The Overwatch Depth Dataset contains pixel-aligned RGB images and dense depth maps captured from
diverse synthetic outdoor environments. Streets, buildings, vegetation, open spaces, and long-range sky
regions provide varied geometry for training and evaluating monocular depth estimation models.
- Paired RGB and depth images are provided at a preview resolution of 1024 × 576.
- Depth is stored on a shared 8-bit dataset scale: black represents nearer geometry and white
represents farther geometry. The images are not independently min–max normalized per frame.
- The scale is relative and its conversion to meters is not available. The dataset is therefore intended
for relative-depth training and scale-aligned evaluation rather than reporting metric errors in meters.
- Spectral color is used only for visualization and does not change the underlying grayscale depth
values.
- Five environments form the training split, while Midtown is reserved for testing.