Depth Holes

Pixels where a depth camera can’t measure distance, usually recorded as zero in the depth map.

Depth holes are pixels in a depth map with no valid measurement; cameras like the RealSense record them as zero, which shows up as black patches. Taking stereo-based depth cameras as an example, Intel RealSense’s official white paper lists several causes: occlusion (the left and right cameras can’t both see the same point), lack of surface texture, repeating patterns causing multiple matches, over- or under-exposure, and objects closer than the minimum working distance; transparent surfaces, specular reflections, and very distant surfaces also often fail to register. Holes directly affect downstream processing: missing chunks in a point cloud cause grasp detection and obstacle avoidance to misjudge, and feeding a raw zero into a network as if it were a real distance makes the network think the object is pressed right up against the camera. Common fixes include post-processing filters (filling holes from neighboring pixels, or from previous frames’ values), learned depth completion, or masking out missing regions during training.

ExamplePhotographing a stainless-steel spoon and a glass on a table with a stereo depth camera: the highlight on the spoon and the whole glass often come back as a patch of zeros in the depth map, so these two objects are nearly missing once converted to a point cloud.

Also called

Missing Depth, Invalid Depth Pixels

Related

Depth Completion · Depth Camera · Active Stereo · Flying Pixels · Occlusion · Transparent & Reflective Object Perception

Sources

RealSense: Depth Post-Processing for Intel RealSense Depth Camera D400 Series
librealsense: Post-processing filters