresponding depth maps. The directory tree of the dataset is as follows:

Dataset:

• sequence_1
  • log.txt
  • complete.json
  • Images
    • img_0_depth.png
    • img_Y_depth.png
    • img_left0_color.png
    • img_leftY_color.png
    • img_right0_color.png
    • img_rightY_color.png
• sequence_1_sgbm
  • img_0_offdepth.png
  • img_Y_offdepth.png
• sequence_1_gcnet
  • img_0_offdepth.png
  • img_Y_offdepth.png
• sequence 2
  • ...
• ...

The Log file

The TXT file named "log" (log.txt) stores the camera settings. This data was obtained using the ZED API. A log file is provided for each sequence. The information provided is listed below:
For each camera:

• Optical center along x axis (pixels)
• Optical center along y axis (pixels)
• Focal length along x axis (pixels)
• Focal length along y axis (pixels)
• Vertical field of view after stereo rectification (angle in degrees)
• Horizontal field of view after stereo rectification (angle in degrees)
• Diagonal field of view after stereo rectification (angle in degrees)
• Distortion factor of the right cam before calibration
• Distortion factor of the right cam after calibration

For each sequence:

• Confidence threshold
• Depth min and max range values (millimeters)
• Resolution of the images (pixels)
• Camera FPS
• Frame count

The Manifest File

The "manifest" file (complete.json) packs the core information for each sequence. The information provided is listed below:

• Filename of the left color image
• Filename of the right color image
• Filename of the depth map as provided by the SGM algorithm
• Filename of the depth map provided by the GC-NET method
• Translation matrix (3x1)
• Orientation matrix (3x1)
• M matrix (4x4) which contains the rotation and translation
• Timestamp (ms)

Any question?

Behold the best methods for depth prediction (stereo and monocular):

Method	Mean Depth Error respect to SGM (mm)	Mean Depth Error respect to GC-Net[1] (mm)
Laina et. al[2]	available soon	available soon