Depth bleeding

Depth bleeding is a depth-map artifact in which an estimated distance incorrectly spreads across an object boundary. For example, pixels belonging to a nearby person may receive the farther depth of the background—or the person’s depth may extend into the background. It is especially important in depth estimation, where each image pixel should represent the distance to its corresponding surface.

Link to this sectionHow Depth Bleeding Appears#

A correct depth map contains sharp transitions where one surface occludes another. Depth bleeding replaces that transition with misplaced, smoothed, or mixed values. The CVPR study on segmentation and depth constraints describes a common form in which foreground depth expands outward into background regions. (openaccess.thecvf.com)

Typical causes include:

• Low-resolution prediction followed by upsampling
• Convolutional smoothing that removes fine edge information
• Loss functions dominated by large, smooth regions rather than narrow boundaries
• Occlusions and ambiguous stereo correspondences
• Misalignment between RGB images and depth labels
• Sensor measurements that mix foreground and background surfaces

In block-based stereo vision, matching windows can cover both sides of an edge. The OpenCV stereo depth guide notes that this can produce speckles near object boundaries, while its disparity post-filtering documentation explains that errors often concentrate near depth discontinuities and half-occlusions. (docs.opencv.org)

Link to this sectionRelated Terms and Differences#

Depth bleeding overlaps with boundary blur, but the terms emphasize different aspects: boundary blur describes an overly smooth transition, while depth bleeding highlights incorrect values crossing into another surface. Research on sharp occlusion-boundary prediction addresses both poor sharpness and misplaced contours. (openaccess.thecvf.com)

Flying pixels are usually isolated or unstable sensor points near edges, whereas bleeding may form a continuous band. Depth bleeding is also unrelated to data leakage, which occurs when training improperly uses information unavailable during real-world inference.

Link to this sectionReal-World Impact#

• Robotics and object manipulation: If a box’s background depth bleeds onto its edge, a picking robot may estimate the wrong width or choose an unsafe grasp point.
• Autonomous vehicles: Bleeding around a pedestrian can distort obstacle distance, point-cloud shape, and downstream 3D object detection.
• Augmented reality: Incorrect boundaries can make virtual objects appear in front of real furniture when they should be hidden behind it. Depth-edge errors are particularly harmful to augmented reality and novel-view synthesis. (openaccess.thecvf.com)

Link to this sectionDetection and Mitigation#

Do not rely only on average error metrics. Inspect depth-edge overlays and reconstructed point clouds, and evaluate boundary quality with datasets such as iBims-1, which includes masks for distinct depth transitions. (asg.ed.tum.de)

Practical mitigation includes using higher-resolution inputs, edge-aware losses, segmentation-guided refinement, confidence filtering, and accurately aligned training labels. The Ultralytics depth estimation task guide supports dense per-pixel prediction with YOLO26 depth models, while the depth dataset guide explains valid target formatting. Teams can manage annotation, training, deployment, and monitoring through Ultralytics Platform, with particular attention to thin objects and foreground-background boundaries.