Depth Maps
Read 16-bit PNG, TIFF, NPY, PFM and EXR depth with percentile windowing, colormaps, a metres readout under the cursor, and unprojection into the 3D viewer.
Potato renders depth maps legibly, reports the distance under the cursor in metres, and — given camera intrinsics — turns a depth map into a point cloud you can put 3D boxes on. Depth is how monocular estimators, stereo rigs and RGB-D sensors deliver 3D, and it is the format most robot datasets ship alongside their video.
Runnable example: examples/spatial/depth-eval/.
Three things make depth different from an image
A depth map opened as an image is a black rectangle. If it were not, you would still be looking at colours with no idea what distance any of them means.
The file does not know its own unit. Millimetres for RealSense, Azure Kinect and NYU-Depth; 1/256 m for the KITTI completion benchmark; metres for anything written from a float array. depth_scale is metres per stored unit and defaults to 0.001. Getting it wrong is silent, because the window rescales and the picture looks the same, which is exactly why the metres readout exists.
Zero is not a distance. It is the near-universal "no return" code. Read as depth, it paints a bright wall across every hole in the sensor's coverage. Potato carries non-measurements as NaN and paints them magenta, a colour in none of the colormaps, so a hole cannot be mistaken for near or far depth. The info line reports what fraction of the map is holes. A stereo rig facing a textureless wall really does return 80% holes, and an annotator who cannot see that will read them as geometry.
The interesting range is almost never the full range. The near/far window defaults to the 2nd and 98th percentile of the valid pixels.
Configuration
instance_display:
fields:
- key: depth
type: depth_map
label: "Depth"Supported formats: 16-bit PNG, 16-bit TIFF, NPY and NPZ, PFM, and EXR.
Windowing is measurable, not cosmetic
The percentile window is the substantive part, and its effect can be stated as a number. A 16-bit scan whose content sits between 1200 and 1800 renders with 4 grey levels of separation out of 255 under a naive 8-bit cast, and 254 under the percentile window. The same applies to scientific TIFF generally. See media ingest.
Unprojection
Given intrinsics, a depth map unprojects into the same 3D viewer the point cloud surface uses, so the same cuboid tools work on it. This is what makes an RGB-D dataset annotatable in 3D without a separate conversion step.